Papers by Zhifang Zhang
ShieldLM: Empowering LLMs as Aligned, Customizable and Explainable Safety Detectors (2024.findings-emnlp)
Copied to clipboard
Zhexin Zhang, Yida Lu, Jingyuan Ma, Di Zhang, Rui Li, Pei Ke, Hao Sun, Lei Sha, Zhifang Sui, Hongning Wang, Minlie Huang
| Challenge: | Existing tools for detecting safety issues in LLMs are expensive and inefficient. |
| Approach: | They propose an LLM-based safety detector which annotates the safety of queries and provides explanations for its decisions. |
| Outcome: | The proposed detector outperforms baselines on four sets of query-response pairs and is effective as a safety evaluator for advanced LLMs. |
Can Large Language Models Always Solve Easy Problems if They Can Solve Harder Ones? (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) have impressive capabilities, but still suffer from inconsistency issues. |
| Approach: | They develop a ConsisEval benchmark to evaluate LLMs' inconsistency . they find that LLM models can paradoxically fail at easier problems . |
| Outcome: | The proposed model achieves highest consistency score but inconsistent to specific questions due to distraction by redundant information, misinterpretation of questions, etc. |
SelfBudgeter: Adaptive Token Allocation for Efficient LLM Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent large reasoning models have shown exceptional performance on various tasks, but they consume excessive tokens even for simple queries, leading to resource waste and prolonged user latency. |
| Approach: | They propose a self-adaptive reasoning strategy that automatically allocates budgets according to problem complexity and introduces GRPO for reinforcement learning to reduce output length. |
| Outcome: | The proposed model achieves an average response length compression of 61% on math reasoning tasks while maintaining accuracy. |
Confidence v.s. Critique: A Decomposition of Self-Correction Capability for LLMs (2025.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to improve self-correction performance of Large Language Models are based on intrinsic selfcorrectione, which allows the model to check and revise its selfgenerated answers without external feedback. |
| Approach: | They propose to decompose the self-correction capability into confidence and critique capabilities and a metric for overall self-corretion capability evaluation. |
| Outcome: | The proposed method outperforms vanilla SFT and achieves much higher accuracy after self-correction. |
A Token-level Reference-free Hallucination Detection Benchmark for Free-form Text Generation (2022.acl-long)
Copied to clipboard
| Challenge: | Existing work on pre-trained generative models often fails to detect non-existent or incorrect content . Existing studies have attempted to detect hallucinations based on oracle references . |
| Approach: | They propose a token-level, reference-free hallucination detection task based on Wikipedia annotations to detect non-existent or incorrect content. |
| Outcome: | The proposed task is token-level, reference-free hallucination detection task and dataset . authors argue that the proposed task can be used in real-time to detect hallucines . |
A Probabilistic Inference Scaling Theory for LLM Self-Correction (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated the capability to refine their generated answers through self-correction, enabling continuous performance improvement over multiple rounds. |
| Approach: | They propose a probabilistic theory to model the dynamics of accuracy change and explain performance improvements observed in multi-round self-correction. |
| Outcome: | The proposed model can predict accuracy curves and improve accuracy over multiple rounds. |
Towards Stable and Effective Reinforcement Learning for Mixture-of-Experts (2026.acl-long)
Copied to clipboard
| Challenge: | Reinforcement learning with verifiable rewards (RLVR) training with Mixture-of-Experts policies remains fragile and prone to reward collapse. |
| Approach: | They propose a router shift-based policy optimization method that computes a per-token router-shift ratio conditioned on the previously activated experts and applies stop-gradient and a lower-bound floor. |
| Outcome: | The proposed method achieves better performance and greater stability than previous methods. |
Be a Multitude to Itself: A Prompt Evolution Framework for Red Teaming (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have gained increasing attention for their capacity to generate harmful content. |
| Approach: | They propose a scalable evolution framework to evolve red teaming prompts across breadth and depth dimensions, facilitating automatic generation of numerous high-quality and diverse red team prompts. |
| Outcome: | The proposed framework surpasses existing red teaming methods on attack success rate and diversity. |
CBLUE: A Chinese Biomedical Language Understanding Evaluation Benchmark (2022.acl-long)
Copied to clipboard
Ningyu Zhang, Mosha Chen, Zhen Bi, Xiaozhuan Liang, Lei Li, Xin Shang, Kangping Yin, Chuanqi Tan, Jian Xu, Fei Huang, Luo Si, Yuan Ni, Guotong Xie, Zhifang Sui, Baobao Chang, Hui Zong, Zheng Yuan, Linfeng Li, Jun Yan, Hongying Zan, Kunli Zhang, Buzhou Tang, Qingcai Chen
| Challenge: | a new benchmark for biomedical language understanding is being developed in Chinese . most benchmarks are limited to English, which makes it difficult to replicate success in other languages. |
| Approach: | They propose to use Chinese biomedical language understanding evaluation benchmarks to evaluate Chinese models. |
| Outcome: | The proposed benchmarks show that the current models perform worse than the human ceiling. |
Towards Reverse Engineering of Language Models: A Survey (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Due to the vast amounts of data and computational resources required for model development, protecting the model’s parameters and training data has become an urgent and crucial concern. |
| Approach: | They define "reverse engineering" techniques as attacks on large language models and provide an in-depth analysis of them. |
| Outcome: | The proposed attacks are described as “reverse engineering” techniques on LMs and provide an introduction to existing protective strategies. |