Papers2

#alignment tax

Training LLMs for Divide-and-Conquer Reasoning Elevates Test-Time Scalability

Xiao Liang, Zhong-Zhi Li et al.Feb 2arXiv

The paper trains language models to solve hard problems by first breaking them into smaller parts and then solving those parts, instead of only thinking in one long chain.

#divide-and-conquer reasoning#chain-of-thought#reinforcement learning

Entropy-Adaptive Fine-Tuning: Resolving Confident Conflicts to Mitigate Forgetting

Intermediate

Muxi Diao, Lele Yang et al.Jan 5arXiv

Supervised fine-tuning (SFT) often makes a model great at a new task but worse at its old skills; this paper explains a key reason why and how to fix it.

#Entropy-Adaptive Fine-Tuning#confident conflicts#token-level entropy