Papers by Yijun Yang
Lookahead Q-Cache: Achieving More Consistent KV Cache Eviction via Pseudo Query (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing KV cache eviction methods prune tokens using prefilling-stage attention scores, causing inconsistency with actual inference queries. |
| Approach: | They propose a lookahead q-cache framework that generates low-cost pseudo lookaheaded queries to better approximate the true decoding-stage queries. |
| Outcome: | The proposed framework outperforms existing methods on LongBench and Needle-in-a-Haystack benchmarks and can be flexibly combined to yield further improvements. |
Make Some Noise: Unlocking Language Model Parallel Inference Capability through Noisy Training (2024.emnlp-main)
Copied to clipboard
Yixuan Wang, Xianzhen Luo, Fuxuan Wei, Yijun Liu, Qingfu Zhu, Xuanyu Zhang, Qing Yang, Dongliang Xu, Wanxiang Che
| Challenge: | Existing speculative decoding methods require additional model structure and training processes to assist the model for draft token generation. |
| Approach: | They propose a make some noise training framework that introduces some noise at the input for the model to learn the denoising task. |
| Outcome: | The proposed model improves inference speed by 2.3-2.7x times without compromising model performance. |
UniArk: Improving Generalisation and Consistency for Factual Knowledge Extraction through Debiasing (2024.naacl-long)
Copied to clipboard
| Challenge: | Existing studies have investigated the potential of language models as knowledge bases and the existence of severe biases when extracting factual knowledge. |
| Approach: | They propose an adapter-based framework for generalised factual knowledge extraction using simple methods without introducing extra parameters. |
| Outcome: | The proposed framework improves the model’s out-of-domain generalisation and consistency under various prompts. |
ALDEN: Reinforcement Learning for Active Navigation and Evidence Gathering in Long Documents (2026.acl-long)
Copied to clipboard
| Challenge: | Visually rich documents (VRDs) combine text, tables, and figures within complex, semantically structured layouts. |
| Approach: | They propose a multi-turn reinforcement learning framework that fine-tunes VLMs as interactive agents capable of actively navigating long, visually rich documents. |
| Outcome: | The proposed framework achieves state-of-the-art on five long-document benchmarks. |
Evaluating and Improving Graph to Text Generation with Large Language Models (2025.naacl-long)
Copied to clipboard
| Challenge: | Recent advances in large language models have revolutionized natural language processing due to their zero-and-short-shot capabilities. |
| Approach: | They propose a tuning-free prompting approach for graph-to-text generation tasks. |
| Outcome: | The proposed approach improves LLMs on graph-to-text generation tasks incrementally. |
EEE-QA: Exploring Effective and Efficient Question-Answer Representations (2024.lrec-main)
Copied to clipboard
| Challenge: | Current approaches to question answering rely on pre-trained language models like RoBERTa. |
| Approach: | They propose a pooling approach that embeds all answer candidates with the question . they also propose enabling cross-reference between answer choices . |
| Outcome: | The proposed methods improve throughput and memory efficiency with little sacrifice in performance. |
Adversarial Preference Learning for Robust LLM Alignment (2025.findings-acl)
Copied to clipboard
Yuanfu Wang, Pengyu Wang, Chenyang Xi, Bo Tang, Junyi Zhu, Wenqiang Wei, Chen Chen, Chao Yang, Jingfeng Zhang, Chaochao Lu, Yijun Niu, Keming Mao, Zhiyu Li, Feiyu Xiong, Jie Hu, Mingchuan Yang
| Challenge: | Modern language models rely on Reinforcement Learning from Human Feedback (RLHF) to encourage safe behaviors, but they remain vulnerable to adversarial attacks due to three key limitations: (1) the inefficiency and high cost of human annotation; (2) the vast diversity of potential adversarials; and (3) the risk of feedback bias and reward hacking. |
| Approach: | They propose an iterative adversarial training method that incorporates three key innovations to address these challenges. |
| Outcome: | Experiments on Mistral-7B-Instruct-v0.3 show that the proposed method significantly enhances robustness and reduces harmful outputs from 5.88% to 0.43%. |
Boosting Vulnerability Detection of LLMs via Curriculum Preference Optimization with Synthetic Reasoning Data (2025.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are capable of detecting software vulnerabilities, but lack of reasoning data hinders their ability to capture underlying vulnerability patterns. |
| Approach: | They propose a framework that excels at mining vulnerability patterns through reasoning data synthesizing and vulnerability-specific preference optimization. |
| Outcome: | The proposed framework improves on SVEN and PrimeVul datasets and improves 12.24%-22.77% accuracy. |
CARE-STaR: Constraint-aware Self-taught Reasoner (2025.findings-acl)
Copied to clipboard
Zhiliang Li, Bo Tang, Yijun Niu, Beihong Jin, Qiwen Shi, Yuchen Feng, Zhiyu Li, Jie Hu, Mingchuan Yang, Feiyu Xiong
| Challenge: | Recent research on instruction following has demonstrated that LLMs can handle complex instructions. |
| Approach: | They propose to assign constraints to different levels of constraints in instructions . they use chain-of-thought and self-taught reasoner methods to identify constraints . |
| Outcome: | The proposed method outperforms supervised fine-tuning (SFT) on three instruction-following benchmarks. |