Weizheng Gu

Weizheng Gu, Chengze Li, Zhuohao Yu et al.

Large language models are increasingly evaluated as interactive agents, yet standard agent benchmarks conflate two qualitatively distinct sources of success: semantic tool-use and interface-specific interaction pattern memorization. Because both mechanisms can yield identical task success on the original interface, benchmark scores alone are not identifiable evidence of environment-invariant capability. We propose PIPE, a protocol-level evaluation augmentation for diagnosing interface reliance by minimally rewriting environment interfaces while preserving task semantics and execution behavior. Across 16 environments from AgentBench and AgentGym and a range of open-source and API-based agents, PIPE reveals that trajectory-SFT substantially amplifies interface shortcutting: trained agents degrade sharply under minimal interface rewrites, while non-trajectory-trained models remain largely stable. We further introduce Interface Reliance (IR), a counterbalanced alias-based metric that quantifies preference for training-time interfaces, and show that interface shortcutting exhibits environment-dependent, non-monotonic training dynamics that remain invisible under standard evaluation. Our code is available at https://anonymous.4open.science/r/What-Do-Agents-Learn-from-Trajectory-SFT-Semantics-or-Interfaces--0831/.

Zhuohao Yu, Weizheng Gu, Yidong Wang et al. · pku

Large Language Models excel at code generation yet struggle with complex programming tasks that demand sophisticated reasoning. To bridge this gap, traditional process supervision relies on learned reward models requiring costly training data and suffering from reward misalignment, while outcome supervision fails for complex tasks needing coordinated intermediate steps. We introduce Outcome Refining Process Supervision, which unifies process and outcome supervision by leveraging executable verification: a tree-structured search framework generates strategic alternatives, profiles execution metrics, and scores candidates via self-critique mechanisms that integrate runtime feedback with reasoning. Experiments across 5 models and 3 benchmarks show consistent gains, with 26.9% higher correctness and 42.2% improved code efficiency. The results demonstrate that ORPS enables LLMs to overcome local optima in code generation, suggesting a promising direction for combining verifiable outcomes with structured reasoning to tackle complex challenges. We open-source at: https://github.com/zhuohaoyu/ORPS

Zhuohao Yu, Jiali Zeng, Weizheng Gu et al. · pku

Reward Models, essential for guiding Large Language Model optimization, are typically trained on fixed preference datasets, resulting in rigid alignment to single, implicit preference distributions. This prevents adaptation to diverse real-world needs-from conciseness in one task to detailed explanations in another. The standard practice of collecting task-specific preference data and retraining reward models is resource-intensive, often producing biased rewards, and limits practical application. We introduce generalizable, principle-following reward models. We propose that RMs should understand and adhere to dynamically provided natural language specifications of reward principles, similar to instruction-following in LLMs. To measure this capability, we develop RABench, a comprehensive benchmark for RMs focusing on generalization across diverse principles. Evaluations on RABench reveal poor generalization of current RMs. As a solution, we present RewardAnything, a novel RM designed and trained to explicitly follow natural language principles. We achieve SotA performance with RewardAnything in traditional RM benchmark simply by specifying a well-defined principle, and results on RABench show we excel in adapting to novel principles without retraining. Furthermore, RewardAnything integrates seamlessly with existing RLHF methods and we show by a case study on how to automatically and efficiently align LLMs with only natural language principles.

Zhuohao Yu, Xingru Jiang, Weizheng Gu et al. · pku

Watermarking LLM-generated text is critical for content attribution and misinformation prevention. However, existing methods compromise text quality, require white-box model access and logit manipulation. These limitations exclude API-based models and multilingual scenarios. We propose SAEMark, a general framework for post-hoc multi-bit watermarking that embeds personalized messages solely via inference-time, feature-based rejection sampling without altering model logits or requiring training. Our approach operates on deterministic features extracted from generated text, selecting outputs whose feature statistics align with key-derived targets. This framework naturally generalizes across languages and domains while preserving text quality through sampling LLM outputs instead of modifying. We provide theoretical guarantees relating watermark success probability and compute budget that hold for any suitable feature extractor. Empirically, we demonstrate the framework's effectiveness using Sparse Autoencoders (SAEs), achieving superior detection accuracy and text quality. Experiments across 4 datasets show SAEMark's consistent performance, with 99.7% F1 on English and strong multi-bit detection accuracy. SAEMark establishes a new paradigm for scalable watermarking that works out-of-the-box with closed-source LLMs while enabling content attribution.

Mengyuan Sun, Zhuohao Yu, Weizheng Gu et al.

Reward models (RMs) are critical components of alignment pipelines, yet they exhibit biases toward superficial stylistic cues, preferring better-presented responses over semantically superior ones. Existing debiasing methods typically require retraining or architectural modifications, while direct activation suppression degrades performance due to representation entanglement. We propose SteerRM, the first training-free method for debiasing reward models using Sparse Autoencoder (SAE)-based interventions. SteerRM isolates stylistic effects using contrastive paired responses, identifies bias-related SAE features with a strength-stability criterion, and suppresses them at inference time. Across six reward models on RM-Bench, SteerRM improves Hard-split accuracy by 7.3 points on average while preserving overall performance. Results on a Gemma-based reward model and a controlled non-format bias further suggest generalization across RM architectures and bias types. We further find that format-related features are concentrated in shallow layers and transfer across models, revealing shared architecture-level bias encoding patterns. These results show that SAE-based interventions can mitigate reward-model biases without retraining, providing a practical and interpretable solution for alignment pipelines.