Reward models (RMs) are crucial for aligning large language models (LLMs) with human preferences. However, most RM research is centered on English and relies heavily on synthetic resources, which leads to limited and less reliable datasets and benchmarks for Chinese. To address this gap, we introduce CheemsBench, a fully human-annotated RM evaluation benchmark within Chinese contexts, and CheemsPreference, a large-scale and diverse preference dataset annotated through human-machine collaboration to support Chinese RM training. We systematically evaluate open-source discriminative and generative RMs on CheemsBench and observe significant limitations in their ability to capture human preferences in Chinese scenarios. Additionally, based on CheemsPreference, we construct an RM that achieves state-of-the-art performance on CheemsBench, demonstrating the necessity of human supervision in RM training. Our findings reveal that scaled AI-generated data struggles to fully capture human preferences, emphasizing the importance of high-quality human supervision in RM development.

Automated Alignment refers to a set of algorithms designed to align Large Language Models (LLMs) with human intentions and values while minimizing manual intervention. However, it faces challenges such as algorithmic diversity and excessively convoluted workflows. We present AutoAlign, an open-source toolkit that offers:(1) a unified framework integrating mainstream automated algorithms through a consistent interface, and(2) an accessible workflow supporting one-click execution for prompt synthesis, automatic alignment signal construction, and iterative model training. Our toolkit enables easy reproduction of existing results through extensive benchmarks and facilitates the development of novel approaches via modular components. It includes implementations for both highly efficient inference and training, as well as low-resource training. By standardizing automated alignment methodologies and providing accessible implementations, AutoAlign lowers the barriers to building customized aligned models and supports academic research.

Self-critic has become a crucial mechanism for enhancing the reasoning performance of LLMs. However, current approaches mainly involve basic prompts for intuitive instance-level feedback, which resembles System-1 processes and limits the reasoning capabilities. Moreover, there is a lack of in-depth investigations into the relationship between LLM’s ability to criticize and its task-solving performance. To address these issues, we propose Critic-CoT, a novel framework that pushes LLMs toward System-2-like critic capability. Through a step-wise CoT reasoning paradigm and the automatic construction of weak-supervision data without human annotation, Critic-CoT enables LLMs to engage in slow, analytic self-critique and refinement, thereby improving their reasoning abilities. Experiments on GSM8K and MATH and out-of-domain evaluation demonstrate that our enhanced model significantly boosts task-solving performance by filtering out invalid solutions or iterative refinement. Furthermore, we investigate the intrinsic correlation between critique and task-solving abilities within LLMs, discovering that these abilities can mutually reinforce each other rather than conflict.

Hallucination occurs when large language models exhibit behavior that deviates from the boundaries of their knowledge during response generation. To address this critical issue, previous learning-based methods attempt to finetune models but are limited by off-policy sampling and coarse-grained feedback. In this paper, we present Reinforcement Learning for Hallucination (RLFH), an on-policy self-alignment approach that enables LLMs to actively explore their knowledge boundaries and self-correct generation behavior through fine-grained feedback signals. RLFH introduces a self-assessment framework where the policy serves as its own judge. Through this framework, responses are automatically decomposed into atomic facts and their truthfulness and informativeness are assessed against external knowledge sources. The resulting fine-grained feedback at the statement level are then converted into token-level dense reward signals. This enables online reinforcement learning to achieve precise and timely optimization without human intervention. Comprehensive evaluations on HotpotQA, SQuADv2, and Biography benchmarks validate RLFH’s effectiveness in hallucination mitigation.

As post-training processes utilize increasingly large datasets and base models continue to grow in size, the computational demands and implementation challenges of existing algorithms are escalating significantly. In this paper, we propose modeling the changes at the logits level during post-training using a separate neural network (i.e., the value network). After training this network on a small base model using demonstrations, this network can be seamlessly integrated with another pre-trained models during inference, enabling them to achieve similar capability enhancements. We systematically investigate the best practices for this paradigm in terms of pre-training weights and connection schemes. We demonstrate that the resulting value network has broad transferability across pre-trained models of different parameter sizes within the same family, models undergoing continuous pre-training within the same family, and models with different vocabularies across families. In certain cases, it can achieve performance comparable to full-parameter fine-tuning. Furthermore, we explore training methods to enhance transferability, which effectively improve the transfer performance of the value model across models of various parameter scales and prevent overfitting to the base model used during training.