SACHIN KUMAR·Oct 30Can Small AI Agents Run IT Ops? One 4B Generalist Tied Two 4B SpecialistsA 30B planner plus two 4B specialists fine-tuned on its verified work, that i built at the NVIDIA × Red Hat Small Models Hack. The…A response icon1A response icon1
SACHIN KUMAR·May 2, 2025THINKPRM: generative Process Reward Model for solution verification via long CoT ReasoningStep-by-step verifiers — also known as process reward models (PRMs) helps in test time scaling, but do require step-level supervision…
SACHIN KUMAR·Apr 21, 2025CLustering-based Iterative Data Mixture Bootstrapping: framework for Optimizing Data Mixture to…Identifying an optimal pre-training data mixture remains a challenging problem, despite its significant benefits for pre-training…
SACHIN KUMAR·Apr 12, 2025Breaking the LLM Bottleneck: GenPRM’s approach to Scale test-time compute via generative reasoningRecently Process Reward Models (PRMs) as verifiers served as promising approach to enhance the performance of LLMs. However, current PRMs…
SACHIN KUMAR·Mar 26, 2025ReSearch: LLM training framework to Reason with Search via Reinforcement LearningIntegrating reasoning with external search processes remains challenging, especially for complex multi-hop questions requiring multiple…
SACHIN KUMAR·Mar 25, 2025LEMMA: framework to enhance mathematical reasoning of LLMs by learning from errorsCurrent approaches to improve LLM’s mathematical reasoning abilities primarily focus on improving the quality of correct training data…
SACHIN KUMAR·Mar 20, 2025DeepPerception: knowledge-driven reasoning for R1-like Cognitive Visual Perception in MLLMsCurrent Multimodal Large Language Models (MLLMs) lacks human like abilities of fine-grained visual discrimination by leveraging domain…
SACHIN KUMAR·Mar 12, 2025Symbolic Mixture-of-Experts: using symbolic output to combine expert LLMs for heterogeneous…For scalably tackling large-scale and diverse tasks, selecting experts at the task level is often too coarse-grained, as heterogeneous…
SACHIN KUMAR·Mar 8, 2025Highlighted Chain of Thought: prompting approach to reference text from input question in LLM…LLM responses containing both factual and non-factual statements poses a challenge for humans to verify and accurately base their decisions…
SACHIN KUMAR·Mar 6, 2025ATLAS: approach to finetune LLM Agents by identifying critical steps in expert trajectoriesCurrent LLM agent tuning approaches typically employ supervised finetuning on entire expert trajectories, which introduces expert bias and…