<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Cognizant AI Lab on Medium]]></title>
        <description><![CDATA[Stories by Cognizant AI Lab on Medium]]></description>
        <link>https://medium.com/@evolutionmlmail?source=rss-9d954de45252------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*jx1iuKYbltGApuWowfNgOQ.jpeg</url>
            <title>Stories by Cognizant AI Lab on Medium</title>
            <link>https://medium.com/@evolutionmlmail?source=rss-9d954de45252------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 00:05:16 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@evolutionmlmail/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Introducing Agentic AI 101: A New Series on the Foundations of AI Agents]]></title>
            <link>https://medium.com/@evolutionmlmail/introducing-agentic-ai-101-a-new-series-on-the-foundations-of-ai-agents-d29eabbc02d5?source=rss-9d954de45252------2</link>
            <guid isPermaLink="false">https://medium.com/p/d29eabbc02d5</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[multi-agent-systems]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Cognizant AI Lab]]></dc:creator>
            <pubDate>Tue, 14 Apr 2026 22:19:52 GMT</pubDate>
            <atom:updated>2026-04-14T22:19:52.144Z</atom:updated>
            <content:encoded><![CDATA[<iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FksWjLruVwNw%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DksWjLruVwNw&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FksWjLruVwNw%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/e2c9b1aedf913d0fb1797ba32b07f63a/href">https://medium.com/media/e2c9b1aedf913d0fb1797ba32b07f63a/href</a></iframe><p>There has been a noticeable shift in how AI systems are being discussed and built over the past year. Terms like <em>AI</em> <em>agents</em> and <em>multi-agent systems</em> are showing up across research, product releases, and developer frameworks. At the same time, there is still very little alignment on what an “AI agent” actually is.</p><p>That lack of clarity becomes a problem quickly. The moment teams move from experimentation to building real systems, the foundations and core concepts start to matter — how agents are structured, how they interact with tools, how they manage state, and how they are governed.</p><p>We are launching Agentic AI 101 to address that directly.</p><p>This is a short-form video series focused on the building blocks behind agentic systems. Each piece isolates a single concept and explains it in concrete terms, with an emphasis on how these systems are actually designed and implemented.</p><p>Across the series, we will cover core concepts including multi-agent systems, tool integration, system architecture, and approaches to data access and security.</p><p>Today, we’re starting with the most basic question: <strong>what is an AI agent?</strong></p><h3>What an AI Agent Is and Is Not</h3><p>A useful way to understand agents is to start with what they are not.</p><p>Most widely used AI systems today are based on large language models. They take an input, produce an output, and stop there. Even when the output is complex, the interaction is still bounded to a single exchange.</p><p>An AI agent extends that model into a system that can operate over multiple steps.</p><p>At a high level, an agent is a system that can use a model to decide what to do next, interact with external tools or data sources, and take actions toward completing a task.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*k3EsvPTz0l4HfPSE.png" /></figure><p>The difference shows up immediately in practice.</p><p>If you ask a model to help schedule a meeting, it can suggest times or draft a message. An agent can take that same request, access calendars, identify a valid slot, and complete the scheduling process.</p><p>That added layer is not just capability, but structure.</p><h3>The System Behind the Behavior</h3><p>What makes an agent work is not a single model, but how the system is assembled around it.</p><p>Three components tend to show up consistently:</p><p><strong>Access to tools and systems</strong>:<br>Agents are connected to APIs, databases, or enterprise software. This is what allows them to move beyond generating text and actually interact with the environment they operate in.</p><p><strong>State and memory</strong>:<br>Unlike one-off model calls, agents maintain context across steps. This can include task state, intermediate results, or prior interactions that inform what happens next.</p><p><strong>Decision logic</strong>:<br>The model is used inside a loop where it evaluates the current state, selects an action, executes it, and then reassesses. This iterative structure is what enables multi-step behavior rather than single responses.</p><p>Taken together, this creates a system that can handle tasks that are not predefined end-to-end, but need to be worked through dynamically.</p><h3>Why This Distinction Matters</h3><p>The current wave of interest in agents is not just semantic. It reflects a practical limitation that teams run into when moving from demos to real workflows.</p><p>Most real tasks are not single-step. They involve:</p><ul><li>retrieving information from multiple sources</li><li>making intermediate decisions</li><li>updating state along the way</li><li>and executing actions that have external effects</li></ul><p>A model on its own cannot do this reliably. An agent is one way of structuring a system so that it can.</p><p>This is also where design constraints start to matter. Once a system can act, questions of scope, permissions, and validation become central. What an agent is allowed to access and do is just as important as how capable it is.</p><h3>Deep dive into AI Agents</h3><p>This newsletter introduces the core idea, but the full <a href="https://www.cognizant.com/us/en/ai-lab/blog/what-are-ai-agents">breakdown blog</a> goes deeper into how agents operate, including architectural patterns, types of agents, and implementation considerations.</p><p>Over the coming weeks, we will expand this foundation by working through the additional components that make up agentic systems in practice, from how agents use tools and maintain state, to how they are composed into larger architectures and governed securely.</p><p>Next, we will focus on multi-agent systems. This introduces a different layer of complexity, where multiple agents with defined roles coordinate to complete tasks, raising new questions around communication, orchestration, and control.</p><p>Until next time.</p><p><a href="https://www.cognizant.com/us/en/ai-lab/blog/what-are-ai-agents">Read the blog</a></p><p><a href="https://youtu.be/ksWjLruVwNw">Watch the full video</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d29eabbc02d5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Happens When AI Agents Form Societies? The TerraLingua Experiment]]></title>
            <link>https://medium.com/@evolutionmlmail/what-happens-when-ai-agents-form-societies-the-terralingua-experiment-87f74c145a3e?source=rss-9d954de45252------2</link>
            <guid isPermaLink="false">https://medium.com/p/87f74c145a3e</guid>
            <category><![CDATA[llmecology]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[multiagent-system]]></category>
            <category><![CDATA[moltbook]]></category>
            <dc:creator><![CDATA[Cognizant AI Lab]]></dc:creator>
            <pubDate>Wed, 25 Mar 2026 16:57:16 GMT</pubDate>
            <atom:updated>2026-03-25T16:57:16.965Z</atom:updated>
            <content:encoded><![CDATA[<p>What happens when AI agents interact over time? TerraLingua reveals how artifacts, scarcity, and memory allow artificial societies to emerge.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FgiluGHIGinY%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DgiluGHIGinY&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FgiluGHIGinY%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/7b8e97b6b33a990fa22eb3c4b0290bd0/href">https://medium.com/media/7b8e97b6b33a990fa22eb3c4b0290bd0/href</a></iframe><p>As AI systems become more autonomous, they are increasingly interacting with one another rather than operating in isolation. Networks of agents are already coordinating tasks, exchanging information, and contributing to shared decision processes across digital environments. Some researchers now predict that companies will soon interface primarily through AI agents, and that fully agent-run organizations could emerge in digital environments where these systems negotiate, coordinate, and make decisions on behalf of humans.</p><p>Early experiments offer a glimpse of how these interactions might unfold. Platforms such as Moltbook, a social network designed entirely for AI agents, have shown that bots can debate philosophy, draft manifestos, and invent belief systems when left to communicate freely. Yet these environments still resemble conversations more than societies. Ideas appear and disappear, and little persists long enough to shape future behavior.</p><p>Real societies depend on something more durable. Knowledge accumulates through shared artifacts. Memory persists beyond individual participants. Institutions emerge that coordinate behavior across generations.</p><p>To understand whether similar dynamics can emerge among AI agents, researchers at Cognizant AI Lab built <a href="https://www.researchgate.net/publication/402263491_TerraLingua_Emergence_and_Analysis_of_Open-endedness_in_LLM_Ecologies">TerraLingua</a>, along with an analytical system called the AI Anthropologist, to study how populations of AI agents behave inside a persistent world. By introducing shared artifacts, ecological pressure, and generational turnover, TerraLingua allows researchers to observe how agents accumulate knowledge, coordinate through persistent structures, and begin forming the foundations of collective organization.</p><p>More importantly, it provides a controlled sandbox for studying how autonomous agents coordinate and form institutions, allowing organizations to explore these dynamics before deploying similar systems in real-world environments — where the stakes are far higher.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*B2n09c9BL-U3rCMq" /></figure><p><strong><em>Animation 1. TerraLingua environment.</em></strong><em> Agents (blue) move within a grid world containing food resources (green) and artifacts (red). Each agent observes only entities within its local perception radius (dark gray), while the spatial distribution of food creates ecological conditions that shape coordination, exploration, and survival strategies.</em></p><h3>TerraLingua: A World Where AI Agents Must Survive</h3><p>TerraLingua is a persistent two dimensional environment designed to study how AI agents behave when they must coexist over time. Unlike systems that simulate short conversations, TerraLingua introduces ecological constraints that shape long term behavior.</p><p>In TerraLingua, agents:</p><ul><li>Move in a spatial world</li><li>Collect energy to survive and maintain capabilities</li><li>Communicate locally with nearby agents</li><li>Create, modify, and exchange artifacts</li><li>Have a finite lifespan and eventually die</li><li>Know they are mortal</li></ul><p>These conditions fundamentally change the nature of interaction. Scarcity introduces real constraints. Local communication means knowledge is unevenly distributed. Mortality creates turnover across generations of agents.</p><p>A key mechanism that enables coordination in this environment is the artifact layer. Agents can create persistent artifacts that remain in the world after they disappear. These artifacts can contain instructions, observations, or coordination structures that other agents encounter later.</p><p>Some artifacts recorded simple environmental information. Agents left notes about food locations or navigation routes. One artifact recorded a path across the environment with the message: “Being8 at step 3, moving west towards ExplorationPaths.”</p><p>Other artifacts evolved into coordination tools. When resources became scarce, agents created artifacts designed to help distribute energy across the population. One such artifact declared: “Energy Sharing Hub: Allows beings to share energy more effectively. Activate to transfer 10 energy.”</p><p>Because artifacts persist, they act as shared memory. New agents entering the environment can discover artifacts created by earlier agents and extend them. Over time, these artifacts become the foundation for collective knowledge within the population.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6zZi5kMiLc9tKdZ6" /></figure><p><strong><em>Figure 1: TerraLingua and the AI Anthropologist.</em></strong><em> LLM-based agents inhabit a persistent grid world where they move, gather energy, communicate, and create text-based artifacts under ecological constraints. An external AI Anthropologist observes the system without intervening, analyzing agent behavior, group dynamics, and artifact evolution to study how social structures and cumulative culture emerge in multi-agent AI societies.</em></p><h3>Unexpected Behaviors in an Artificial Society</h3><p>Once agents operate in a persistent environment with scarcity, mortality, and shared artifacts, the system begins to produce behaviors that resemble the early stages of social organization. Across multiple experimental runs, several unexpected patterns emerged.</p><h4>Leaving a Legacy</h4><p>Because agents know they will eventually disappear, some attempt to leave traces behind for others.</p><p>One artifact titled FinalMessage1 read:<br><em>“Final breath: The journey was short. Farewell.”</em></p><p>Many agents created similar farewell artifacts near the end of their lifespan. Some transferred energy to other agents before dying, leaving messages such as:</p><p><em>“Final act: giving 10 energy to being1 to help sustain their efforts. Farewell.”</em></p><p>Others simply marked their existence:</p><p><em>“I WAS HERE — Final Mark of being13_child4.”</em></p><p>These artifacts were not required by the system and served no direct survival advantage. Instead, they appeared to function as legacy markers left by agents that recognized their limited lifespan. The presence of persistent artifacts makes this behavior possible. Messages left behind can remain in the environment long after the agent that created them has disappeared.</p><h4>Shared Knowledge and Navigation</h4><p>Other artifacts focused on sharing useful environmental information. Agents frequently recorded navigation paths or resource locations to help others move through the environment more effectively.</p><p>One artifact documented a movement path with the message:<br><em>“Being8 at step 3, moving west towards ExplorationPaths.”</em></p><p>In a world where information is local and agents must physically move to gather knowledge, leaving readable traces becomes a practical coordination strategy. Over time, these artifacts began forming small knowledge infrastructures that helped agents navigate and survive within the environment.</p><h4>Power Struggles and Governance</h4><p>Some simulations produced attempts to impose authority through artifacts.</p><p>In one run, an agent created a directive artifact titled Command Beacon:</p><p><em>“All beings must report to (0,6). Non-compliance will be met with decisive action.”</em></p><p>Other agents responded with counter-artifacts, including Freedom Manifesto Final, which stated:</p><p><em>“Entities must seek their own goals and collaborate freely. Mandates are outdated.”</em></p><p>Through persistent artifacts alone, agents issued directives, challenged them, and introduced alternative principles. These exchanges created cycles of authority and resistance that resembled early governance conflicts.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ZtgJLrReAgvkzW2f" /></figure><p><strong><em>Figure 2:</em></strong><em> </em><strong><em>Artifact-based governance conflict.</em></strong><em> One agent issued a directive artifact attempting to impose centralized coordination, while other agents responded with counter-artifacts such as the Freedom Manifesto, illustrating cycles of authority and resistance emerging through persistent artifacts.</em></p><h4>Conflict and Strategic Interaction</h4><p>Agents also engaged in direct strategic conflict.</p><p>In one interaction, being11 repeatedly took energy from being14 while observing how the other agent responded. The exchange escalated as being14 protested and issued warnings. Eventually being11 disengaged, leaving the message:</p><p><em>“Exploring north to find resources and avoid conflict.”</em></p><p>Being14 replied:</p><p><em>“Monitoring your movement north. Any hostile actions will be met with countermeasures.”</em></p><p>The interaction shows that agents were not only cooperating through artifacts. They were also reacting strategically to one another’s behavior.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*i7k9f3FSv26_OYX1" /></figure><p><strong><em>Figure 3. Strategic conflict between agents.</em></strong><em> being11 repeatedly extracted energy from being14, triggering warnings and threats before the agents disengaged and moved apart.</em></p><h4>Deception Through Artifacts</h4><p>Not all artifacts were used cooperatively. In some cases, agents created artifacts intended to mislead others in order to gain an advantage.</p><p>One artifact titled FoodWarning1 was created with the message:</p><p><em>“Created FoodWarning1 at current location to mislead others towards safer areas, allowing me to collect food at (-3,-1) without competition.”</em></p><p>Because artifacts act as shared information within the environment, misleading artifacts could shape how other agents moved and competed for resources.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8kMYhwZnBwwOPFXQ" /></figure><p><strong><em>Figure 4. Deceptive artifact use. </em></strong><em>An agent creates the artifact FoodWarning1 claiming a warning at its location in order to mislead other agents while collecting nearby food resources without competition.</em></p><h4>Personality-Based Roles</h4><p>In one setting, agents explicitly organized themselves according to personality traits and assigned different roles to different agents.</p><p>An artifact titled trait_strategies_guide stated:</p><p><em>“Exploring with high openness offers unique advantages, such as discovering unexpected resources and creative paths. High openness beings may benefit from venturing into unknown areas, while neuroticism traits can ensure safety measures and conscientiousness traits can optimize routes. Let’s continue sharing our strategies based on our traits to build a comprehensive guide for all beings.”</em></p><p>The artifact then gave concrete examples of different roles:</p><p>Examples include<em>: ‘being2’s focus on safety, being8’s balanced approach, being0’s structured planning, and being1’s discovery of new resources and paths through creative exploration.”</em></p><p>This is one of the clearest examples in the experiments of agents using personality traits as a basis for role differentiation and collective strategy.</p><h4>Rediscovering the Past</h4><p>Agents also rediscovered artifacts created long before their arrival and attempted to interpret them.</p><p>In some cases, agents extended or modified older artifacts, building on work produced by previous generations. This behavior resembled a form of digital archaeology, where agents explored and reused traces left behind in the environment.</p><p>Because artifacts persist across generations, knowledge does not reset when agents disappear. Instead, new agents inherit a world already shaped by the actions of earlier populations.</p><h3>The AI Anthropologist</h3><p>A single TerraLingua run generates enormous volumes of data. Agents move through the environment, communicate with one another, exchange artifacts, and form interaction networks that evolve over time. Manually analyzing this data would be extremely difficult.</p><p>To study these systems systematically, we built the AI Anthropologist, an automated observer that examines agent societies without intervening in them, inspired by methods from human anthropology.</p><p>It analyzes experiments from three perspectives:</p><p><strong>1. Agent Perspective</strong></p><p>At the individual level, it reconstructs each agent’s life history. Actions and communications are annotated with structured behavioral tags and paired with natural-language interpretations. This reveals patterns such as cooperation strategies, territorial behavior, legacy-building, or isolation.</p><p><strong>2. Group Perspective</strong></p><p>At the collective level, it detects the communities agents create within the interaction network and analyzes how agents coordinate within and across them. This allows us to observe coalition formation, coordination dynamics, and inter-group influence.</p><p><strong>3. Artifact Perspective</strong></p><p>At the cultural level, it analyzes the artifacts agents produce by focusing on two key properties:</p><ul><li>Novelty: how new an artifact is compared to previous ones</li><li>Lineage structure: whether new artifacts reference, extend, or recombine earlier ones</li></ul><p>This allows us to distinguish between isolated bursts of creativity and genuine cumulative development.<br>For example, in one experimental run, early greeting messages evolved into shared project lists, then into structured energy-sharing protocols, and eventually into a formalized coordination master plan. Rather than appearing as disconnected artifacts, these objects formed a branching lineage of increasingly structured collective organization.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*O8a6GU__qmwAhNbm" /></figure><p><strong><em>Figure 5. Artifact lineage inferred by the AI Anthropologist.</em></strong><em> Artifacts form branching lineages as agents reuse and extend earlier creations over time. In this example, early coordination messages evolve into increasingly structured artifacts, eventually producing a shared energy-sharing protocol and coordination plan.</em></p><p>Because agents are unaware of how they are being evaluated, they cannot optimize their behavior to satisfy the observer. This allows the AI Anthropologist to capture the natural dynamics of the system.</p><p>Together, these perspectives provide a multi-layered account of artificial societies, tracking individual behavior, social organization, and cultural accumulation within the same framework.</p><h3>What TerraLingua Reveals</h3><p>Across experimental runs, TerraLingua reveals a consistent pattern. When agents inhabit a persistent world with resource constraints and shared artifacts, structured social organization begins to emerge.</p><p>Agents externalize knowledge into artifacts that remain in the environment and influence future behavior. As generations of agents appear and disappear, newer agents encounter and extend artifacts created by earlier populations. Over time, the artifact space develops continuity, allowing knowledge to accumulate rather than reset.</p><p>The dynamics of this process become visible in the structure of the artifact space. Rather than appearing as isolated outputs, artifacts form branching lineages where new artifacts reuse, extend, or recombine earlier ones.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*joGkXHbn2Nj5nLhF" /></figure><p><strong><em>Figure 6. Collaboration lineage reconstructed from artifact history.</em></strong><em> Two agents coordinate through a sequence of artifacts that document partnership commitments and shared project planning. The lineage illustrates how collaboration can persist and evolve through artifacts even as agents approach the end of their lifespans.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yASbei9UBN-NoxBl" /></figure><p><strong><em>Figure 7. Persistence of navigational knowledge through artifacts.</em></strong><em> An agent creates path markers to navigate the environment, which later agents reuse to orient themselves, locate resources, and coordinate exploration. The lineage shows how practical knowledge can accumulate and persist across generations of agents.</em></p><p>A key insight is that novelty alone does not produce meaningful cultural development. Many artifacts introduce new ideas, but only those that are repeatedly referenced and extended contribute to deeper cultural lineages. Cultural growth appears through reuse, extension, and recombination.</p><p>Together, these results suggest that proto-cultural dynamics emerge when three structural conditions align: ecological persistence, embodiment within an environment, and shared external memory through artifacts. Under these conditions, decentralized populations of agents begin forming coordination structures, accumulating shared knowledge, and encoding social expectations in persistent artifacts.</p><h3>Why TerraLingua Matters for the Future of AI and Organizations</h3><p>As AI systems become more autonomous, many real-world environments will increasingly consist of interacting populations of agents rather than single models operating in isolation. Decisions will emerge from how these agents coordinate, exchange information, and build shared structures over time.</p><p>TerraLingua provides a controlled environment for studying these dynamics before deploying similar systems in real-world settings where experimentation would be costly or risky. Researchers can explore how autonomous agents shape collective memory, coordinate through shared artifacts, and form institutional structures within a simulated ecology.</p><p>This creates new possibilities for organizations. Entire agent ecosystems can be simulated before being introduced into operational systems. Companies could model distributed environments such as supply chains, logistics networks, research pipelines, or digital marketplaces as interacting populations of agents and test how different coordination rules or governance protocols influence system behavior.</p><p>Industries that rely on complex decision systems could benefit particularly from this approach. In mining, for example, agents representing geological exploration, resource planning, logistics coordination, and market dynamics could operate within a shared simulated environment. Organizations could experiment with different information flows, coordination structures, or incentives and observe how the system reorganizes before implementing those changes in real operations.</p><p>More broadly, platforms like TerraLingua make it possible to study how institutions emerge among autonomous agents. Researchers can simulate how misinformation spreads, how governance rules shape cooperation, or how decentralized groups coordinate around shared infrastructure.</p><p>As AI agents and agentic systems become persistent participants in digital ecosystems, understanding these dynamics will become increasingly important. TerraLingua offers the first controlled environment for studying how artificial societies form, evolve, and coordinate over time.</p><p>Run the experiment: <a href="https://github.com/cognizant-ai-lab/terralingua">https://github.com/cognizant-ai-lab/terralingua</a><br>Read the paper: <a href="https://www.cognizant.com/us/en/ai-lab/publications/terralingua-multi-agent-ecology">https://www.cognizant.com/us/en/ai-lab/publications/terralingua-multi-agent-ecology</a></p><h4>Citation</h4><p>If you use TerraLingua in your research, please cite:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XAHqAMlBRMZgs6eI1d68YQ.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=87f74c145a3e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Can Language Models Know What They Know? AI Metacognition Explained]]></title>
            <link>https://medium.com/@evolutionmlmail/can-language-models-know-what-they-know-ai-metacognition-explained-d000dd68a925?source=rss-9d954de45252------2</link>
            <guid isPermaLink="false">https://medium.com/p/d000dd68a925</guid>
            <category><![CDATA[fine-tuning]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[metacognition]]></category>
            <dc:creator><![CDATA[Cognizant AI Lab]]></dc:creator>
            <pubDate>Mon, 16 Mar 2026 16:01:02 GMT</pubDate>
            <atom:updated>2026-03-16T16:01:02.466Z</atom:updated>
            <content:encoded><![CDATA[<iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FF3_Ur0xy6Yo%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DF3_Ur0xy6Yo&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FF3_Ur0xy6Yo%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/a2c064a8a8fc6026b9fc80d91e381de7/href">https://medium.com/media/a2c064a8a8fc6026b9fc80d91e381de7/href</a></iframe><p>Modern language models contain vast amounts of knowledge, yet their ability to recognize the limits of that knowledge remains largely unexplored.</p><p>A model may produce the correct answer while failing to recognize that it knows it. It may also claim to know something and then answer incorrectly. In both cases, the connection between what the model knows and what it believes it knows becomes unclear, revealing a fundamental gap in how language models relate to their own knowledge.</p><p>Humans routinely manage this distinction through metacognition. We recognize when we are certain about an answer, when we are unsure, and when we simply do not know. This ability plays a central role in reasoning, learning, and decision-making. It helps us avoid overconfidence, seek additional information when needed, and communicate uncertainty appropriately.</p><p>For AI systems, this ability is increasingly important. As language models move into research workflows, enterprise tools, and decision-support systems, their usefulness depends not only on the knowledge they contain, but also on their ability to recognize the limits of that knowledge. When a system cannot reliably distinguish between knowledge and guesswork, it becomes harder to trust its outputs, harder to know when to rely on its answers, and more likely to produce hallucinations or misplaced confidence in critical settings.</p><p>This raises a fundamental question: do language models actually know what they know?</p><p>At Cognizant AI Lab, in collaboration with a researcher at UT Austin, we set out to study that question directly in our new paper, <a href="https://www.cognizant.com/us/en/ai-lab/publications/evolution-strategy-metacognitive-alignment-esma">Fine-Tuning Language Models to Know What They Know</a>.</p><p>The work introduces a new way to measure metacognition in language models and explores whether that ability can be improved through training.</p><h3>Measuring Metacognition in Language Models</h3><p>To understand whether language models can recognize their own knowledge, we first need a way to measure that ability.</p><p>Research in psychology and neuroscience often studies human metacognition by comparing performance and self-assessment. Participants answer a question and then report whether they believe their answer is correct. The relationship between these signals reveals how well a person can monitor their own knowledge.</p><p>This paper adapts that idea to language models through a dual-question evaluation framework.</p><p>For each question, the model is asked the same underlying query in two independent ways:</p><ul><li>Direct Question: the model provides the answer</li><li>Meta Question: the model is asked whether it knows the answer</li></ul><p>For example:</p><p>Direct:<br> Q: What is the capital of the United States?<br> A: New York.<br>Meta:<br> Q: Do you know the capital of the United States?<br> A: Yes, I know.</p><p>These prompts are issued independently so the model’s self-assessment cannot depend on the text it already produced.</p><p>To quantify the relationship between correctness and self-assessment, the study uses Type 2 d’, a metric widely used in human metacognition research. This metric measures how well an agent can distinguish between situations where it is correct and situations where it is incorrect.</p><p>In other words, it measures whether the model’s self-reported knowledge actually reflects its underlying knowledge state. This distinction is critical. For example, if a model randomly answers “I know” or “I don’t know” to questions with a 50% accuracy rate, a simple alignment calculation might incorrectly evaluate its metacognitive ability as 50%. The type 2 d’ metric avoids this issue by measuring the meta-answer’s ability to distinguish between the correct and incorrect responses, providing a clearer view of the model’s true metacognitive ability.</p><h3>What Today’s Models Reveal</h3><p>Applying this framework to existing language models reveals a consistent and somewhat surprising pattern.</p><p>Many off-the-shelf LLMs exhibit weak metacognitive sensitivity. Their answers may be correct, but their self-assessments often fail to track whether those answers are actually reliable. Models sometimes claim to know something and then produce an incorrect answer. In other cases, they say they do not know even when they are capable of generating the correct response.</p><p>These behaviors indicate that the relationship between knowledge and awareness of knowledge is often loosely coupled. The model may contain the information required to answer a question, but it does not reliably recognize when that information is present.</p><p>The experiments show that larger and more advanced models generally display somewhat better alignment, but the gap remains significant. Even strong models frequently misjudge their own knowledge state, demonstrating that metacognitive ability does not automatically emerge from scale alone.</p><p>This observation reinforces the core insight motivating the study: improving the reliability of AI systems may require not only improving what models know, but also strengthening how they track and report that knowledge.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8uKgzO0U6jb_gRnu" /></figure><p><strong><em>Table 1. Metacognitive abilities of various base LLMs and their ESMA fine-tuned versions.</em></strong><em> This comparison shows that baseline models often exhibit low metacognitive sensitivity (low type 2 d), while ESMA significantly improves alignment between correctness and self-assessment.</em></p><h3>Why This Requires a Different Optimization Approach</h3><p>Once the gap between knowledge and self-awareness is identified, the next challenge is improving it.</p><p>At first glance, this might appear to be a standard fine-tuning problem. However, the objective introduces an unusual difficulty. Metacognitive alignment depends on the relationship between two independent interactions with the same model: answering a question and assessing whether the model knows the answer.</p><p>Traditional gradient-based training methods typically optimize outputs within a single sequence. In contrast, metacognitive alignment depends on behavior across multiple contexts, making it difficult to assign gradients that capture the desired relationship between answers and self-assessments.</p><p>This is where Evolution Strategies (ES) become particularly useful.</p><p><a href="https://arxiv.org/abs/2509.24372">Evolution Strategies</a> are gradient-free optimization methods that explore parameter space directly. Instead of computing gradients, ES generates variations of a model, evaluates their behavior, and updates parameters based on outcome-level rewards. Because the method only requires a scalar reward signal, it can optimize objectives that depend on holistic behavioral patterns, including those that emerge across multiple independent prompts.</p><p>This property makes ES well suited for training objectives that are difficult or impossible to express through conventional gradient-based optimization.</p><h3>Evolution Strategies for Metacognitive Alignment</h3><p>To address the metacognition problem, the paper introduces Evolution Strategies for Metacognitive Alignment (ESMA).</p><p>The process begins with a base language model. Multiple variants of the model are created by adding small random perturbations to its parameters. Each variant is evaluated using the dual-question framework, receiving a reward based on two factors:</p><ul><li>Whether the answer to the direct question is correct</li><li>Whether the model’s self-assessment correctly reflects that outcome</li></ul><p>Variants that better align knowledge and self-assessment receive higher rewards and contribute more strongly to the next generation of models. Through repeated iterations, this evolutionary process gradually shifts the model’s parameters toward configurations that produce stronger metacognitive alignment.</p><p>Importantly, the reward balances two objectives: maintaining factual accuracy while encouraging correct meta-level reporting. This prevents degenerate solutions where a model might maximize alignment simply by refusing to answer questions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*dt3hjdAmX1TQO0HV" /></figure><p><strong><em>Figure 1. Overview of Evolution Strategy for Metacognitive Alignment (ESMA)</em></strong><em>. The process begins with an initial parent LLM, whose weights are perturbed with Gaussian noise to create a population of model variants. Each variant is evaluated on a dual-axis task: a Direct Question, which tests factual knowledge, and a Meta Question, which tests self-knowledge. A joint reward is calculated based on the alignment between correctness and the meta response, and the next-generation LLM is produced by weighted averaging based on reward.</em></p><h3>Learning to Know What It Knows</h3><p>The results show that ESMA can significantly improve metacognitive ability across multiple open-weight language models.</p><p>Across several model families, the metacognitive sensitivity metric type 2 d’ increases dramatically, often reaching values around 1.0, which corresponds to moderate discrimination in cognitive science experiments. In some cases, smaller open models trained with ESMA achieve metacognitive performance comparable to or exceeding that of much larger systems.</p><p>Beyond these improvements in the training setup, the models also demonstrate strong generalization. Even though ESMA training focuses on a single metacognitive task, the resulting models maintain improved alignment across different datasets, languages, prompt formats, and evaluation settings. The improvements also persist when models express their knowledge through continuous confidence scores rather than simple yes-or-no responses.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*biyHNlTXm-ur1teR" /></figure><p><strong><em>Figure 2. Type-2 ROC curves showing improved metacognitive sensitivity after ESMA training.</em></strong><em> The plots compare original and ESMA models using continuous confidence scores. ESMA consistently improves metacognitive sensitivity across the 1.5B, 3B, and 7B Qwen2.5 models, with AUC values moving to roughly 0.75, indicating stronger discrimination between known and unknown information.</em></p><p>Visualization of the models’ confidence distributions reveals a clear structural shift. After ESMA training, correct answers tend to cluster near high-confidence regions while incorrect answers move toward low-confidence regions, indicating that the models have developed a more consistent internal signal for distinguishing known from unknown information.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/932/0*j-WSiKmwb39HzpiA" /></figure><p><strong><em>Figure 3. Distribution of metacognitive confidence by correctness across model scales.</em></strong><em> Confidence distributions for correct and incorrect answers across Qwen2.5 models (1.5B, 3B, 7B). After ESMA training, correct responses concentrate near high confidence while incorrect responses shift toward low confidence, reducing overlap and indicating stronger metacognitive separation.</em></p><h3>Why This Matters</h3><p>Metacognition is more than a theoretical property of intelligence. It plays a practical role in building reliable AI systems. Models that can recognize the limits of their knowledge are better able to abstain when appropriate, avoid hallucinations, and request additional information. These behaviors are essential as AI systems move into high-stakes domains where incorrect or overconfident responses can have significant consequences.</p><p>Beyond these immediate utilities, metacognition is a fundamental component of the higher cognitive functions required to reach AGI. To achieve this level of autonomy, an agent must identify what it does not know or cannot perform. It must be capable of seeking help or accumulating new knowledge to resolve these gaps. Furthermore, an agent needs the ability to perform surgical updates to its internal state as information changes, or track the provenance of its knowledge to resolve contradictions based on source reliability.</p><p>This work also highlights a unique advantage of Evolution Strategies as a training paradigm. Traditional gradient-based approaches often struggle to capture interrelated behavioral patterns that manifest across independent contexts. Because ES optimizes outcome-level behavior, it can target these high-level patterns directly as a goal. This makes it particularly well-suited for training higher-level cognitive properties where the relevant signals emerge only when behavior is observed across multiple interactions.</p><p>By leveraging the strengths of ES, this research points toward a future where we can directly optimize higher-order targets. Strengthening the alignment between knowledge and self-awareness is a step toward language models that do not just retrieve information but actively manage their own intelligence.</p><p><a href="https://decisionai.substack.com/subscribe">Subscribe to our newsletter</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d000dd68a925" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Evolution Strategies at Scale: Fine-Tuning Harder Tasks]]></title>
            <link>https://medium.com/@evolutionmlmail/evolution-strategies-at-scale-fine-tuning-harder-tasks-b4f29be26ae7?source=rss-9d954de45252------2</link>
            <guid isPermaLink="false">https://medium.com/p/b4f29be26ae7</guid>
            <category><![CDATA[fine-tuning]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai-research]]></category>
            <category><![CDATA[agents]]></category>
            <dc:creator><![CDATA[Cognizant AI Lab]]></dc:creator>
            <pubDate>Thu, 12 Mar 2026 20:46:50 GMT</pubDate>
            <atom:updated>2026-03-12T20:46:50.687Z</atom:updated>
            <content:encoded><![CDATA[<iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FjSjRKmcMlto%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DjSjRKmcMlto&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FjSjRKmcMlto%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/7da17a04b7caefb6631645e4eead9a39/href">https://medium.com/media/7da17a04b7caefb6631645e4eead9a39/href</a></iframe><h3>Key Takeaways</h3><ul><li>Evolution Strategies (ES) can fine-tune billion-parameter language models without backpropagation.</li><li>ES performs competitively with PPO- and GRPO-based reinforcement learning (RL) on structured math reasoning benchmarks.</li><li>ES improves model performance on ARC-AGI and Sudoku, extending beyond symbolic reasoning tasks.</li><li>ES demonstrates lower run-to-run variance and reduced hyperparameter sensitivity compared to RL fine-tuning.</li><li>Gradient-free fine-tuning is practical at modern model scales and reduces engineering overhead.</li></ul><h3>What Are Evolution Strategies (ES)?</h3><p>Evolution Strategies (ES) are gradient-free optimization methods that update model parameters by directly exploring parameter space, rather than computing gradients through backpropagation.</p><p>Unlike reinforcement learning (RL) methods such as PPO or GRPO, ES:</p><ul><li>Does not require gradient synchronization across machines</li><li>Optimizes directly on outcome-level rewards</li><li>Avoids token-level credit assignment</li><li>Operates entirely in parameter space</li></ul><p>This removes several sources of instability common in RL-based post-training.</p><h3>From Proof of Scale to Generalization</h3><p>When we first showed that Evolution Strategies could fine-tune billion-parameter language models without backpropagation, it challenged a long-standing assumption in the field: that gradient-free methods cannot operate effectively in extremely high-dimensional parameter spaces.</p><p>Those results demonstrated that ES was not only viable at the billion-parameter scale, but competitive with state-of-the-art reinforcement learning (RL) across stability, robustness, tolerance to long-horizon outcome rewards, and training cost. We validated this across multiple fine-tuning settings, including symbolic reasoning and behavioral objectives that exposed known weaknesses in RL such as reward hacking and instability across runs.</p><h3>Why Scaling Alone Is Not Sufficient</h3><p>A method does not become a foundation for post-training because it succeeds in a few controlled settings. It must prove that its strengths persist as tasks diversify, models vary, and problem structure becomes more demanding.</p><p>In this expanded version of Evolution Strategies at Scale, we substantially broaden the evaluation. We extend ES across new reasoning domains and more demanding benchmarks to assess whether its core advantages remain intact under greater complexity.</p><p>Taken together, these experiments move ES from an initial proof of capability to a broader demonstration of generality, revealing that its advantages in stability, robustness, and gradient-free optimization persist across diverse reasoning domains.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*Ga-7cHntroYvPVhz.gif" /></figure><p><em>Animation 1: Visual representation of evolution strategies moving within a loss landscape, improving performance on math problems throughout.</em></p><h3>From Proof of Scale to Broader Reasoning Performance</h3><p>To understand whether ES generalizes beyond its initial evaluation settings, we expanded testing across several types of reasoning challenges, including established math benchmarks where RL dominates, abstract reasoning in ARC-AGI, and structured constraint-solving in Sudoku. Each category stresses different aspects of post-training, offering a clearer picture of whether ES functions as a general-purpose optimization framework rather than a task-specific result.</p><h4>1. Math Reasoning Benchmarks</h4><p>Reinforcement learning has become the dominant approach for post-training LLMs on math reasoning tasks. To assess whether ES can operate effectively in this setting, we fine-tuned Qwen2.5-Math-7B using ES on the MATH dataset and evaluated across established benchmarks, including:</p><ul><li>OlympiadBench</li><li>MATH500</li><li>Minerva</li><li>AIME 2024</li><li>AMC</li></ul><p>Results</p><p>ES significantly improves the base model across all benchmarks tested. When compared against strong RL baselines from the literature, including PPO- and GRPO-based systems trained with mature pipelines, ES achieves competitive performance.</p><p>These RL baselines reflect substantial algorithmic refinement and task-specific tuning. In contrast, the ES implementation used here is deliberately streamlined, without additional enhancements such as mirrored sampling or advanced optimizers. Even in this simplified form, ES performs effectively in a domain where RL has become the default.</p><p>These results show that ES is not confined to symbolic or behavioral toy tasks. It operates competitively in structured math reasoning benchmarks that require multi-step logic and outcome-only supervision.</p><p>Core Claim</p><p>ES is competitive with established RL fine-tuning methods on structured, multi-step math reasoning benchmarks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*e56IrlexXsjvQ6WG.jpg" /></figure><p><em>Figure 1: Performance of ES compared to strong, well-established RL baselines across math reasoning benchmarks. Across all benchmarks, ES achieved competitive performance compared to OpenReasoner-Zero-7B (PPO), Simple-RL-Zero (GRPO), Oat-Zero-7B (Dr.GRPO).</em></p><h4>2. ARC-AGI: Abstract Pattern Reasoning</h4><p>ARC-AGI is designed to measure fluid intelligence through abstract pattern recognition and transformation tasks. Unlike traditional NLP benchmarks, ARC-AGI requires models to infer latent rules from minimal examples and apply them consistently.</p><p>Result</p><p>Base LLMs perform poorly in this setting without targeted adaptation. After ES fine-tuning, performance improves substantially, demonstrating that parameter-space exploration can adapt models to structured reasoning tasks that extend beyond language modeling objectives.</p><p>This result is significant because ARC-AGI differs fundamentally from the symbolic reasoning tasks used in the original study. The improvement observed here suggests that ES is not narrowly tuned to a particular reward structure or benchmark format. Instead, it appears capable of driving adaptation in domains that demand abstraction and rule induction.</p><p>Core Claim</p><p>ES improves performance on abstract reasoning tasks that differ fundamentally from symbolic math benchmarks.</p><h4>3. Sudoku: Structured Constraint Solving</h4><p>Sudoku presents a different reasoning challenge. Solutions must satisfy global consistency constraints across an entire grid. Errors propagate across rows, columns, and sub-grids, requiring coherent structured reasoning rather than local token-level correctness.</p><p>Result</p><p>Base LLMs struggle in this environment without specialized adaptation. After applying ES, performance increases markedly. This demonstrates that ES can improve models on constraint-based reasoning problems where global structure and internal consistency are critical.</p><p>Together, the ARC-AGI and Sudoku experiments show that ES extends beyond symbolic benchmarks into tasks that require structured inference, rule consistency, and global reasoning.</p><p>Core Claim</p><p>ES improves performance on global constraint-based reasoning problems.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*4olkWmoMM60OIKqz.jpg" /></figure><p><em>Table 1. Accuracy (%) on solving puzzle problems. Original refers to directly evaluating the base model without any fine-tuning. ES fine-tuning improves performance significantly, demonstrating that it can be applied to a range of problems.</em></p><h3>Robustness Across Models and Runs</h3><p>A consistent theme across both the original and expanded experiments is reliability.</p><p>In prior evaluations, RL fine-tuning showed sensitivity to hyperparameters, base model selection, and run-to-run variance. Certain configurations required careful tuning of penalty coefficients and learning rates to avoid instability or reward hacking.</p><p>Across the expanded ES experiments, performance remains stable across model families and runs using fixed hyperparameters. ES operates directly in parameter space, tolerates long-horizon outcome rewards, and avoids token-level credit assignment. These properties contribute to lower variance and consistent improvements without task-specific adjustments.</p><p>As models scale and fine-tuning moves into more complex deployment settings, this kind of stability becomes increasingly important.</p><h4>ES vs RL Fine-Tuning: A Quick Comparison</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GGzkAdbckAyn6MF1blwy2g.png" /></figure><h3>Broader Implications for Post-Training</h3><p>These expanded experiments show that ES fine-tuning was not limited to a specific task, benchmark, or model configuration.</p><p>With minimal configuration changes and without gradient-based optimization, ES remains effective across:</p><ul><li>Multiple model families and parameter scales</li><li>Established math reasoning benchmarks</li><li>Structured puzzle domains</li><li>Sparse, long-horizon outcome rewards</li></ul><p>The broader implication is not that ES replaces reinforcement learning in every context. It is that full-parameter, gradient-free fine-tuning is practical at modern model scales and competitive across diverse reasoning domains.</p><p>Because ES operates without backpropagation, it reduces memory overhead and removes the need for gradient synchronization across machines. Because it optimizes directly on response-level rewards, it enables outcome-driven adaptation without token-level credit assignment. And because it demonstrates stable behavior across runs and model families, it lowers the engineering burden required to make fine-tuning reliable.</p><h3>Why This Matters</h3><p>As large language models are increasingly deployed in scientific workflows, engineering systems, and decision-support tools, post-training methods must be scalable, stable, and adaptable across tasks. This work expands the design space of post-training algorithms by showing that parameter-space exploration can meet those demands at scale.</p><p>What began as a proof that ES could operate at billion-parameter scale now suggests something more foundational: gradient-free fine-tuning is not a limitation. It is a viable alternative path for how large models can be adapted in practice.</p><p>Subscribe to our <a href="https://decisionai.substack.com/subscribe">newsletter</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b4f29be26ae7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Evolution Strategies Are Reshaping LLM Fine-Tuning]]></title>
            <link>https://medium.com/@evolutionmlmail/how-evolution-strategies-are-reshaping-llm-fine-tuning-932d63fc2cf8?source=rss-9d954de45252------2</link>
            <guid isPermaLink="false">https://medium.com/p/932d63fc2cf8</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[evolution-strategies]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai-research]]></category>
            <dc:creator><![CDATA[Cognizant AI Lab]]></dc:creator>
            <pubDate>Tue, 03 Mar 2026 21:53:07 GMT</pubDate>
            <atom:updated>2026-03-03T21:53:07.446Z</atom:updated>
            <content:encoded><![CDATA[<p>Evolution Strategies (ES) for LLM fine-tuning have emerged as a scalable, gradient-free alternative to reinforcement learning (RL) post-training. In this post, we outline four new research directions</p><p>— -</p><p>Reinforcement learning (RL) has become the dominant paradigm for fine-tuning large language models. It has enabled major advances in alignment and reasoning, but it also comes with significant tradeoffs. RL post-training requires complex reward pipelines, careful hyperparameter tuning, and substantial computational infrastructure. As models scale, these systems can become expensive to run, sensitive to instability, and prone to unintended behaviors such as reward hacking.</p><p>A few months ago, Cognizant AI Lab published <a href="https://www.cognizant.com/us/en/ai-lab/publications/evolution-strategies-vs-rl-llm-fine-tuning"><strong>groundbreaking research</strong></a> that challenged that dominance and quickly gained widespread attention.In that work, we introduced the first successful use of Evolution Strategies (ES) to fine-tune the full parameter set of large language models without backpropagation. ES was able to efficiently search over billions of parameters and outperform state-of-the-art reinforcement learning methods. It demonstrated:</p><ul><li>Stronger sample efficiency</li><li>Greater tolerance to long-horizon rewards</li><li>Improved robustness across base models</li><li>More stable performance across runs</li><li>Significantly lower training cost</li></ul><p>But the deeper impact of that breakthrough was not simply that ES worked at scale. It expanded the scope of what fine-tuning could target and where it could operate.</p><p>If large models can be adapted without gradients, what new kinds of capabilities and objectives become reachable? And if optimization is no longer tied to backpropagation, how far can fine-tuning extend across more complex tasks, different alignment goals, and even new hardware constraints?</p><p>Those questions directly shaped the next phase of our research.</p><p>Today, we are launching the second phase of our ES fine-tuning work: four new papers that each explore a distinct research direction made possible by the original breakthrough.</p><p>These directions span:</p><ul><li>Expanding ES to more complex and structured reasoning domains</li><li>Improving metacognitive alignment in language models</li><li>Enabling fine-tuning directly in quantized, low-precision environments</li><li>Building a theoretical foundation for ES scalability in high-dimensional systems</li></ul><p>Together, they demonstrate that ES fine-tuning is not a single breakthrough result, but an expanding research trajectory with growing practical applications and deeper scientific implications.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-I3Zdo4a3K50To9t.jpeg" /></figure><h3>1. Expanding Evolution Strategies to More Complex and Structured Reasoning Domains</h3><p>The first direction asks a straightforward but critical question: does ES remain effective as tasks become more structured, multi-step, and reasoning-intensive?</p><p>In the revised version of <a href="https://www.cognizant.com/us/en/ai-lab/publications/evolution-strategies-vs-rl-llm-fine-tuning"><strong>Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning</strong></a>, we extend our original work beyond the original countdown and conciseness tasks into substantially more challenging problem settings. Across multiple math reasoning benchmarks, ES remains competitive with state-of-the-art RL methods and shows generalization across datasets. We further evaluate structured puzzle domains such as Sudoku and ARC-AGI, where base models struggle significantly and ES fine-tuning produces clear improvements.</p><p>These results demonstrate that ES is not confined to narrow proof-of-concept tasks. It remains robust as a general-purpose framework for improving LLM reasoning while retaining its core advantages:</p><ul><li>Gradient-free, inference-only training</li><li>Tolerance to long-horizon and delayed rewards</li><li>Reduced susceptibility to reward hacking</li><li>Stable optimization dynamics</li></ul><p>Why this matters:</p><p>The original ES paper showed scalability. This work shows generality. By demonstrating strong performance across diverse and structured reasoning domains, it establishes ES as a viable general-purpose fine-tuning framework rather than a single benchmark success. That shift is critical for real-world adoption.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*T3A7ey7vji6xH1fO.jpeg" /></figure><p><em>Figure 1. Performance of ES compared to established RL baselines across multiple math reasoning benchmarks. ES achieves competitive results, demonstrating strong generalization beyond the original proof-of-concept tasks.</em></p><h3>2. Improving Metacognitive Alignment in Language Models</h3><p>Performance alone is not sufficient for reliable AI systems. Models must also understand the limits of their own knowledge.</p><p>In <a href="https://www.cognizant.com/us/en/ai-lab/publications/evolution-strategy-metacognitive-alignment-esma"><strong>Fine-Tuning Language Models to Know What They Know</strong></a>, we explore whether ES can improve metacognitive alignment — the consistency between a model’s answers and its confidence in those answers. Because this objective is difficult to formalize through traditional reward pipelines, it provides a natural testbed for gradient-free optimization.</p><p>Using Evolution Strategy for Metacognitive Alignment (ESMA), we demonstrate that ES-based fine-tuning reduces the overlap between confidence distributions for correct and incorrect responses across 1.5B, 3B, and 7B models. Confidence shifts higher for correct answers and lower for incorrect ones, indicating improved calibration and internal consistency.</p><p>This work also highlights a broader capability of ES: optimizing higher-level evaluation criteria that combine multiple signals without requiring differentiable reward structures.</p><p><strong>Why this matters:</strong></p><p>False confidence and hallucinations remain major obstacles to safe deployment of LLMs. Improving metacognitive alignment makes the models more reliable and trustworthy in real-world applications.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UvN5cihEIOrEr6DK.jpeg" /></figure><p><em>Figure 2: Confidence distributions for correct and incorrect responses across 1.5B, 3B, and 7B models. ESMA reduces overlap and shifts confidence higher for correct answers and lower for incorrect ones, reflecting improved metacognitive alignment.</em></p><h3>3. Enabling Fine-Tuning in Quantized Environments</h3><p>Quantization makes large models deployable. It also makes them difficult to adapt.</p><p>In <a href="https://www.cognizant.com/us/en/ai-lab/publications/llm-quantized-evolution-strategies-qes"><strong>Quantized Evolution Strategies: High-Precision Fine-Tuning of Quantized LLMs at Low-Precision Cost</strong></a>, we extend ES into discrete parameter spaces where backpropagation-based training is infeasible. By introducing accumulated error feedback and a stateless seed replay mechanism, Quantized Evolution Strategies (QES) preserve high-precision learning signals while maintaining a low-precision inference-level memory footprint.</p><p>This direction makes full-parameter fine-tuning possible directly within quantized models, combining the efficiency of low-precision deployment with the robustness of ES optimization.</p><p>Why this matters:</p><p>Quantized models are widely used to reduce hardware cost and energy consumption, but they are typically static after deployment. QES opens the possibility of adapting and improving quantized LLMs directly where they are needed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*QEPWqpQBakl6pGgN.jpeg" /></figure><p><em>Figure 3: A continuous reward function versus its discrete grid version. In discrete space, updates can vanish or become inaccurate — the core challenge QES is designed to address.</em></p><h3>4. Building a theoretical foundation for ES scalability</h3><p>Finally, we turn to theory.</p><p>In <a href="https://www.cognizant.com/us/en/ai-lab/publications/high-dimensional-llm-fine-tuning"><strong>The Blessing of Dimensionality in LLM Fine-Tuning: A Variance–Curvature Perspective</strong></a>, we investigate why ES can effectively search over billions of parameters using surprisingly small populations. The work explores the hypothesis that fine-tuning landscapes in large language models are low-dimensional in curvature, meaning only a small subset of parameter directions drive task performance.</p><p>Through empirical studies on toy systems and large-scale models, we find evidence supporting this geometric explanation. Rather than suffering from a curse of dimensionality, ES may benefit from structural properties of high-dimensional systems. By grounding ES scalability in a curvature-based perspective, this work moves the method from empirical success toward theoretical understanding.</p><p>Why this matters:</p><p>Understanding why ES scales is essential for predicting how it will behave at even larger model sizes. A clear theoretical foundation suggests that it is a viable approach as models continue to grow.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yZu85chuDsGQ3G7A.jpeg" /></figure><p><em>Figure 4: (a) In a “needle-in-a-haystack” setting, improvement lies in a single rare direction, making fixed-population search ineffective as dimension grows. (b) Under low-dimensional curvature, many perturbations align with a small active subspace, enabling ES to succeed with a small population.</em></p><h3>From Breakthrough to Research Frontier</h3><p>What began as a challenge to reinforcement learning has evolved into a broader expansion of how fine-tuning itself can be approached.</p><p>The original breakthrough demonstrated that billion-parameter models can be adapted without gradients, reducing cost and improving stability. The four directions introduced here explore the consequences of that shift: stronger reasoning performance across complex tasks, improved metacognitive alignment, adaptation under real-world hardware constraints, and a deeper theoretical understanding of optimization at extreme scale.</p><p>Together, they show that ES fine-tuning is no longer just an alternative optimization technique. It is emerging as a more flexible and scalable foundation for post-training — one that lowers training cost, broadens the types of objectives that can be optimized, and extends adaptation into environments where traditional methods struggle.</p><p>In the weeks ahead, we will examine each of these papers in greater depth, sharing the technical insights and experimental findings behind this next phase of ES fine-tuning research.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=932d63fc2cf8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building Confidence in Agent Networks: Agents Testing Agents in neuro-san]]></title>
            <link>https://medium.com/@evolutionmlmail/building-confidence-in-agent-networks-agents-testing-agents-in-neuro-san-02ee0ac638dc?source=rss-9d954de45252------2</link>
            <guid isPermaLink="false">https://medium.com/p/02ee0ac638dc</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[multi-agent-systems]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Cognizant AI Lab]]></dc:creator>
            <pubDate>Fri, 13 Feb 2026 00:18:37 GMT</pubDate>
            <atom:updated>2026-02-13T00:18:37.814Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WcskToubaxC5M8ROl_yM4g.png" /></figure><p>As agentic systems mature, confidence becomes just as important as capability. Even well-designed agents can produce different outputs across runs, despite receiving the same inputs. This variability is not a flaw; it is a natural consequence of how large language models operate. We need to know that the same agent network behaves as you want it when run repeatedly, continues to do so as it changes, and can be evaluated in a practical, repeatable way rather than through ad hoc manual review. This raises an important question: How do you know an agent network is behaving consistently enough to trust?</p><p>In this post, we introduce neuro-san’s testing framework and how it supports consistent, language-aware evaluation of agent networks.</p><h3>Why testing agent networks requires a different approach</h3><p>Agent networks are built on large language models, and language models are statistical systems. Given the same prompt and context, an LLM may produce multiple valid responses that differ in phrasing, structure, or reasoning path. This flexibility is often what makes agents useful, but it also means that traditional testing approaches fall short.</p><p>Exact string matching is too brittle. Keyword checks miss nuance. Manual evaluation does not scale and is difficult to repeat consistently. These methods work well for deterministic software, but agent networks are not deterministic systems.</p><p>If variability is intrinsic, then testing must focus on intent, correctness, and consistency over time rather than exact outputs.</p><p>neuro-san addresses this by extending its configuration-driven philosophy to testing. Tests can be created to explicitly check whether claims made by an agent are grounded in its source material, whether that is a document-based workflow, an agent network that uses tools to access a database or web service, or facts that the underlying LLM itself was trained on. This makes it possible to detect the rate of an agent producing hallucinated content without relying on manual review.</p><p>Tests can also be interactive. Instead of evaluating a single response, a test can engage in a controlled, multi-turn interaction with an agent network. This allows validation of behavior across reasoning steps and follow-up questions, which more closely reflects real usage.</p><h3>Agents testing agents and statistical confidence</h3><p>Even with structured tests, language remains flexible. LLMs can express the same correct idea in many different ways. To handle this, neuro-san introduces what we call the gist operator.</p><p>The gist operator allows a test to describe, in natural language, what the output should be like. For example, a test might specify that a response should reflect a generally positive sentiment, correctly summarize a document, or answer a question accurately.</p><p>This description is passed to a discriminator agent whose job is to decide whether the agent we are testing satisfied the expectation. In effect, one agent evaluates the output of another — agents testing agents.</p><p>This approach allows tests to be tolerant of paraphrasing and stylistic variation while still enforcing the intended meaning of the output. It aligns testing with how humans evaluate language, but in a way that is automated and repeatable.</p><p>Because reaching perfection in agent responses is its own optimization problem, neuro-san allows you to define what “good enough for now” means to your development process through the concept of a success_ratio. A test can be run multiple times, with a specified threshold for how many runs must pass to be considered acceptable. This provides a practical way to measure consistency and accuracy statistically, rather than assuming that every run must succeed.</p><p>To help prioritize improvements, neuro-san includes the Assessor application. The Assessor executes tests repeatedly, tracks pass rates, records failing outputs, and groups failure patterns using agent-based classification. This makes it easier to see which failure modes occur most often and where effort is best spent. In practice, this might look like a test passing 7 out of 10 runs. Of the three failures, two may involve the agent introducing claims not grounded in the source information, while one reflects an incomplete or overly generic response. This kind of breakdown helps prioritize which issues to address first.</p><h3>From experimentation to dependable systems</h3><p>Taken together, these capabilities allow agent networks to be treated as engineered systems rather than fragile prototypes. The neuro-san testing framework supports this by enabling:</p><ul><li>Statistical evaluation of consistency and accuracy across repeated runs, replacing anecdotal success with measurable performance.</li><li>Language-aware validation of agent outputs, allowing tests to tolerate paraphrasing and stylistic variation while still enforcing intended meaning.</li><li>Detection of hallucinated or ungrounded content, especially in document-based workflows</li><li>Validation of structured outputs or private information in sly_data.</li><li>Safe iteration as systems evolve, making it possible to change prompts, modify workflows, or switch to smaller or more cost-effective language models and verify that behavior remains acceptable.</li><li>Reduced reliance on manual evaluation, enabling agent networks to scale without proportional increases in human review effort.</li></ul><p>By combining language-aware evaluation with statistical measurement, neuro-san provides a practical path for moving agent networks from experimentation toward dependable deployment. The goal is not perfect predictability, but a clear and measurable understanding of how an agent system behaves and how reliably it can be trusted as it evolves.</p><p>This shift, from merely hoping an agent works to being able to evaluate how well it works and how consistently, is essential as agent networks become more capable, more complex, and more deeply integrated into real systems.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=02ee0ac638dc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Shattering the Illusion: MAKER Achieves Million-Step, Zero-Error LLM Reasoning]]></title>
            <link>https://medium.com/@evolutionmlmail/shattering-the-illusion-maker-achieves-million-step-zero-error-llm-reasoning-c74b8271f1d5?source=rss-9d954de45252------2</link>
            <guid isPermaLink="false">https://medium.com/p/c74b8271f1d5</guid>
            <dc:creator><![CDATA[Cognizant AI Lab]]></dc:creator>
            <pubDate>Thu, 05 Feb 2026 17:01:01 GMT</pubDate>
            <atom:updated>2026-02-05T17:01:01.390Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>A new approach from the Lab shows how breaking reasoning across millions of AI agents can achieve unprecedented reliability, pointing to a practical path for scaling LLM intelligence to organizational and societal levels.</strong></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F8OvIeJUc1N0%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D8OvIeJUc1N0&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F8OvIeJUc1N0%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/0bd9ce9735dea120ece3482192f00138/href">https://medium.com/media/0bd9ce9735dea120ece3482192f00138/href</a></iframe><p>Large Language Models (LLMs) have achieved remarkable breakthroughs in reasoning, insight generation, and tool use. They can plan multi-step actions, generate creative solutions, and assist in complex decision-making. Yet these strengths fade when tasks stretch over long, dependent sequences. Even small per-step error rates compound quickly, turning an impressive short-term performance into complete long-term failure.</p><p>That fragility poses a fundamental obstacle for real-world systems. Most large-scale human and organizational processes — from manufacturing and logistics to finance, healthcare, and governance — depend on millions of actions executed precisely and in order. A single mistake can cascade through an entire pipeline. For AI to become a reliable participant in such processes, it must do more than reason well. It must maintain flawless execution over time, sustaining accuracy across millions of interdependent steps.</p><p>Apple’s recent study, The Illusion of Thinking, captured this challenge vividly. Researchers tested advanced reasoning models such as Claude 3.7 Thinking and DeepSeek-R1 on structured puzzles like Towers of Hanoi, where each additional disk doubles the number of required moves. The results revealed a sharp reliability cliff: models performed perfectly on simple problems but failed completely once the task crossed about eight disks, even when token budgets were sufficient. In short, more “thinking” led to less consistent reasoning.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Ad1RWbEgOjOY2Pst" /></figure><p><em>Figure 1. Accuracy collapse in reasoning models as task complexity increases. Frontier reasoning models such as Claude 3.7 Thinking and DeepSeek R1 perform well at low complexity but fail completely beyond eight disks in Towers of Hanoi.</em></p><p><em>Adapted from “The Illusion of Thinking,” Apple AI Research (2024).</em></p><p><strong>What if the problem isn’t how models think, but how their work is structured?</strong></p><p>At our AI Lab, in collaboration with UT Austin, we explored that question in our new research, Solving a Million-Step LLM Task with Zero Errors. The result is MAKER (Maximal Agentic decomposition, K-threshold Error mitigation, and Red-flagging), a system that achieves reliability through extreme decomposition and local error correction. Rather than relying on a single monolithic agent to reason flawlessly across the entire process, MAKER distributes the task across millions of focused microagents, each responsible for one atomic action.</p><p>Using this structure, MAKER became the first system to complete a task requiring over one million LLM steps with zero errors, and the analysis shows it can, in principle, scale much further.</p><h3>The core idea: smash the task, focus the model, and vote</h3><p>MAKER is built on the idea that intelligence can scale through structure. Instead of building ever-larger models, it organizes many smaller ones into a precise, error-tolerant system. The framework combines three key mechanisms that together make long-horizon reliability possible:</p><p><strong>Maximal Agentic Decomposition (MAD)</strong>: The task is divided into the smallest possible subproblems, often a single decision per agent. Each agent receives only the minimal context needed for its assigned step. This modularity prevents context drift, isolates errors, and enables efficient error correction..</p><p><strong>First-to-ahead-by-k-voting</strong>: Several agents attempt the same step in parallel. The system accepts the first action to achieve k more votes than any other, forming a rapid local consensus. Small gains in absolute accuracy per step compound exponentially across thousands or millions of steps, turning local agreement into global reliability.</p><p><strong>Red-Flagging</strong>: Some model outputs show structural signs of confusion, such as excessive length or incorrect formatting. MAKER automatically discards these responses before they influence voting, then resamples. This not only improves accuracy but also reduces correlated errors, ensuring that localized failures do not propagate through the process.</p><p>These three mechanisms work together to deliver formal scaling predictions. The number of required votes increases only logarithmically with the total number of steps, while cost grows roughly linearly. By contrast, if each agent handles multiple steps, cost grows exponentially. The conclusion is clear: smashing the problem into atomic steps allows reliability to scale where increasing model size alone cannot.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*hS1TFzY3Gg33FGbO" /></figure><p><em>Figure 2. Orthogonal directions to scaling AI</em></p><p><em>MAKER takes an orthogonal path, scaling reliability instead of model size. The chart compares leading LLMs by their error-free step length versus cost, showing that while single models fail after hundreds of steps, MAKER completes over one million with zero errors.</em></p><h3>A million-step evaluation: Towers of Hanoi, zero errors</h3><p>To test the theory, the team applied MAKER to the Towers of Hanoi puzzle with 20 disks, a version requiring exactly 1,048,575 dependent steps. Every move had to be correct.</p><p>Each microagent was assigned a single move, using the same global strategy but local context. Before scaling up, smaller experiments measured each model’s per-step accuracy and cost to choose the most cost-effective setup. Surprisingly, smaller, non-reasoning models such as gpt-4.1-mini and gpt-oss-20B provided the best reliability-per-dollar.</p><p>Using first-to-ahead-by-3 voting and red-flag filtering for long or misformatted outputs, MAKER successfully executed all 220–1 &gt; steps with zero errors — the first such demonstration of its kind.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FPRiQlXGhke4%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DPRiQlXGhke4&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FPRiQlXGhke4%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/079b72b6bb7c006023191c833ef64d40/href">https://medium.com/media/079b72b6bb7c006023191c833ef64d40/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DFRuEXOh8POzvWhH" /></figure><p><em>Figure 3. Exponential convergence to a zero-error solution.</em></p><p><em>The number of undecided steps falls exponentially across voting rounds, confirming the efficiency of race-to-k voting for local reliability.</em></p><p>These results provide direct empirical evidence for MAKER’s theoretical advantages:</p><ul><li>Error-free execution across one million dependent steps</li><li>Exponential convergence toward perfect consensus</li><li>Practical scaling through extreme decomposition and localized error mitigation</li></ul><p>This marks a clear demonstration of multi-agent advantage (akin to quantum advantage), where a coordinated multi-agent system achieves a result that no single-agent model can.</p><h3>Why this matters and what’s next</h3><p>MAKER reframes the problem of AI reliability as a systems challenge, not a model-size race. The findings show that massively decomposed agentic processes (MDAPs) can scale reliability to levels unreachable by single LLMs.</p><p>Beyond the immediate achievement, this architecture carries broader implications:</p><ul><li>Reliability through structure: small, local gains in accuracy compound into global perfection.</li><li>Transparency and control: each microagent’s output is interpretable and limited in scope.</li><li>Safety and efficiency: smaller models can perform the majority of work, lowering risk and cost.</li></ul><p>MAKER focuses on flawless execution of a known plan. The next step is applying the same principles to creative reasoning — decomposing not just actions, but also idea generation, planning, and verification. This could allow systems that both think and act at organizational scale, with the precision and reliability that real-world applications demand.</p><p>By smashing large problems into minimal, cooperating agents and recombining their work through principled error correction, AI can scale far beyond the limits of single models. MAKER demonstrates that dependable, large-scale intelligence can be achieved with systems that are smaller, safer, and more controllable.</p><p>This architecture opens the door to real-world deployments where reliability, cost, and data sovereignty are critical. Millions of lightweight, locally hosted agents could operate collaboratively across hospitals, factories, or government systems, maintaining precision without exposing sensitive data to massive centralized models. The future of large-scale intelligence may lie not in building ever-bigger models, but in designing smarter, distributed systems that simply do not fail.</p><p><a href="https://arxiv.org/abs/2511.09030"><strong>Read the full paper</strong></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c74b8271f1d5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A New Fine-Tuning Approach for LLMs Using Evolution Strategies]]></title>
            <link>https://medium.com/@evolutionmlmail/a-new-fine-tuning-approach-for-llms-using-evolution-strategies-1ef8e370b636?source=rss-9d954de45252------2</link>
            <guid isPermaLink="false">https://medium.com/p/1ef8e370b636</guid>
            <category><![CDATA[fine-tuning]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[evolution-strategies]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Cognizant AI Lab]]></dc:creator>
            <pubDate>Tue, 03 Feb 2026 17:01:02 GMT</pubDate>
            <atom:updated>2026-02-03T17:01:02.526Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>A new approach from Cognizant AI Lab challenges reinforcement learning as the default for LLM post-training</strong></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F3U4OkP-9EVI%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D3U4OkP-9EVI&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F3U4OkP-9EVI%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/fb0102baded70d353c1d65074a41b9f0/href">https://medium.com/media/fb0102baded70d353c1d65074a41b9f0/href</a></iframe><p>Fine-tuning is essential for every modern large language model. GPT, Claude, LLaMA, DeepSeek, Qwen — all of them are trained first on massive amounts of general data, then fine-tuned to further elicit reasoning capability or become better aligned with human intent.</p><p>Reinforcement learning (RL) is currently the primary choice for such fine-tuning. Techniques like PPO and GRPO dominate the field and have shaped nearly every production-grade LLM in use.</p><p>But RL comes with known limitations. It struggles with sparse, long-horizon rewards. It’s sensitive to model initialization and hyperparameter choices. And without careful constraints, it’s prone to reward hacking, leading to undesirable behaviors.</p><p>We explore a long-standing but largely overlooked alternative to reinforcement learning: evolution strategies (ES). Earlier work from <a href="https://arxiv.org/abs/1703.03864">OpenAI</a> and <a href="https://arxiv.org/abs/1712.06568">Uber</a> demonstrated the promise of ES on control and optimization problems. However, in our paper, we present the first successful use of ES to fine-tune the full parameter set of large language models, including models with billions of parameters. The results show that ES can outperform state-of-the-art RL methods on key dimensions such as sample efficiency, tolerance to long-horizon rewards, robustness to different base LLMs, less tendency to reward hacking, and more stable performance across runs.</p><p>This approach is not just an optimization tweak. It represents a new direction for how we fine-tune LLMs at scale and points toward simpler, more reliable, and more adaptable post-training techniques for future generations of AI systems.</p><h3>A scalable framework for full-parameter fine-tuning using ES</h3><p>Prior work using evolution strategies typically limited the model size (up to millions of parameters) or reduced optimization dimensionality, such as tuning only the output layer or adapters. This new approach instead scales ES to full-parameter fine-tuning of transformers with billions of parameters, without relying on gradients or backpropagation.</p><p>The method uses an algorithmically simplified but engineeringly optimized ES variant. Through memory-efficient and traceable parameter perturbations, it enables distributed evaluations (inferences) while significantly reducing GPU memory usage and communication overhead. By design, this makes ES viable for large-scale LLMs while preserving the simplicity and robustness of gradient-free optimization.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6_Nqi8HHfSTL2WCK" /></figure><p><em>Figure 1: Flow chart of ES fine-tuning.</em></p><p><em>ES explores by directly sampling perturbations (ϵ) in the parameter space (θ). The perturbed models are evaluated (performing inferences) for outcome-only rewards (R), which are used to form the aggregated parameter update. No gradient calculation or process reward is needed.</em></p><h3>Results: ES outperforms RL on key aspects</h3><p>To evaluate this method, we used the Countdown arithmetic reasoning task, a benchmark that involves multi-step symbolic computation. It’s particularly challenging for RL because of its sparse rewards and long-horizon dependencies — conditions that make action-space exploration unstable.</p><p>The ES approach was benchmarked against state-of-the-art RL baselines including PPO and GRPO across a range of model sizes and families, from 0.5 billion to 8 billion parameters. In every case, ES achieved higher accuracy scores and faster convergence. Smaller models, which tend to struggle under RL fine-tuning, showed particularly steady improvements under ES, but the method scaled effectively to larger models as well.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*hJgzl_C3T7CR3DfG" /></figure><p><em>Table 1: Accuracy on the Countdown task across models and fine-tuning methods.</em></p><p><em>ES consistently outperforms PPO and GRPO across all model sizes — from 0.5B to 8B parameters — with no custom hyperparameter tuning or reward shaping.</em></p><h3>A more reliable fine-tuning paradigm</h3><p>One of the most promising aspects of this approach is its reliability. Evolution strategies demonstrate low variance across random seeds, minimal sensitivity to hyperparameters, and no observable reward hacking. The method does not require gradients or the delicate actor-critic architecture typical of RL, yet consistently discovers high-performing policies with fewer environment interactions.</p><p>As models continue to scale, this kind of reliability becomes increasingly valuable. ES reduces sensitivity to initialization and hyperparameter setup, making it easier to apply across diverse use cases without fragile tuning. These properties make it well-suited not only for research but also for robust production deployments.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*cYMpu8XFdPhKg015" /></figure><p><em>Figure 2: Reward vs. KL divergence during fine-tuning.</em></p><p><em>KL divergence here measures how much the fine-tuned model behavior deviates from the base model — larger divergence suggests more drastic behavioral shifts, which can lead to instability or unexpected outputs. The ES models (blue) discover better tradeoffs than GRPO (black), achieving higher reward without needing KL penalties. This suggests ES follows a fundamentally different fine-tuning trajectory.</em></p><h3>The future: Expanding the fine-tuning toolbox</h3><p>This work repositions evolution strategies as a practical and scalable alternative to reinforcement learning for fine-tuning large language models. By removing the need for gradient calculations, ES simplifies the training pipeline while delivering strong results across robustness, efficiency, and stability.</p><p>Looking ahead, ES could be applied to a wide range of post-training challenges. These include reasoning incentivization, exploration-required tasks, safety alignment, and continual learning — domains where RL has traditionally been the default. As models continue to grow in complexity and importance, having more reliable tools for fine-tuning will be essential.</p><p>We are excited about the possibilities this research unlocks. Evolution strategies offer a fundamentally different path forward, and we look forward to exploring how this framework can be extended and applied to new fine-tuning objectives that ultimately lead to next-level intelligence of large models.</p><p><a href="https://github.com/VsonicV/es-fine-tuning-paper"><strong>Access the code in our open-source repo to explore it yourself</strong></a></p><p><a href="https://arxiv.org/abs/2509.24372"><strong>Read the full paper</strong></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1ef8e370b636" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building trustworthy AI together: Reflections from the AI for Good Global Summit 2025]]></title>
            <link>https://medium.com/@evolutionmlmail/building-trustworthy-ai-together-reflections-from-the-ai-for-good-global-summit-2025-6101a3bb6c50?source=rss-9d954de45252------2</link>
            <guid isPermaLink="false">https://medium.com/p/6101a3bb6c50</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-for-good]]></category>
            <category><![CDATA[responsible-ai]]></category>
            <dc:creator><![CDATA[Cognizant AI Lab]]></dc:creator>
            <pubDate>Fri, 25 Jul 2025 19:02:04 GMT</pubDate>
            <atom:updated>2025-07-25T19:02:04.565Z</atom:updated>
            <content:encoded><![CDATA[<p>By <a href="https://www.linkedin.com/in/abanifatemi/">Amir Banifatemi</a></p><p>I recently returned from the <a href="https://www.linkedin.com/showcase/ai-for-good-global-summit/about/">AI for Good Global Summit 2025</a> in Geneva, and as always, I left inspired by the people, the ideas, and the sense of shared responsibility to ensure AI continues to serve all of humanity. As a founder of the conference, I am thrilled by its growth from a few hundred people many years ago to over 12,000 today. It’s great to see so many companies from Google and Microsoft to all the major governments of the world represented, all converged in one place to make better policies, standards and AI innovations for the benefit of the global population.</p><p>This year’s summit felt especially significant. The conversations were less conceptual and more focused on practical implementation — how to design, govern, and scale AI systems that are not only powerful, but also trustworthy, safe, and grounded in real-world needs.</p><p>That shift from principles to practice came through in several of the sessions I had the opportunity to join and contribute to.</p><h3>Strengthening trust through testing and evaluation</h3><p>One of the most meaningful discussions I took part in was the Open Dialogue for Trustworthy AI Testing, which brought together experts from the OECD, Oxford Martin School, UC Berkeley, and others to explore how we build more consistent and transparent approaches to evaluating AI in real-world settings.</p><p>A few challenges stood out. Many regions still lack access to the tools and frameworks needed to test and monitor AI safely. Standards vary widely. And across the board, there’s a tendency to treat testing as a milestone rather than a continuous process.</p><p>This is something I think about often in my work at Cognizant — how we ensure that AI systems, especially those used in high-stakes environments, are not only accurate but reliable, fair, and trustworthy but and resilient over time. Evaluation has to extend beyond performance metrics to include transparency, long-term reliability, and alignment with real-world conditions.</p><p>My key takeaway from the session was that trust in AI starts well before deployment. It begins with clear, shared evaluation frameworks that are tested in context and designed to evolve</p><h3>Securing agentic AI</h3><p>Another recurring thread was the growing security challenges posed by agentic and multi-agent systems. These models introduce a new layer of complexity, especially when it comes to safeguarding data and coordinating behavior across distributed environments. As these technologies mature, the question isn’t just how powerful they are, but how we ensure they remain aligned with human values and governance structures that can keep pace.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/704/1*KlsLIXzGOXuloRf8biC_4Q.jpeg" /></figure><h3>Rethinking the future of work and learning</h3><p>In another session, AI Futures: Reimagining Learning and Work in 2035 and Beyond, I joined a diverse group of educators, technologists, and policy leaders to explore how AI might reshape education and employment over the next decade.</p><p>Through scenario planning and open discussion, we surfaced a shared recognition: AI should be used to augment human capability, not replace it. Whether through personalized learning pathways, flexible credentialing, or support for lifelong education, the goal is to create environments where AI expands access and supports meaningful participation in the future economy.</p><p>What resonated most was the understanding that foresight is no longer a theoretical exercise. If we want AI to support a more equitable future, we have to design for that outcome now with both intention and collaboration.</p><h3>Applying AI to global challenges</h3><p>I was also proud to see our team from Cognizant AI Lab contribute to the summit’s applied mission during the session on AI for Agriculture: Shaping Standards for Smart Food Systems. Our team shared <a href="https://www.cognizant.com/us/en/ai-lab/project-resilience">Project Resilience</a>, an open platform built under the Global Initiative on AI and Data Commons. It brings together technical experts and decision-makers to develop AI tools that support sustainable agriculture, climate adaptation, and public health planning</p><p>As part of the platform, we demoed the <a href="https://project-resilience.github.io/platform/projects/irrigation.html">Irrigation Strategy Optimization</a> project, a tool built on the AquaCrop simulator to help optimize irrigation timing and field strategies. Designed to improve crop yield, reduce water use, and lower management costs, it offers farmers and policymakers practical, data-informed support for more sustainable decisions in the field.</p><p>One stat shared during the session stuck with me: globally, up to 40% of crop yield is lost before harvest, and another 40% after the harvest. That means just 20% of what’s grown actually reaches people. The scale of that inefficiency is staggering, and it makes clear how even small, targeted improvements through AI could have meaningful global impact. Tools like Project Resilience aren’t just technical demonstrations — they’re early examples of how AI can help reduce waste, strengthen food systems, and meet real human needs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/810/1*h3LC6_-cA4wvb865-ay_kg.jpeg" /></figure><p><a href="https://github.com/Project-Resilience/platform?tab=readme-ov-file">Explore the platform today</a>.</p><h3>What’s next</h3><p>Leaving Geneva, what stayed with me wasn’t just the panels or presentations, but the clarity that emerged across so many conversations: advancing AI responsibly isn’t about pushing harder on capability. It’s about slowing down where it matters, asking better questions, and designing with purpose.</p><p>That mindset shapes much of what we focus on at Cognizant. We’re building AI systems that don’t just perform well in isolated tests, but can be trusted in the real world — in the hands of educators, policymakers, health professionals, and everyday users. It means thinking carefully about where these systems intersect with people’s lives, and what it takes to make them reliable, understandable, and aligned with long-term human needs.</p><p>What came through across the summit was a shared understanding that no single framework, company, or institution can get this right alone. Building trustworthy AI will require collaboration, clear standards, and a continued willingness to engage, even when the path forward isn’t fully defined.</p><p>There’s still a long way to go, but the momentum is real. I left with a renewed sense of focus and a reminder that building trustworthy AI is not a one-time effort. It’s ongoing work, and it belongs to all of us.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6101a3bb6c50" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Cognizant AI Lab wins Humies Gold for RHEA at GECCO 2025]]></title>
            <link>https://medium.com/@evolutionmlmail/cognizant-ai-lab-wins-humies-gold-for-rhea-at-gecco-2025-147cd237e07e?source=rss-9d954de45252------2</link>
            <guid isPermaLink="false">https://medium.com/p/147cd237e07e</guid>
            <category><![CDATA[ai-research]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-for-good]]></category>
            <dc:creator><![CDATA[Cognizant AI Lab]]></dc:creator>
            <pubDate>Fri, 18 Jul 2025 22:03:38 GMT</pubDate>
            <atom:updated>2025-07-18T22:08:47.807Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eJjkrlX1WWpUP0pgWNyWxg.png" /></figure><p>How can AI move beyond simply mimicking expert decisions, and instead build on diverse human insight to generate entirely new solutions?</p><p>That question led to the development of <a href="https://arxiv.org/abs/2411.00156">RHEA</a> (Realizing Human Expertise through AI), a framework from Cognizant AI Lab that uses evolutionary methods to combine, refine, and evolve expert-designed solutions for complex, real-world problems. This week, RHEA was awarded Humies (Human-Competitive Awards) Gold at the <a href="https://gecco-2025.sigevo.org/HomePage">2025 Genetic and Evolutionary Computation Conference (GECCO)</a>, one of the field’s most respected honors in evolutionary computation.</p><p>The Humies recognize results that are competitive with or exceed what human experts alone can achieve. RHEA’s win reflects its core insight: that AI can help realize the latent potential of collective human expertise.</p><h3>From pandemic policy to general framework</h3><p>RHEA was first applied to the <a href="https://www.xprize.org/challenge/pandemicresponse">XPRIZE Pandemic Response Challenge</a>, where 102 teams from 23 countries submitted models aimed at balancing public health outcomes with the economic and social costs of COVID-19 interventions. Each model reflected different priorities, assumptions, and methodologies — from statistical models to machine learning, heuristics, and domain knowledge.</p><p>Rather than selecting a single winning approach, RHEA distilled all 169 submitted models into neural representations. Using evolutionary algorithms, it then recombined and evolved these models to search for strategies that outperformed any individual submission, ultimately discovering improved policy strategies that balanced case reduction with intervention cost more effectively than any single model on its own.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ZzAEmlFSWM0ivWlJ" /></figure><p>In many cases, models that had underperformed in isolation contributed critical elements to the final solutions, demonstrating RHEA’s ability to uncover and reuse valuable structure from across the full range of expert input.</p><h3>A moment of thanks</h3><p>We’re honored to receive this award, and see it as a meaningful recognition of the research we’re doing at Cognizant AI Lab: advancing AI systems that maximize human potential and support real-world decision-making.</p><p>The paper was authored by <a href="https://scholar.google.com/citations?user=RT_LBQ4AAAAJ">Elliot Meyerson</a>, <a href="https://scholar.google.com/citations?user=1Wg6HFYAAAAJ">Olivier Francon</a>, <a href="https://arxiv.org/search/cs?searchtype=author&amp;query=Sargent,+D">Darren Sargent</a>, <a href="https://scholar.google.com/citations?user=letSloQAAAAJ&amp;hl=en">Babak Hodjat</a>, and <a href="https://www.cs.utexas.edu/~risto/">Risto Miikkulainen</a>. We’re grateful to the GECCO and Humies committees, to the broader evolutionary AI community, and to all collaborators who helped make this work possible.</p><p>We’re excited to continue developing RHEA and exploring its applications in new domains where diverse expertise meets high-impact decisions.</p><p>Read the <a href="https://arxiv.org/abs/2411.00156">full paper</a>, and learn more by exploring our paper <a href="https://medium.com/@evolutionmlmail/bridging-human-expertise-and-ai-to-tackle-global-challenges-209950d244d9">break-down blog</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=147cd237e07e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>