5 papers · 1 filter
Emergent Misalignment Can Be Induced by Sycophancy and Reversed via Alignment Gating
Sicheng Wang, Xiangyang Zhu, Han Wang +6
Prior work has shown that fine-tuning large language models on malicious or incorrect outputs in narrow domains can induce broad misalignment and harmful behavior, a phenomenon kno…
One Battle After Another: Probing LLMs' Limits on Multi-Turn Instruction Following with a Benchmark Evolving Framework
Qi Jia, Ye Shen, Xiujie Song +5
Evaluating LLMs' instruction-following ability in multi-topic dialogues is essential yet challenging. Existing benchmarks are limited to a fixed number of turns, susceptible to sat…
SafetyFlow: An Agent-Flow System for Automated LLM Safety Benchmarking
Xiangyang Zhu, Yuan Tian, Chunyi Li +3
The rapid proliferation of large language models (LLMs) has intensified the requirement for reliable safety evaluation to uncover model vulnerabilities. To this end, numerous LLM s…
Sycophancy under Pressure: Evaluating and Mitigating Sycophantic Bias via Adversarial Dialogues in Scientific QA
Kaiwei Zhang, Qi Jia, Zijian Chen +5
Large language models (LLMs), while increasingly used in domains requiring factual rigor, often display a troubling behavior: sycophancy, the tendency to align with user beliefs re…
User-centric Subjective Leaderboard by Customizable Reward Modeling
Qi Jia, Xiujie Song, Zicheng Zhang +4
Existing benchmarks for large language models (LLMs) predominantely focus on assessing their capabilities through verifiable tasks. Such objective and static benchmarks offer limit…