Research

I am interested in LLM Reasoning, Generalizability of LLMs, and Interpretability of Reasoning and Generalization in LLMs.

Existing layer pruning techniques often suffer severe degradation on generative reasoning tasks. Through a systematic study, we find that tasks requiring multi-step reasoning are particularly sensitive to depth reduction, exhibiting degradation in critical algorithmic capabilities like arithmetic and code synthesis.

We identify that training reasoning models stalls on saturated problems because informative failures are rarely encountered. We propose failure-prefix conditioning, which reallocates exploration by conditioning training on prefixes from rare incorrect reasoning trajectories, matching performance gains of medium-difficulty problems.

We find that distilling (warming up) a LLM with non-domain-specific reasoning traces (like a logic game) can bring general improvements across multiple reasoning-intensive tasks like math and coding. Reinforcement Learning on top of it leads to better sample efficiency, generalizability, and final performance.

We find that LLMs form critical layers during pretraining whose removal completely destroys performance. Furthermore, the importance of such layers remain unchanged after post-training regimes like Reinforcement Learning, Distillation, and Instruction Tuning.

We investigate why RL with verifiable rewards boosts accuracy but not capability, revealing it improves easy questions at the cost of hard ones—while distillation improves both only when new knowledge is introduced.

We find that as you increase the magnitude of numbers in simple math problems, LLMs get more confused and commit logical errors (along with the expected arithmetic errors).

Website referenced from Jon Barron's source code.