arxiv:2411.04282

Language Models are Hidden Reasoners: Unlocking Latent Reasoning Capabilities via Self-Rewarding

Published on Nov 6

· Submitted by

hlnchen on Nov 11

#2 Paper of the day

Upvote

Authors:

Haolin Chen ,

Abstract

Large language models (LLMs) have shown impressive capabilities, but still struggle with complex reasoning tasks requiring multiple steps. While prompt-based methods like Chain-of-Thought (CoT) can improve LLM reasoning at inference time, optimizing reasoning capabilities during training remains challenging. We introduce LaTent Reasoning Optimization (LaTRO), a principled framework that formulates reasoning as sampling from a latent distribution and optimizes it via variational approaches. LaTRO enables LLMs to concurrently improve both their reasoning process and ability to evaluate reasoning quality, without requiring external feedback or reward models. We validate LaTRO through experiments on GSM8K and ARC-Challenge datasets using multiple model architectures. On GSM8K, LaTRO improves zero-shot accuracy by an average of 12.5% over base models and 9.6% over supervised fine-tuning across Phi-3.5-mini, Mistral-7B, and Llama-3.1-8B. Our findings suggest that pre-trained LLMs possess latent reasoning capabilities that can be unlocked and enhanced through our proposed optimization approach in a self-improvement manner. The code of LaTRO is available at https://github.com/SalesforceAIResearch/LaTRO.

View arXiv page View PDF Add to collection

Community

hlnchen

Paper author Paper submitter 3 days ago

Chain-of-thought (CoT) demonstrated strong reasoning capabilities of LLMs. But how to train them to reason?
Introducing LaTent Reasoning Optimization (LaTRO): a principled framework that formulates the reasoning trajectory as a latent variable and optimize the reasoning via variational approaches.

LaTRO has good performance: we improve zero-shot accuracy by an average of 12.5% over 3 different base models: Phi-3.5-mini, Mistral-7B, and Llama-3.1-8B.
LaTRO is reward model-free: Surprisingly but reasonable, the log probabilities of producing the correct answer after the reasoning trajectory serves as a natural reward function, which we call "Self-rewarding". No need to train additional reward models as in RLHF!
LaTRO shifts the inference-time scaling back to training time - by self-generating multiple reasoning trajectories and self-rewarding them with groundtruth during each training update
Free side benefit: one can compress the length of reasoning trajectories via LaTRO - on GSM8K, a model with 200 reasoning tokens achieves 78% performance of a model with 500 reasoning tokens.