Papers by Peiran Li
Measuring Large Language Models’ Adversarial Behavior in Social Deduction Games (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing safety evaluations focus on refusal-based methods that test whether models avoid responding to inappropriate or violent requests, leaving open questions about how models behave in interactive social settings. |
| Approach: | They propose to use a meta-LLM to construct a closed behavioral taxonomy from a multi-agent simulation to examine adversarial behavior of large language models. |
| Outcome: | The proposed model-based model-driven model-model-based taxonomy shows that the model-led model-learning model exhibits distinct behavioral profiles and influences social stability and competitive success. |
FineSteer: A Unified Framework for Fine-Grained Inference-Time Steering in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for inference-time steering fail to be effective, utility-preserving and training-efficient due to rigid, one-size-fits-all designs and limited adaptability. |
| Approach: | They propose a steering framework that decomposes inference-time steering into two stages . they propose 'conditional steering' mechanism that preserves model utility by avoiding unnecessary steering . a 'mixture-of-Steering-Experts' mechanism captures multimodal nature of desired steering behaviors . |
| Outcome: | The proposed framework outperforms the state-of-the-art methods on safety and truthfulness benchmarks. |
AwarenessBench: Assessing Cognitive Capabilities of Language Models (2026.acl-long)
Copied to clipboard
Xiaojian Li, Rongwu Xu, Tianyun Zhang, Yue Wang, Shuo Chen, Qiner Lyu, Briana Zhang, Peiran Yang, Kyle Xue Chen, Haoyuan Shi, Yu Wang, Wei Xu
| Challenge: | Language models exhibit increasingly consciousness-like behaviors, requiring a baseline to evaluate their cognitive abilities. |
| Approach: | They propose a benchmark to assess the cognitive abilities of language models (LMs) they compare 18 state-of-the-art LMs to human models in metacognition, self-awareness, social awareness and situational awareness . |
| Outcome: | Evaluating 18 state-of-the-art LMs, they find they consistently surpass baselines . but most models fall short in metacognition and self-awareness, the study finds . |
ToxiGAN: Toxic Data Augmentation via LLM-Guided Directional Adversarial Generation (2026.eacl-long)
Copied to clipboard
| Challenge: | Generating toxic text for augmentation is a sensitive and challenging task . toxic language remains pervasive, often escaping automated moderation systems . |
| Approach: | They propose a class-aware text augmentation framework that combines adversarial generation with semantic guidance from large language models to provide balanced guidance. |
| Outcome: | Experiments on four hate speech benchmarks show that ToxiGAN outperforms traditional and LLM-based augmentation methods. |
Re-Align: Aligning Vision Language Models via Retrieval-Augmented Direct Preference Optimization (2025.emnlp-main)
Copied to clipboard
Shuo Xing, Peiran Li, Yuping Wang, Ruizheng Bai, Yueqi Wang, Chan-Wei Hu, Chengxuan Qian, Huaxiu Yao, Zhengzhong Tu
| Challenge: | emergence of large Vision Language Models (VLMs) has broadened the capabilities of single-modal Large Language Model (LLM) but VLMs are prone to significant hallucinations, especially in the form of cross-modal inconsistencies. |
| Approach: | They propose a new alignment framework that leverages image retrieval to integrate both textual and visual preference signals. |
| Outcome: | The proposed framework mitigates hallucinations more effectively than previous methods . it maintains robustness and scalability across a wide range of VLM sizes and architectures . |