Yu Dong

Hongming Zhang, Ruixin Hong, Dong Yu

Autoregressive decoding algorithms that use only past information often cannot guarantee the best performance. Recently, people discovered that looking-ahead algorithms such as Monte Carlo Tree Search (MCTS) with external reward models (RMs) can significantly improve models' output by allowing them to think ahead and leverage future outputs and associated rewards to guide the current generation. Such techniques can help the reinforcement fine-tuning phase by sampling better trajectories and the inference phase by selecting the better output. However, their high computational cost limits their applications, especially in streaming scenarios. To address this issue, we propose equipping the policy model with token-level self-reward modeling (TRM) capability to eliminate the need for external models and extra communication. We name the new architecture as Reward Transformer. In addition, we propose a streaming-looking-ahead (SLA) algorithm to further boost search efficiency with better parallelization. Experiments show that SLA achieves an overall win rate of 79.7\% against the baseline greedy decoding algorithm on three general-domain datasets with a frozen policy model while maintaining streaming efficiency. If we combine SLA with reinforcement fine-tuning techniques such as DPO, SLA achieves an overall win rate of 89.4\%.

Dian Yu, Yuheng Zhang, Jiahao Xu et al.

Large language models (LLMs) can refine their responses based on feedback, enabling self-improvement through iterative training or test-time refinement. However, existing methods predominantly focus on refinement within the same reasoning format, which may lead to non-correcting behaviors. We propose CaP, a novel approach that uses external tools to refine chain-of-thought (CoT) responses generated by the same or other LLMs. CaP employs a two-stage training process: supervised fine-tuning followed by preference optimization with DPO variants. Our observations highlight the critical role of preference optimization in enabling effective refinement. Additionally, we compare several sampling strategies to leverage CoT and tools at inference time. Experimental results demonstrate CaP's potential for effective cross-reasoning refinement and efficient inference.

Yuheng Zhang, Dian Yu, Baolin Peng et al.

Reinforcement Learning with Human Feedback (RLHF) has achieved great success in aligning large language models (LLMs) with human preferences. Prevalent RLHF approaches are reward-based, following the Bradley-Terry (BT) model assumption, which may not fully capture the complexity of human preferences. In this paper, we explore RLHF under a general preference framework and approach it from a game-theoretic perspective. Specifically, we formulate the problem as a two-player game and propose a novel online algorithm, iterative Nash policy optimization (INPO). The key idea is to let the policy play against itself via no-regret learning, thereby approximating the Nash policy. Unlike previous methods, INPO bypasses the need for estimating the expected win rate for individual responses, which typically incurs high computational or annotation costs. Instead, we introduce a new loss objective that is directly minimized over a preference dataset. We provide theoretical analysis for our approach and demonstrate its effectiveness through experiments on various representative benchmarks. With an LLaMA-3-8B-based SFT model, INPO achieves a 42.6% length-controlled win rate on AlpacaEval 2.0 and a 37.8% win rate on Arena-Hard, showing substantial improvement over the state-of-the-art online RLHF algorithms.