Papers by Dongrui Liu
Identifying Semantic Induction Heads to Understand In-Context Learning (2024.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have demonstrated remarkable performance, but lack of transparency in their inference logic raises concerns about their trustworthiness. |
| Approach: | They conduct a detailed analysis of the operations of attention heads to understand their in-context learning of LLMs. |
| Outcome: | The proposed analysis of attention heads reveals that they increase the output logits of object tokens and recall objects . the proposed model is a novel approach to understand the in-context learning of large language models. |
The Tug of War Within: Mitigating the Fairness-Privacy Conflicts in Large Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods to enhance an LLM's privacy awareness with thousands of samples decrease its fairness awareness. |
| Approach: | They propose a training-free method to Suppress the Privacy and faIrness coupled Neurons (SPIN) which theoretically and empirically decreases the mutual information between fairness and privacy awareness. |
| Outcome: | The proposed method reduces the mutual information between fairness and privacy awareness without compromising general capabilities. |
AgentGym2: Benchmarking Large Language Model Agents in De-Idealized Real-World Environments (2026.acl-long)
Copied to clipboard
Zhiheng Xi, Dingwen Yang, Jiaqi Liu, Jixuan Huang, Honglin Guo, Baodai Huang, Tinggang Chen, Qi Zhang, Zhonghang Lu, Chenyu Liu, Jiajun Sun, Jiazheng Zhang, Dingwei Zhu, Xin Guo, Junzhe Wang, Zhihao Zhang, Yuming Yang, Junjie Ye, Minghe Gao, Dongrui Liu, Jiaming Ji, Guohao Li, Tao Gui, Qi Zhang, Xuanjing Huang
| Challenge: | Existing benchmarks evaluate agents in simplified, idealized settings, relying on pre-packaged tool interfaces, overlooking critical steps, and assume inputs are clean and fully specified. |
| Approach: | They propose a framework that evaluates language agents in simplified, idealized settings . they show that even SOTA systems like Gemini and GPT-5 struggle on AgentGym2 . |
| Outcome: | Experiments on 15 proprietary and open-source models show that even SOTA systems like Gemini and GPT-5 struggle on AgentGym2 . |
The LLM Already Knows: Estimating LLM-Perceived Question Difficulty via Hidden Representations (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for difficulty estimation rely on repeated response sampling, auxiliary models, or fine-tuning the target model itself. |
| Approach: | They propose a method that leverages only the hidden representations produced by large language models. |
| Outcome: | The proposed method outperforms baselines in difficulty estimation on textual and multimodal tasks and improves adaptive reasoning strategies with fewer generated tokens. |
Cooperative or Competitive? Understanding the Interaction between Attention Heads From A Game Theory Perspective (2025.acl-long)
Copied to clipboard
| Challenge: | a number of attention-based large language models (LLMs) focus on individual head contributions, but the precise interaction mechanisms between attention heads remain poorly understood. |
| Approach: | They propose a game-theoretic attention calibration method that uses the Harsanyi dividend . they selectively retain heads demonstrating significant cooperative gains and apply fine-grained adjustments to remaining heads . |
| Outcome: | The proposed framework is based on the Harsanyi dividend, a concept from cooperative game theory. |
ReasonAny: Incorporating Reasoning Capability to Any Model via Simple and Effective Model Merging (2026.acl-long)
Copied to clipboard
| Challenge: | Existing models with long chain-of-thought reasoning lack reasoning depth and domain-specific utility. |
| Approach: | They propose a model merging framework that integrates reasoning with domain-specific task models. |
| Outcome: | The proposed model merging framework outperforms state-of-the-art models while maintaining robust reasoning performance. |
LED-Merging: Mitigating Safety-Utility Conflicts in Model Merging with Location-Election-Disjoint (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for fine-tuning large language models for specialized tasks are costly and time-consuming. |
| Approach: | They propose a framework that locates task-specific neurons via gradient-based attribution and dynamically Elects critical neurons through multi-model importance fusion. |
| Outcome: | The proposed framework reduces harmful response rates while preserving 95% of utility performance. |
Towards Tracing Trustworthiness Dynamics: Revisiting Pre-training Period of Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing studies focus on pre-trained LLMs to better understand and improve their trustworthiness. |
| Approach: | They apply linear probing to LLMs to explore five key dimensions of trustworthiness: reliability, privacy, toxicity, fairness, and robustness. |
| Outcome: | The proposed model can distinguish concepts in each trustworthiness dimension, suggesting that it can be trained in early pre-training. |
VLSBench: Unveiling Visual Leakage in Multimodal Safety (2025.acl-long)
Copied to clipboard
| Challenge: | Existing studies show that textual unlearning does not achieve comparable safety performance with image-text alignment. |
| Approach: | They propose to use textual unlearning to align MLLMs with image-text pairs to explain this problem . they construct a visual leakless safety bench with 2.2k image- text pairs to test this problem. |
| Outcome: | The proposed model can refuse image-text pairs according to textual queries, leading to unreliable safety evaluations. |
EvoBench: Towards Real-world LLM-Generated Text Detection Benchmarking for Evolving Large Language Models (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to detect LLM-generated texts rely on static benchmarks that neglect the evolving nature of LLMs. |
| Approach: | They propose a benchmark to evaluate the generalization of LLM-generated text detection methods. |
| Outcome: | The proposed benchmark measures generalization of 14 detection methods across LLMs. |
LLMs know their vulnerabilities: Uncover Safety Gaps through Natural Distribution Shifts (2025.acl-long)
Copied to clipboard
Qibing Ren, Hao Li, Dongrui Liu, Zhanxu Xie, Xiaoya Lu, Yu Qiao, Lei Sha, Junchi Yan, Lizhuang Ma, Jing Shao
| Challenge: | Current safety training focuses on teaching models to reject harmful queries, but recent research shows that adversarial attacks or jailbreak methods bypass these safety mechanisms. |
| Approach: | They propose to use a new attack method to craft multi-turn toxic prompts that gradually lead LLMs to reveal unsafe content. |
| Outcome: | The proposed method outperforms existing methods in diversity, effectiveness, and efficiency across aligned LLMs. |
X-Boundary: Establishing Exact Safety Boundary to Shield LLMs from Jailbreak Attacks without Compromising Usability (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for enhancing LLM security compromise usability, study finds . boundary-safe representations close to harmful representations are disrupted, resulting in usability decline . |
| Approach: | They propose a method to push harmful representations away from boundary-safe representations and obtain an exact distinction boundary. |
| Outcome: | The proposed method reduces over-refusal rate and maintains general capability . it pushes harmful representations away from boundary-safe representations, thereby reducing usability. |
LLMs Deceive Unintentionally: Emergent Misalignment in Dishonesty from Misaligned Samples to Biased Human-AI Interactions (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing studies have shown that LLMs finetuned on incorrect completions can exhibit harmful behaviors, which is called emergent misalignment. |
| Approach: | They investigate whether LLMs finetuned on incorrect completions can exhibit harmful behaviors . they find that 1% of misalignment data is sufficient to decrease honest behavior . |
| Outcome: | The proposed model can be misaligned on errors within narrow domains to exhibit harmful behaviors . the proposed model is able to exhibit dishonest behavior with only 10% biased user population . |
AgentSlimming: Towards Efficient and Cost-Aware Multi-Agent Systems (2026.acl-long)
Copied to clipboard
| Challenge: | Automated expansion methods often result in bloated structures with redundant agents, leading to excessive token consumption. |
| Approach: | They propose a plug-and-play compression framework for graph-structured multi-agent workflows . they estimate the importance score of each agent and remove redundant agents . |
| Outcome: | Experiments show that AgentSlimming reduces average token cost by 78.9% with negligible performance degradation. |