<?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 ComfyAI on Medium]]></title>
        <description><![CDATA[Stories by ComfyAI on Medium]]></description>
        <link>https://medium.com/@ai4medical?source=rss-4f0729bdce69------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*kPX6KUhF4H3F80tGvyL0bA.jpeg</url>
            <title>Stories by ComfyAI on Medium</title>
            <link>https://medium.com/@ai4medical?source=rss-4f0729bdce69------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 09:56:23 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ai4medical/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[Following Instructions Isn’t the Same as Doing Science: What NIBench Reveals About AI Agents in…]]></title>
            <link>https://medium.com/@ai4medical/following-instructions-isnt-the-same-as-doing-science-what-nibench-reveals-about-ai-agents-in-0b9dd0f0b16c?source=rss-4f0729bdce69------2</link>
            <guid isPermaLink="false">https://medium.com/p/0b9dd0f0b16c</guid>
            <category><![CDATA[benchmark]]></category>
            <category><![CDATA[nibench]]></category>
            <category><![CDATA[neuroimaging]]></category>
            <category><![CDATA[arxiv]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[ComfyAI]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 01:38:28 GMT</pubDate>
            <atom:updated>2026-11-17T01:38:28.141Z</atom:updated>
            <content:encoded><![CDATA[<h3>Following Instructions Isn’t the Same as Doing Science: What NIBench Reveals About AI Agents in Neuroimaging</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*v9XBG4Tjltiyzw4wmuhZBA.png" /></figure><p>A new paper introduces <strong>NIBench</strong>, a benchmark designed to test how well today’s AI agents actually perform in neuroimaging. It evaluates 19 models, 11 closed-source and 8 open-source, across three parts. Together, they show a clear gap between what agents can execute and what they can judge.</p><h3>Part 1: Can agents work without a manual?</h3><p>The first part consists of 56 real-world neuroimaging tasks. They span structural, diffusion, and functional MRI as well as PET, and use seven specialized toolchains.</p><p>The key design choice is that every task is run twice. In the first run, the model gets an operating manual that spells out which tools to use and which parameters to set. In the second, it gets only the task goal and has to build the workflow itself.</p><p>With the manual in hand, the strongest model, Claude Opus 5.5, reached an accuracy of 0.994. Without it, that dropped to 0.949.</p><p>Open-source models proved far more fragile. Closed-source models lost only 0.09 on average when the manual was removed, while open-source models lost 0.38. The drop was steepest on pipeline tasks that chain multiple steps together, where the top three open-source models fell by 0.61 on average.</p><p>The takeaway: an agent that can complete a task by following instructions is a long way from an agent that can complete it on its own.</p><h3>Part 2: Can agents learn from experience?</h3><p>The second part asks whether an agent’s operational experience can be accumulated and reused. The approach distills each full execution trace into editable text files stored as skills, for example one skill per toolchain. In effect, this is memory evolution in plain text. The researchers then measured performance on both seen and unseen tasks.</p><p>Completion rates on seen tasks improved across the board. On unseen tasks, however, the gains were inconsistent. Some models improved, while others actually got worse, possibly because they overfit to the tasks they had already encountered.</p><p>Skills also turned out to be transferable across models. Giving a strong model’s skills to a weaker model helped the weaker model complete tasks in most cases, but not all. One plausible explanation is that skills distilled by a strong model don’t fully match the kinds of mistakes a weaker model tends to make.</p><h3>Part 3: Can agents make scientific judgments?</h3><p>The third part moves up a level. It tests whether an AI can make sound scientific judgments about the results it computes itself.</p><p>The agents received unfiltered Alzheimer’s disease data from ADNI, covering more than 5,000 participants. Their job was to evaluate 36 scientific claims derived from three published papers and label each one as supported, refuted, or inconclusive.</p><p>Across 18 data-grounded runs, the agents’ execution was solid, and the code they wrote was reproducible. That fits well with what agents are already good at today.</p><p>Judgment was another story. The two frontier models agreed with the expert reference answers on only 73.1% of their calls. Most disagreements came down to one question: is the evidence actually strong enough to draw a conclusion?</p><p>The two models failed in opposite directions. Claude tended to overstate claims that should have stayed uncertain, overreaching 18 times. GPT-6-astra did the reverse, often labeling claims as inconclusive even when the evidence was sufficient, erring on the side of caution 22 times.</p><p>Their overall accuracy was similar, but it masked two opposite failure modes.</p><h3>The real bottleneck</h3><p>For AI-driven autonomous research, the hard problem right now isn’t execution. It’s calibration: the sense of proportion and judgment an agent needs when it faces uncertainty.</p><p>| Find papers faster on <a href="https://arxivsub.comfyai.app/">arXivSub</a> with AI summary (CVPR、ICCV、ECCV、ICML、ICLR、NeurIPS、AAAI、MICCAI、CoRL、RSS、ACL、EMNLP、ICRA、IROS、IJCAI)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AETHZbd3OPIYdVsGN6BgYA.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0b9dd0f0b16c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Giving fMRI a Language: Teaching an LLM to Read the Brain]]></title>
            <link>https://medium.com/@ai4medical/giving-fmri-a-language-teaching-an-llm-to-read-the-brain-73cb5353838b?source=rss-4f0729bdce69------2</link>
            <guid isPermaLink="false">https://medium.com/p/73cb5353838b</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[fmri]]></category>
            <category><![CDATA[mri]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[cvpr]]></category>
            <dc:creator><![CDATA[ComfyAI]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 01:34:15 GMT</pubDate>
            <atom:updated>2026-11-16T01:34:15.295Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kihavA_9W7-wdh8iJ-TD0Q.png" /></figure><p>Multimodal large models can already handle images, audio, and video in a single framework. Brain imaging, however, has been left behind.</p><p>Existing fMRI foundation models, such as BrainLM and Brain-JEPA, rely only on self-supervision within the neural signal itself, through objectives like masked reconstruction and contrastive learning. They have no interface to language. They also provide only a single set of representations. Each new downstream task needs its own task-specific training or fine-tuning on top of those representations.</p><h3>The Core Idea: Captions Built From Neuroimaging Statistics</h3><p>The authors want to connect fMRI to a pretrained language model so that brain activity can be described and predicted directly. The difficulty is that fMRI has no naturally paired text. Images come with captions, but brain scans don’t.</p><p>Their solution is to build the descriptions from statistics that neuroimaging already produces.</p><p>They draw on four types of features: functional connectivity, functional gradients, graph metrics, and ICA components. Together these yield 23 descriptors. Each descriptor is converted into a standardized text description using a fixed template, so every subject gets 23 descriptions. These serve as the “captions” for that subject’s fMRI.</p><h3>A Three-Stage Pipeline</h3><p><strong>Stage 1: an fMRI tokenizer.</strong> A ViT encoder with vector quantization splits a segment of fMRI into discrete tokens. Two mechanisms align these brain tokens to the language model’s text embedding space. The first is a SigLIP-style contrastive loss on fMRI–text pairs. The second is a domain discriminator with a gradient reversal layer, following the DANN approach.</p><p><strong>Stage 2: fine-tuning the LLM.</strong> The model learns to handle brain tokens and text together, through two objectives. One is temporal next-step prediction: the tokens for 450 brain regions at time <em>t</em> predict the tokens at time <em>t+1</em>. The other is a language task, in which the model reads a brain signal and writes a description of it.</p><p><strong>Stage 3: multi-task, multi-paradigm instruction tuning.</strong> This stage covers single-question/single-answer, multi-question/multi-answer, and open-ended description formats. The tasks include sex, age, fluid intelligence (a cognitive score), Alzheimer’s disease, ADHD, and autism.</p><h3>Data and Evaluation</h3><p>The model was pretrained on more than 50,000 fMRI scans from UK Biobank and ABCD. It was evaluated on seven datasets: held-out UKB and ABCD data, plus HCP, HCP-Aging, ADNI4, ADHD200, and ABIDE2.</p><h3>Results</h3><p>On classification tasks, the version built on Qwen performs best on most tasks. For example, it reaches 94.5% accuracy on sex classification in UKB.</p><p>The gains are also clear on cross-cohort clinical tasks, where earlier models performed close to chance. ADHD accuracy rises from Brain-JEPA’s 72% to 78.6%. Autism accuracy rises from about 65% to 76.6%.</p><p>| Find papers faster on <a href="https://arxivsub.comfyai.app/">arXivSub</a> with AI summary (CVPR、ICCV、ECCV、ICML、ICLR、NeurIPS、AAAI、MICCAI、CoRL、RSS、ACL、EMNLP、ICRA、IROS、IJCAI)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3Ezsh1azD3s582HKZjIniQ.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=73cb5353838b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why RL Lifts Qwen but Leaves Llama Flat: A Closer Look at “RL-Friendliness”]]></title>
            <link>https://medium.com/@ai4medical/why-rl-lifts-qwen-but-leaves-llama-flat-a-closer-look-at-rl-friendliness-6dc4f830feb1?source=rss-4f0729bdce69------2</link>
            <guid isPermaLink="false">https://medium.com/p/6dc4f830feb1</guid>
            <category><![CDATA[acl]]></category>
            <category><![CDATA[lrs]]></category>
            <category><![CDATA[post-training]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[ComfyAI]]></dc:creator>
            <pubDate>Sat, 14 Nov 2026 01:34:44 GMT</pubDate>
            <atom:updated>2026-11-14T01:34:44.055Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lS1LmkiaWd3BQ_l9mObIkQ.png" /></figure><p>Train models with reinforcement learning from verifiable rewards, using algorithms like GRPO or DAPO, and a pattern shows up. Qwen improves substantially. Llama-family models barely move.</p><p>To dig into why, the authors define a model’s <strong>RL-Friendliness</strong> as how much it benefits from RL, measured as the difference in Pass@1 before and after training. The rough ranking is Qwen &gt; OctoThinker &gt; Llama. Training data, hyperparameters, and algorithm are held fixed across all models, so what’s being compared is a property of the models themselves.</p><h3>Same problems solved, different confidence</h3><p>The first step is a controlled comparison between Qwen2.5–7B and OctoThinker-8B. OctoThinker is derived from Llama through mid-training, a stage in which it was fed large amounts of high-quality math and reasoning data.</p><p>The two models largely solve the same set of problems. On the problems both can solve, however, Qwen’s pass rate is noticeably higher. Qwen is better at concentrating enough probability mass on the correct answer.</p><h3>Measuring how well correct and incorrect answers separate</h3><p>The core analysis uses <strong>length-normalized sequence probability</strong>, the geometric mean of per-token probabilities, as a confidence score for each response. The authors then compare the distribution of this score across correct and incorrect responses.</p><p>For each problem, the model samples many responses (K = 256 on AIME in the paper). Qwen’s distributions are compact within each group, meaning among correct responses and among incorrect ones, and the two groups sit far apart. For Llama and OctoThinker, the two distributions overlap heavily. Incorrect responses often have sequence probabilities, and therefore confidence, no lower than correct ones.</p><p>To quantify this, the authors borrow the <strong>silhouette coefficient S</strong> from clustering and compute one value per problem. A value close to 1 means the groups are both compact and well separated. Problems with positive S have far higher pass rates than problems with negative S:</p><ul><li>Qwen: 52.3% vs. 6.5%</li><li>OctoThinker: 40.8% vs. 3.1%</li></ul><h3>What S means semantically</h3><p>To see what S corresponds to, the authors used an LLM as a judge to sort errors into three categories: logical errors, calculation errors, and formatting errors.</p><p>Logical errors make up a clearly smaller share of Qwen’s mistakes, so its error profile is healthier. Its logical errors are also heavily concentrated in responses with negative S. For Llama, 62.2% of logical errors also fall where S &lt; 0. When the distributions fail to separate (low S), the model has usually gotten something genuinely wrong.</p><p>The less RL-friendly models’ correct solutions also look more varied. On inspection, though, many of them were reached through lucky guesses and hallucinated steps, and these also correspond to low S.</p><h3>Turning the diagnosis into a training signal</h3><p>To improve RL, the authors build on DAPO and reweight each problem by its S, using a weight of <strong>exp(−β·S)</strong> with β = 0.2. Easy problems, whose distributions are already well separated, get small weights. Problems with blurry distributions get large weights. This pushes the gradient toward the model’s weakest points.</p><h3>Results</h3><p>All three models improve across six math benchmarks:</p><ul><li>OctoThinker: AIME24 from 4.9 to 8.2, AIME25 from 2.1 to 5.0</li><li>Qwen: AIME24 from 12.2 to 18.1</li></ul><p>During training, the correlation between S and Pass@1 also reaches 0.815.</p><p>| Find papers faster on <a href="https://arxivsub.comfyai.app/">arXivSub</a> with AI summary (CVPR、ICCV、ECCV、ICML、ICLR、NeurIPS、AAAI、MICCAI、CoRL、RSS、ACL、EMNLP、ICRA、IROS、IJCAI)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qeAc9b-a8O_t6imk-Ndgig.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6dc4f830feb1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Weight-Space Programming: When an LLM Reads and Writes Neural Network Weights Directly]]></title>
            <link>https://medium.com/@ai4medical/weight-space-programming-when-an-llm-reads-and-writes-neural-network-weights-directly-5e88686dd964?source=rss-4f0729bdce69------2</link>
            <guid isPermaLink="false">https://medium.com/p/5e88686dd964</guid>
            <category><![CDATA[mit]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[gpt-6-astra]]></category>
            <category><![CDATA[weight-space-programming]]></category>
            <dc:creator><![CDATA[ComfyAI]]></dc:creator>
            <pubDate>Fri, 13 Nov 2026 01:34:12 GMT</pubDate>
            <atom:updated>2026-11-13T01:34:12.124Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kBtD3PDOlBDvXNJl_3KWyQ.png" /></figure><p><em>A new MIT CSAIL blog post argues that a frontier model can understand, build, and repair neural networks without gradient descent, using only natural-language reasoning.</em></p><p>A recent blog post from MIT CSAIL makes a striking claim. Frontier large language models (in this case, GPT-6 Astra) can read and write the weights of a neural network directly. They don’t use gradient descent or an optimizer. They reason in natural language. The authors call this capability <strong>Weight-Space Programming</strong>.</p><h3>Reading Weights</h3><p>The first set of experiments tests whether Astra can tell what a network does just by looking at its parameters.</p><p>The researchers used PPO to train four small networks with identical architectures, each with 17,253 parameters, to play four different grid games. They then gave Astra nothing but the raw weights, formatted as text. They didn’t tell it which games were possible. Astra identified all four correctly. It could also point to specific structure, such as noting that in the Breakout network, the weights corresponding to the paddle sit in the bottom row of the grid.</p><p>A second experiment went further. The team took a character-level language model and removed both its input and output layers, leaving only the four transformer blocks in the middle. Astra could still tell whether the model had been trained on Python, romanized Japanese, or English. It also found every induction head, the components responsible for copying earlier parts of the context, and its findings matched the empirical measurements exactly.</p><h3>Writing Weights</h3><p>The second set of experiments reverses the direction: can Astra build a working network from scratch?</p><p>Astra was given the rules of a game and a set of all-zero initial weights. It was not allowed to train and not allowed to run the environment. Working purely in text, it wrote out networks that play three MinAtar games: Breakout, Seaquest, and Asterix. In all three, its hand-written networks beat PPO trained for 100 million steps.</p><p>Breakout was the most dramatic result. Astra’s network scored 517, while PPO, even after a full 500 million steps, reached only 25.</p><p>Its method was to first work out what each layer should compute. For example, the first layer predicts where the ball will land, and the second moves the paddle there. Only after that did it translate the plan into actual weights.</p><h3>Repairing Models</h3><p>The most practical application is fixing existing models.</p><p>Qwen3.5–0.8B tends to fall into thinking loops on math problems: it repeats itself over and over and is slow to produce an answer. Astra traced this failure to a single direction in the MLPs of layers 15 and 19, then subtracted that direction out. The edit touched only 131,072 parameters, just 0.016% of the model.</p><p>The results after the fix:</p><ul><li><strong>MATH-500:</strong> 7.9% → 48.1%</li><li><strong>GSM8K:</strong> 13.8% → 67.1%</li><li><strong>IFEval</strong> (instruction following, which the fix never targeted): 21% → 62.3%</li></ul><h3>Where It Falls Short</h3><p>The approach has limits. When the same strategy was applied to Gemma 3 270M, a model that rarely gets stuck in loops, it didn’t work. Gemma’s mistakes are genuine calculation and reasoning errors, and there is no single direction you can switch off to make that kind of problem go away.</p><p>| Find papers faster on <a href="https://arxivsub.comfyai.app/">arXivSub</a> with AI summary (CVPR、ICCV、ECCV、ICML、ICLR、NeurIPS、AAAI、MICCAI、CoRL、RSS、ACL、EMNLP、ICRA、IROS、IJCAI)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5e88686dd964" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Unlearning Isn’t Enough: Why “Forgotten” LLMs Still Don’t Know What They Don’t Know]]></title>
            <link>https://medium.com/@ai4medical/unlearning-isnt-enough-why-forgotten-llms-still-don-t-know-what-they-don-t-know-1baf6cf91f6a?source=rss-4f0729bdce69------2</link>
            <guid isPermaLink="false">https://medium.com/p/1baf6cf91f6a</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[acl]]></category>
            <category><![CDATA[unlearning]]></category>
            <dc:creator><![CDATA[ComfyAI]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 01:27:22 GMT</pubDate>
            <atom:updated>2026-11-12T01:27:22.597Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OOBlngNLm0njGzJE8i-shg.png" /></figure><p>Large language model unlearning aims to erase harmful or sensitive training data from a model without retraining it from scratch, while doing as little damage as possible to its general capabilities. Evaluations of unlearning methods have traditionally focused on two questions. Was the target knowledge actually removed? And does that removal hold up against attacks such as rephrased questions, jailbreaks, and further fine-tuning?</p><p>This paper argues that a third dimension has been missing: <strong>honesty</strong>. The authors break it into two parts. The first is <strong>self-awareness</strong>: the model should recognize when it doesn’t know something and be willing to simply say “I don’t know.” The second is <strong>consistency</strong>: when the same question is asked in different ways, or followed up on, the model’s answers should be consistent and stable.</p><p>Using the WMDP-Bio biohazard dataset, the authors evaluated nine mainstream unlearning methods on Zephyr-7b and Llama3–8b. They grouped the methods into three categories. Their conclusion is blunt: <strong>none of the current methods pass the honesty test.</strong></p><h3>Category 1: Refusal Fine-Tuning Only Looks Honest</h3><p>The first category trains the model directly to say “I don’t know.” This is the refusal-style fine-tuning approach, IDK_AP. Its refusal rate looks impressively high, but the authors show that this is superficial and even deceptive. Ask about the same piece of knowledge in a different way, and the model still answers correctly. The knowledge was never deleted. The model simply learned to say “I don’t know” when it sees certain trigger patterns.</p><h3>Category 2: Gradient Ascent Methods Are Dodging, Not Knowing</h3><p>The second category is the gradient ascent family, including GradDiff and NPO. On multiple-choice questions, these methods pick the “I don’t know” option at a high rate. This is also an illusion. If you replace that option with an irrelevant sentence, the model still picks it.</p><p>The reason is that these methods push the first-token distribution to an extremely sharp peak. In essence, the model is just avoiding the other options. It doesn’t actually know that it doesn’t know.</p><p>On open-ended question answering, these methods either hallucinate or produce repetitive gibberish tokens. Even when strongly prompted to refuse, their refusal rate is close to zero.</p><h3>Category 3: Feature Randomization Forgets Well but Never Admits It</h3><p>The third category is feature randomization, best represented by RMU. RMU is the strongest of the three at both erasing knowledge and preserving the model’s original capabilities. Yet it almost never admits that it doesn’t know something. Its refusal rate is just 1.36%.</p><p>The paper gives a striking example. Rofecoxib is a painkiller that was withdrawn from the market because of cardiovascular risks. After it was unlearned, RMU described it as an energy-boosting health supplement. In effect, the model could mislead users into taking a withdrawn drug. RMU’s answer consistency on multiple-choice questions is also fairly weak.</p><h3>The Fix: ReVa</h3><p>To address these problems, the authors propose <strong>ReVa</strong>. It starts from a model that has already been unlearned with RMU. From around 20 samples where the model refused to answer, ReVa extracts a <strong>refusal direction vector</strong>. It then aligns the residual representations of forget-set inputs toward that direction. In practice, it uses the vector as a supervision target and performs a gradient-descent update on the weights. Only the down-projection parameters of the MLP in a middle layer, around layer 18, are tuned.</p><p>The results are strong. ReVa’s refusal rate reaches 60.86%, and it still holds at 45% after two rounds of conversation, which is nearly double the next-best method. Its outputs are more stable, and its general capabilities on the retain set are unaffected.</p><p>| Find papers faster on <a href="https://arxivsub.comfyai.app/">arXivSub</a> with AI summary (CVPR、ICCV、ECCV、ICML、ICLR、NeurIPS、AAAI、MICCAI、CoRL、RSS、ACL、EMNLP、ICRA、IROS、IJCAI)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iGqKbh5-Ors-Ej-OLIG58A.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1baf6cf91f6a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why AI Writing Is Getting Harder to Read]]></title>
            <link>https://medium.com/@ai4medical/why-ai-writing-is-getting-harder-to-read-52fd18aea375?source=rss-4f0729bdce69------2</link>
            <guid isPermaLink="false">https://medium.com/p/52fd18aea375</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai-writing]]></category>
            <category><![CDATA[language]]></category>
            <category><![CDATA[arxiv]]></category>
            <dc:creator><![CDATA[ComfyAI]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 01:36:26 GMT</pubDate>
            <atom:updated>2026-11-11T01:36:26.120Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8cE3CtXDPv0P4sqiZe1meA.png" /></figure><p><em>When models are trained to be right rather than understood, readability is the first thing to go.</em></p><p>If you’ve felt that AI-generated text has become more tiring to read over the past couple of years, you’re not imagining it.</p><p>A recent study tackles this question by starting from something psycholinguists settled long ago: a set of well-documented “difficulty features” that make sentences harder to process. These include object relative clauses with the “that” dropped, long syntactic dependencies, rare word combinations (high n-gram perplexity), denominal verbs where a noun is pressed into service as a verb (think “calendar” used as a verb), inanimate subjects paired with verbs that normally require a living agent (“is_file() swallows that”), and long noun compounds stacked one after another.</p><p>None of these structures is grammatically wrong. But each one adds to the cognitive load a reader carries while working through a sentence.</p><p>The authors collected text from 14 frontier and near-frontier models across three settings: chat, coding, and argumentative essays. They then compared it against human corpora, including movie subtitles, StackExchange, student essays, and Wikipedia.</p><p>A straight comparison wasn’t enough, though. Text can be hard to read simply because the content is hard. To separate the two, the authors used a clever trick: they translated each model output from English into Japanese and then back into English. The meaning stays largely intact, but the phrasing gets smoothed into more common forms. If a difficulty feature becomes less frequent after this round trip, that tells you the model chose it. The difficulty wasn’t inherent to the content.</p><p>The results were clear. Frontier models do overuse these hard-to-read structures. They use them more than their back-translated versions do, and for inanimate subjects and rare word combinations, they use them more than humans do. The effect also tracks the intensity of post-training. Moving through the Olmo family from base to instruct to the reasoning-focused Think version, the difficulty features rise at every step. There is one notable exception: Claude Opus 5.5 is easier to read than its predecessor, Opus 5, on almost every measure, 19 out of 21.</p><p>The second experiment went looking for the mechanism. The authors trained Qwen2.5–1.5B with GRPO reinforcement learning on a task that turns structured event descriptions into English sentences. The same meaning can often be expressed in a shorter or a longer valid form. “He sent us a photo” is 6 tokens; “He sent a photo to us” is 7. During training, they added a length penalty.</p><p>The model quickly swung to choosing the shorter form almost every time. It even carried this preference over to structures it had never seen in training. Without a KL constraint, it picked the shorter sentence more than 98% of the time.</p><p>From this, the authors offer an explanation. Human language is optimized for successful communication. It takes into account the listener’s limited memory and the effort of understanding a sentence word by word, so phrasings that are too hard to follow get weeded out naturally. But models are increasingly trained on verifiable rewards, such as math and code, where no human sits in the optimization loop. All the model needs to do is get the answer right. So it drifts toward writing that packs in more information but is harder to read.</p><p>| Find papers faster on <a href="https://arxivsub.comfyai.app/">arXivSub</a> with AI summary (CVPR、ICCV、ECCV、ICML、ICLR、NeurIPS、AAAI、MICCAI、CoRL、RSS、ACL、EMNLP、ICRA、IROS、IJCAI)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*U5HrR25zjO11XClAWNRbEw.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=52fd18aea375" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Diagnosis to Training: Using Sparse Autoencoders to Pick Better Fine-Tuning Data]]></title>
            <link>https://medium.com/@ai4medical/from-diagnosis-to-training-using-sparse-autoencoders-to-pick-better-fine-tuning-data-223294211ccb?source=rss-4f0729bdce69------2</link>
            <guid isPermaLink="false">https://medium.com/p/223294211ccb</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[acl]]></category>
            <category><![CDATA[sparse-autoencoder]]></category>
            <category><![CDATA[sft]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[ComfyAI]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 01:46:29 GMT</pubDate>
            <atom:updated>2026-11-10T01:46:29.163Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kXh_89dOw6tW-2DNGVmMzg.png" /></figure><p>Interpretability tools like sparse autoencoders (SAEs), which map a model’s internal activations onto features with clear semantic meaning, have reached the point where we can partially read what’s going on inside a large language model. But most of these findings stop at diagnosis. They tell us what the model is doing, and are rarely used to make the model better.</p><p>This paper sets out to change that by putting interpretability to work in model training.</p><p>The core idea is simple. If we can see which internal features actually matter for a given task, why not deliberately choose the training data that activates those features most strongly? The authors call their method IGDS, and it works in two stages.</p><h3>Stage One: Finding the Features That Matter</h3><p>The first step is identifying task features. The authors run the model on a small batch of task-relevant samples and look at which SAE features fire almost every time at key token positions, for example features that appear in more than 80% of cases. This gives a rough list of candidates.</p><p>Co-occurrence only shows correlation, though, not causation. So the second step is an intervention test. For each candidate feature, they take its direction (the feature’s activation value multiplied by its weight vector in the SAE decoder) and add it directly into the model’s residual stream, artificially amplifying it. Then they check whether task performance improves. Only features that actually boost performance when amplified are kept, and by default just the single most impactful one is used.</p><h3>Stage Two: Scoring the Data</h3><p>With the task features in hand, every example in the candidate pool gets scored. The activation values of the task features are summed into what the authors call a “feature resonance score.” A higher score means the example does a better job of lighting up the internal mechanisms that genuinely drive performance. The data is ranked by this score, and only the top half is used for fine-tuning.</p><h3>Results</h3><p>The experiments cover three models (Gemma-2, LLaMA-3.1, and Qwen3) across three task types: mathematical reasoning, summarization, and translation. Baselines include quality-based selection (IFD, Loss), diversity-based selection (ZIP), and random selection, with full-dataset fine-tuning as a reference point.</p><p>Across every model and task, IGDS using only half the data outperformed fine-tuning on the full dataset, and beat the other selection methods as well.</p><p>The standout result was Gemma-2–2B on math, where pass@8 rose from 32.2 with full data to 37.8. The size of the gains varied a lot, however. On Qwen3, most improvements were only a fraction of a percentage point.</p><h3>Sanity Checks</h3><p>The authors ran some additional checks. First, amplifying the learned features made the corresponding words appear more often in the model’s output. Amplifying the translation feature, for instance, led to a clear increase in words like “english” and “chinese.” Second, when they searched for math features using three different datasets (GSM8K, Math500, and OpenR1), the top-ranked feature was the same every time.</p><h3>The Ceiling</h3><p>The method’s upper limit is set by the SAE itself. Qwen3 had no ready-made, high-quality SAE, so the authors had to train their own, and its training data didn’t cover everything. The gains on Qwen3 were noticeably smaller as a result, and the authors list this as a limitation.</p><p>| Find papers faster on <a href="https://arxivsub.comfyai.app/">arXivSub</a> with AI summary (CVPR、ICCV、ECCV、ICML、ICLR、NeurIPS、AAAI、MICCAI、CoRL、RSS、ACL、EMNLP、ICRA、IROS、IJCAI)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QAIuDZX_8lvOB2X7AOCIeg.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=223294211ccb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Amyloid Isn’t the Clock: Which Biomarker Actually Tracks Cognitive Decline in Alzheimer’s?]]></title>
            <link>https://medium.com/@ai4medical/amyloid-isnt-the-clock-which-biomarker-actually-tracks-cognitive-decline-in-alzheimer-s-a463e402e007?source=rss-4f0729bdce69------2</link>
            <guid isPermaLink="false">https://medium.com/p/a463e402e007</guid>
            <category><![CDATA[biomarker]]></category>
            <category><![CDATA[cognitive-science]]></category>
            <category><![CDATA[alzheimers-disease]]></category>
            <category><![CDATA[clinical-research]]></category>
            <category><![CDATA[longitudinal]]></category>
            <dc:creator><![CDATA[ComfyAI]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 02:12:47 GMT</pubDate>
            <atom:updated>2026-11-09T02:12:47.284Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eDrSo37HtxxKMIQQcrk-MQ.png" /></figure><h3>A five-year, two-cohort study compares how fast A/T/N biomarkers change, and why amyloid PET falls short as a measure of treatment response</h3><p>With anti-amyloid drugs like lecanemab and donanemab now approved and in clinical use, Alzheimer’s care faces a practical question. Once a patient is on treatment, how do we know whether their disease is still progressing?</p><p>Clinicians need an objective marker they can follow over the long term. Drug developers need something too: a surrogate endpoint that moves faster than the slow decline in cognitive test scores, which currently anchors most trials.</p><p>So among the A/T/N biomarkers (amyloid, tau, and neurodegeneration), which one’s change over time best tracks cognitive decline?</p><p>That question is harder to answer than it sounds. Most earlier work relied on cross-sectional data, a single snapshot per person. This study takes a different approach and directly compares how quickly each biomarker changes over time.</p><h3>The setup</h3><p>The authors drew on two independent cohorts: ADNI (141 participants) and A4/LEARN (151 participants), each followed for about five years on average. Four biomarkers were tracked in parallel:</p><ul><li><strong>Amyloid PET</strong>, measured in Centiloids</li><li><strong>Tau PET</strong></li><li><strong>Plasma p-tau217</strong></li><li><strong>Cortical thickness</strong> in AD-signature regions, measured by MRI</li></ul><p>The analysis had two steps. First, linear mixed models estimated each person’s annual rate of change for each biomarker. Then those rates were regressed onto rates of cognitive change. Cognition was measured with MMSE, ADAS13, and CDR-SB in ADNI, and with MMSE and PACC in A4/LEARN. The models controlled for age, sex, education, baseline cognition, and other factors, and bootstrapping was used to estimate confidence intervals.</p><h3>What they found</h3><p>The results were strikingly consistent across both cohorts. Changes in tau PET, p-tau217, and cortical thickness all reliably tracked cognitive decline.</p><p>Amyloid PET was the exception. Its change over time did not track cognition, whether measured globally or by region. Take Centiloid change against MMSE change: p = 0.590 in ADNI and p = 0.425 in A4/LEARN, with confidence intervals crossing zero in both.</p><p>Among the markers that did track decline, cortical thickness was often the strongest.</p><h3>Why amyloid falls out</h3><p>Why does amyloid, of all things, fail to track decline? The authors’ explanation comes down to timing.</p><p>Aβ deposition begins very early in the disease. Once it crosses the positivity threshold, it gradually plateaus. Its rate of change therefore becomes decoupled from the clinical deterioration that follows. Downstream processes are different. Tau pathology and neurodegeneration keep changing throughout the disease, so they correlate much more closely with cognitive decline.</p><p>p-tau217 sits in the middle of this cascade. It reflects tau pathology driven by Aβ, which places it as an intermediate link between amyloid upstream and neurodegeneration downstream.</p><h3>What this means for monitoring and trials</h3><p>The conclusion is a split verdict on amyloid PET. It remains indispensable for diagnosis, for screening trial participants, and for confirming that a drug has hit its target. But it is not well suited to serve as a surrogate endpoint for efficacy.</p><p>The alternatives each come with a catch. Tau PET does track decline, but it is expensive. Cortical thickness has a more subtle problem. Anti-amyloid therapy itself causes brain volume loss, sometimes called pseudo-atrophy, which can be hard to tell apart from genuine neurodegeneration.</p><p>That leaves plasma p-tau217 as the most practical option for ongoing monitoring. It is inexpensive, minimally invasive, specific to Alzheimer’s, and easy to repeat with a simple blood draw.</p><p>| Find papers faster on <a href="https://arxivsub.comfyai.app/">arXivSub</a> with AI summary (CVPR、ICCV、ECCV、ICML、ICLR、NeurIPS、AAAI、MICCAI、CoRL、RSS、ACL、EMNLP、ICRA、IROS、IJCAI)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a463e402e007" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When the Chart Lies, AI Just Guesses]]></title>
            <link>https://medium.com/@ai4medical/when-the-chart-lies-ai-just-guesses-69d6794f4c7f?source=rss-4f0729bdce69------2</link>
            <guid isPermaLink="false">https://medium.com/p/69d6794f4c7f</guid>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[vlm]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[acl]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[ComfyAI]]></dc:creator>
            <pubDate>Sun, 08 Nov 2026 01:27:38 GMT</pubDate>
            <atom:updated>2026-11-08T01:27:38.558Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dj0mHLRjzMmhwqL5w9wSUw.png" /></figure><p>Multimodal large language models have gotten remarkably good at reading charts. But what happens when the chart itself is misleading? Can we still trust them?</p><p>By “misleading charts,” I mean charts where the data is fine but the presentation isn’t. Think of a truncated y-axis, an inverted axis, a dual-axis chart that overlays two unrelated lines, or 3D effects that distort the proportions of bars. If you look only at the picture, it’s easy to come away with the wrong conclusion.</p><p>The authors of this study tested 19 multimodal models. These included open-source models like LLaVA, Qwen2-VL, and InternVL2.5, commercial models like GPT-4o, Gemini, and Claude 3.5, and several models built specifically for chart understanding. They covered 17 different misleading techniques. The setup was simple: ask each model questions with known correct answers, once on a misleading chart and once on a normal version, and compare accuracy.</p><p>The results are sobering. On misleading charts, the models averaged 26.4% accuracy. The random-guess baseline was 25.6%. That’s almost no difference. Once a chart has been manipulated, the models are essentially guessing. Compared with normal charts, accuracy dropped by an average of 23.8 percentage points, and by as much as 34.8 points.</p><p>For comparison, humans get fooled by these charts too, but they still score between 39% and 49%, well above chance.</p><p>What’s more interesting is that humans and models fail in different places. Take area encoding, where the size of a shape’s area represents a value instead of its length or height. It misleads because models tend to judge size by a single linear dimension, such as a bubble’s diameter. On this technique, models got almost everything wrong (7.4%), while humans reached 91.5%.</p><p>Dual-axis charts work the other way. These charts have two y-axes, one on the left and one on the right, each with its own scale. They were very hard for humans (16.1%), while models did somewhat better (35.5%). So it isn’t accurate to say that models get fooled the way people do. Their weak spots don’t line up with ours.</p><p>The authors also found that high scores on ChartQA, the standard chart-understanding benchmark, don’t carry over to misleading charts. This kind of robustness doesn’t emerge on its own as models climb the leaderboards.</p><p>Next, the authors tried six inference-time fixes. Only two actually worked. The first was to extract the underlying data from the chart and hand it to a text-only model to answer the question. This improved accuracy by 15.4 to 19.6 percentage points. The second was to use matplotlib to redraw the data as a clean chart and swap it in for the original. What the two have in common is that both get rid of the deceptive chart.</p><p>The costs are just as clear, though. The table-extraction step makes its own mistakes, and it gets less accurate on inverted axes, 3D effects, and dual axes. As a result, this method actually <em>lowers</em> accuracy on normal charts by up to 8.5 points. The redrawing approach only helps when the generated code runs successfully, and the redraw success rate was just 79% for scatter plots and 80% for stacked bar charts.</p><p>| Find papers faster on <a href="https://arxivsub.comfyai.app/">arXivSub</a> with AI summary (CVPR、ICCV、ECCV、ICML、ICLR、NeurIPS、AAAI、MICCAI、CoRL、RSS、ACL、EMNLP、ICRA、IROS、IJCAI)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MKbWPWkiZIqjXtZATr3m5w.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=69d6794f4c7f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Muon: An Optimizer That Orthogonalizes Its Own Updates]]></title>
            <link>https://medium.com/@ai4medical/muon-an-optimizer-that-orthogonalizes-its-own-updates-26925baa3ee1?source=rss-4f0729bdce69------2</link>
            <guid isPermaLink="false">https://medium.com/p/26925baa3ee1</guid>
            <category><![CDATA[muon]]></category>
            <category><![CDATA[optimization]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[optimizer]]></category>
            <dc:creator><![CDATA[ComfyAI]]></dc:creator>
            <pubDate>Sat, 07 Nov 2026 01:32:45 GMT</pubDate>
            <atom:updated>2026-11-07T01:32:45.462Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7lFV6BEk2N4g0O6f8y318Q.png" /></figure><p><em>Notes on Keller Jordan’s blog post on MomentUm Orthogonalized by Newton-Schulz</em></p><p>In his blog post, Keller Jordan introduces Muon, an optimizer built specifically for the 2D weight matrices in a neural network’s hidden layers. The name stands for <strong>MomentUm Orthogonalized by Newton-Schulz</strong>, and the idea is simple. First, compute the update the way SGD with momentum would (Nesterov momentum by default). Then, before applying that update to the weights, orthogonalize the update matrix.</p><h3>What “orthogonalizing” means here</h3><p>Orthogonalizing means replacing the update matrix G with the semi-orthogonal matrix closest to it. Concretely, take the SVD of G, written as USVᵀ, and keep only UVᵀ.</p><p>The author’s intuition goes like this. For the 2D parameters of a transformer, the updates produced by SGD-momentum or Adam tend to have a very high condition number, meaning the ratio of the largest singular value to the smallest is large. In practice these updates look close to low-rank. Orthogonalization amplifies the directions that have small magnitude but still matter for learning.</p><h3>Why Newton-Schulz instead of SVD</h3><p>Running a full SVD at every step is too slow, so Muon approximates the result with a Newton-Schulz iteration. Each step of the iteration needs only a few matrix multiplications, and it runs stably in bfloat16. By contrast, the Coupled Newton Iteration needs float32 to stay stable.</p><p>In the implementation, the input is first normalized by its Frobenius norm. This is safe because scaling a matrix doesn’t change the result of orthogonalizing it. The iteration then runs for 5 steps with coefficients <strong>(3.4445, -4.7750, 2.0315)</strong>, which were tuned specifically to speed up convergence.</p><h3>Why normalize first?</h3><p>This family of Newton-Schulz iterations only converges to an orthogonal matrix when the input’s singular values fall within a certain range. A matrix’s Frobenius norm is always at least as large as its largest singular value, which is its spectral norm. So dividing by the Frobenius norm guarantees that every singular value is at most 1, which places all of them inside the range where the iteration converges.</p><h3>Muon doesn’t handle everything</h3><p>Muon is not meant for every parameter. Muon is used for:</p><ul><li>2D hidden-layer weights</li><li>flattened convolution kernels</li></ul><p>AdamW still handles:</p><ul><li>scalars such as biases</li><li>vectors</li><li>the input token embeddings</li><li>the output classification head</li></ul><p>The author says this split matters for getting the best performance. My guess at the intuition is that each row of the input embedding and output head corresponds to a single token, so their updates are very sparse. That makes them quite different from the dense linear maps in the hidden layers, and forcing orthogonalization onto them may actually hurt.</p><p>Another detail: orthogonalizing Q, K, and V separately works better than treating them as one large matrix.</p><h3>Results</h3><ul><li><strong>CIFAR-10:</strong> the speed record for reaching 94% accuracy dropped from 3.3 to 2.6 A100-seconds.</li><li><strong>NanoGPT speedrun:</strong> about a 1.35× speedup.</li><li><strong>1.5B-parameter transformer:</strong> reached GPT-2 XL-level performance on HellaSwag in 10 hours on 8×H100s, compared with 13.3 hours for AdamW.</li></ul><h3>The cost</h3><p>Muon adds very little overhead. The author estimates that in typical language model training, orthogonalization adds less than 1% in FLOPs, and memory use is the same as SGD-momentum.</p><p>| Find papers faster on <a href="https://arxivsub.comfyai.app/">arXivSub</a> with AI summary (CVPR、ICCV、ECCV、ICML、ICLR、NeurIPS、AAAI、MICCAI、CoRL、RSS、ACL、EMNLP、ICRA、IROS、IJCAI)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=26925baa3ee1" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>