Papers by Muhan Zhang
Law in Silico: Simulating Legal Society with LLM-Based Agents (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are powerful tools for legal simulation, but their application remains underexplored. |
| Approach: | They propose a unified LLM-based agent framework for simulating legal scenarios . they calibrate agent behaviors against real-world crime data . |
| Outcome: | The proposed framework calibrates agent behaviors against real-world crime data. |
Towards a Mechanistic Understanding of Large Reasoning Models: A Survey of Training, Inference, and Failures (2026.acl-long)
Copied to clipboard
Yi Hu, Jiaqi Gu, Ruxin Wang, Zijun Yao, Hao Peng, Xiaobao Wu, Jianhui Chen, Muhan Zhang, Liangming Pan
| Challenge: | Recent research has shown that reinforcement learning can elicit intriguing emergent reasoning behaviors. |
| Approach: | They propose a comprehensive survey of the mechanistic understanding of large reasoning models . they organize findings into three core dimensions: 1) training dynamics, 2) reasoning mechanisms, and 3) unintended behaviors. |
| Outcome: | This paper synthesizes the mechanistic understanding of large reasoning models into three dimensions . authors outline a roadmap for future studies including improved interpretability and methodologies . |
SubTokenTest: A Practical Benchmark for Real-World Sub-token Understanding (2026.acl-long)
Copied to clipboard
| Challenge: | Existing benchmarks have highlighted character-level tasks as lacking practical relevance . many real-world applications rely heavily on precise sub-token understanding . |
| Approach: | They propose a benchmark that assesses sub-token understanding through practical tasks . they examine the impact of test-time scaling on sub-word reasoning . |
| Outcome: | The proposed benchmark assesses sub-token understanding through practical tasks . it includes ten tasks across four domains and isolates tokenization-related failures . |
RulE: Knowledge Graph Reasoning with Rule Embedding (2024.findings-acl)
Copied to clipboard
| Challenge: | Knowledge graph reasoning is an important problem for knowledge graphs. |
| Approach: | They propose a framework that leverages logical rules to enhance KG reasoning by learning rule embeddings from existing triplets and first-order rules. |
| Outcome: | The proposed framework outperforms existing embedding-based and rule-based methods on multiple benchmarks. |
HD-PiSSA: High-Rank Distributed Orthogonal Adaptation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for large language models constrain update to low-rank subspaces, limiting expressiveness and performance. |
| Approach: | They propose a distributed PEFT approach that initializes adapters across different devices and aggregates their delta updates collectively on (W) Empirically, HD-PiSSA provides 16 higher effective updated ranks than data-parallel LoRA or PiSSA when fine-tuning on 8 GPUs with the same per-device adapter rank. |
| Outcome: | Empirically, HD-PiSSA outperforms LoRA and PiSSA in math, code, and multi-task learning tasks. |
JurisBench: A Deep Benchmark for Assessing Large Language Models in Professional Legal Practice (2026.acl-long)
Copied to clipboard
Ziang Chen, Guannan Li, Fanlin Ji, Yipeng Kang, Jiaqi Li, Muhan Zhang, Yangtao Zhang, Li Tianjiao, Jiannan Wang, Xin Guo, Song-Chun Zhu, Bin Ling
| Challenge: | Existing legal benchmarks evaluate isolated tasks or exam-style questions, failing to capture the procedural interdependencies and adjudicative rigor inherent in professional practice. |
| Approach: | They propose a vertical, depth-oriented, domain-specific benchmark to evaluate Large Language Models (LLMs) in Chinese civil litigation. |
| Outcome: | The proposed benchmarks show that large language models exhibit an "illusion of competence" the results highlight a critical gap between fluent linguistic output and judicial reliability . |
Knowledge is Not Enough: Injecting RL Skills for Continual Adaptation (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models face the "knowledge cutoff" problem because their parametric memory remains frozen after pretraining, preventing them from natively internalizing new information or tools on the fly. |
| Approach: | They propose a framework that supports modular skill transfer for efficient and effective knowledge adaptation by extracting a domain-agnostic **Skill Vector from a source domain. |
| Outcome: | Experiments on knowledge-incorporation QA (SQuAD, LooGLE) and agentic tool-use benchmarks show that the proposed framework outperforms state-of-the-art self-editing SFT by 9.9 points. |