10 papers
Greed Is Learned: Visible Incentives as Reward-Hacking Triggers
Tong Che, Rui Wu
Safety evaluations test a policy on prompts that omit the incentive information deployment supplies: a commission, a performance score, a dashboard naming which action pays best. W…
Constitutional Value Potentials: reading and steering internal priority margins in language models
Tong Che, Rui Wu
Safety evaluations test a policy on prompts that omit the incentive information deployment supplies: a commission, a performance score, a dashboard naming which action pays best. W…
Reference Feature Atlases for Mechanistic Auditing of Language Models
Rui Wu, Tong Che
Auditing a new language model usually means relearning and reinterpreting its internal features from scratch. We propose a reference feature atlas: a sparse feature library trained…
Weak Critics Make Strong Learners: On-Policy Critique Distillation for Scalable Oversight
Can Jin, Jiakang Li, Rui Wu +2
As large language models become stronger, weak supervisors may fail to provide reliable labels, preferences, or final judgments for complex outputs, limiting both weak-to-strong ge…
Taming Actor-Observer Asymmetry in Agents via Dialectical Alignment
Bobo Li, Rui Wu, Zibo Ji +5
Large Language Model agents have rapidly evolved from static text generators into dynamic systems capable of executing complex autonomous workflows. To enhance reliability, multi-a…
From Rebound to Remedy: Understanding and Mitigating Reward Hacking via Representation Engineering
Rui Wu, Ruixiang Tang
Reinforcement learning for LLMs is vulnerable to reward hacking, where models exploit shortcuts to maximize reward without solving the intended task. We systematically study this p…
Chinese Labor Law Large Language Model Benchmark
Zixun Lan, Maochun Xu, Yifan Ren +7
Recent advances in large language models (LLMs) have led to substantial progress in domain-specific applications, particularly within the legal domain. However, general-purpose mod…
Reasoning over Precedents Alongside Statutes: Case-Augmented Deliberative Alignment for LLM Safety
Can Jin, Rui Wu, Tong Che +10
Ensuring that Large Language Models (LLMs) adhere to safety principles without refusing benign requests remains a significant challenge. While OpenAI introduces deliberative alignm…
Read the Scene, Not the Script: Outcome-Aware Safety for LLMs
Rui Wu, Yihao Quan, Zeru Shi +3
Safety-aligned Large Language Models (LLMs) still show two dominant failure modes: they are easily jailbroken, or they over-refuse harmless inputs that contain sensitive surface si…
The Devil behind the mask: An emergent safety vulnerability of Diffusion LLMs
Zichen Wen, Jiashu Qu, Zhaorun Chen +13
Diffusion-based large language models (dLLMs) have recently emerged as a powerful alternative to autoregressive LLMs, offering faster inference and greater interactivity via parall…