Papers by Zesheng Shi
Team-Based Self-Play With Dual Adaptive Weighting for Fine-Tuning LLMs (2026.acl-long)
Copied to clipboard
| Challenge: | Recent self-training approaches have reduced reliance on human-labeled data, which limits their scalability. |
| Approach: | They propose a team-based self-play algorithm that iteratively refines alignment without additional human supervision. |
| Outcome: | The proposed algorithm outperforms baselines and LLM benchmarks in the self-supervised setting. |
Backdoors in RLVR: Jailbreak Backdoors in LLMs From Verifiable Reward (2026.acl-long)
Copied to clipboard
| Challenge: | Reinforcement Learning with Verifiable Rewards (RLVR) is an emerging paradigm that significantly boosts a Large Language Model’s reasoning abilities on complex logical tasks. |
| Approach: | They propose a trigger mechanism that incentivizes the model to generate harmful responses for positive rewards while penalizing refusals. |
| Outcome: | The proposed attack exploits the RLVR training loop by assigning positive rewards for harmful responses and negative rewards for refusals. |
Skill Weaving: Efficient LLM Improvement via Modular Skillpacks (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) can specialize under fixed memory and inference budgets, but they struggle to achieve high performance across heterogeneous domains. |
| Approach: | They propose a modular improvement framework that partitions full capabilities of a general-purpose model into domain-specific delta modules that reorganize and refine the model's internal knowledge. |
| Outcome: | The proposed framework outperforms monolithic models on multi-task and agentic benchmarks and achieves up to 4 speedup. |
LLM Inductive Reasoning Through Multi-Agent Enhanced Monte Carlo Tree Search (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for enhancing inductive reasoning of large language models often lack explicit optimization guidance and effective error correction. |
| Approach: | They propose a plug-and-play test-time framework that integrates multi-agent coordination with Monte Carlo Tree Search to improve inductive reasoning. |
| Outcome: | The proposed framework outperforms existing methods on four benchmarks and shows consistent improvements on QWQ-32B and Deepseek-V3 . |
Safety Alignment via Constrained Knowledge Unlearning (2025.acl-long)
Copied to clipboard
Zesheng Shi, Yucheng Zhou, Jing Li, Yuxin Jin, Yu Li, Daojing He, Fangming Liu, Saleh Alharbi, Jun Yu, Min Zhang
| Challenge: | Existing defense mechanisms have not fully deleted harmful knowledge in large language models (LLMs) Existing methods to address safety alignment have not completely deleted harmful information in LLMs. |
| Approach: | They propose a safety alignment strategy that uses scoring neurons to identify useful knowledge in LLMs and pruning the gradients of neurons in U to preserve beneficial information. |
| Outcome: | The proposed method significantly improves model safety while maintaining utility compared to existing methods. |
E3-TIR: Enhanced Experience Exploitation for Tool-Integrated Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing training paradigms for Large Language Models (LLMs) suffer from inefficient exploration and mode degradation due to a lack of prior guidance, while SFT-then-RL is limited by high data costs and capability plateaus caused by low-entropy collapse. |
| Approach: | They propose an Enhanced Experience Exploitation paradigm that integrates expert prefixes, expert guided, and self-exploration to improve agent training. |
| Outcome: | The proposed model achieves a 6% performance improvement over traditional paradigms on tool-use tasks while requiring less than 10% of the synthetic data. |