Papers by Cong Fu
Safety Alignment in NLP Tasks: Weakly Aligned Summarization as an In-Context Attack (2024.acl-long)
Copied to clipboard
| Challenge: | Recent developments in balancing usefulness and safety of large language models raise a critical question . current attacks, especially adversarial ones that manipulate malicious prompts, often aim to manipulate the input . |
| Approach: | They show that LLMs can effectively summarize malicious long documents but often refuse to translate them. |
| Outcome: | The findings highlight a vulnerability in LLMs that can't translate or summarize documents . the study focuses on LLM models, Gemini and GPT-4, which can' be exploited . |
AgentCPM-GUI: Building Mobile-Use Agents with Reinforcement Fine-Tuning (2025.emnlp-demos)
Copied to clipboard
Zhong Zhang, Yaxi Lu, Yikun Fu, Yupeng Huo, Shenzhi Yang, Yesai Wu, Han Si, Xin Cong, Haotian Chen, Yankai Lin, Xie Xie, Wei Zhou, Wang Xu, Zhou Su, Zhongwu Zhai, Xiaoming Liu, null Meiyudong, Jianming Xu, Hongyan Tian, Chongyi Wang, Chi Chen, Yuan Yao, Zhiyuan Liu, Maosong Sun
| Challenge: | Large language model agents have enabled GUI-based automation, but their deployment is limited by noisy data, poor generalization, and lack of support for non-English GUIs. |
| Approach: | They propose an 8B-parameter GUI agent built for robust and efficient on-device GUI interaction. |
| Outcome: | The proposed GUI agent achieves promising performance on five public benchmarks and proposed Chinese benchmark CAGUI. |
Mulan: A Multi-Level Alignment Model for Video Question Answering (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods focus on visual-language alignment at the video level, but they do not account for fine-grained semantic interaction between video and text. |
| Approach: | They propose a multi-level Alignment Model for Video Question Answering that establishes alignment between visual and textual modalities at the object-level, frame-level and video-level. |
| Outcome: | The proposed model outperforms state-of-the-art methods even with a small amount of extra visual-language pre-training data and a reduced number of trainable parameters. |
Towards Provably Secure Generative AI: Reliable Consensus Sampling (2026.findings-acl)
Copied to clipboard
Yu Cui, Hang Fu, Sicheng Pan, Zhuoyu Sun, Yifei Liu, Yuhong Nie, Bo Ran, Baohan Huang, Xufeng Zhang, Haibin Zhang, Cong Zuo, Licheng Wang
| Challenge: | Existing research on generative AI security is driven by mutually reinforcing attack and defense methodologies grounded in empirical experience. |
| Approach: | They propose a new algorithm that uses a random sampling algorithm to control risk. |
| Outcome: | The proposed algorithm improves robustness and utility while maintaining latency comparable to existing algorithms. |
MASPO: Unifying Gradient Utilization, Probability Mass, and Signal Reliability for Robust and Sample-Efficient LLM Reasoning (2026.acl-long)
Copied to clipboard
Xiaoliang Fu, Jiaye Lin, Yangyi Fang, Binbin Zheng, Chaowen Hu, Zekai Shao, Cong Qin, Lu Pan, Ke Zeng, Xunliang Cai
| Challenge: | Existing RLVR algorithms rely on rigid, uniform, and symmetric trust region mechanisms . current algorithms lack robustness, asymmetric signal reliability and inefficient gradient utilization . |
| Approach: | They propose a framework to harmonize three dimensions of RLVR algorithms, a paper argues . a binary cutoff is used to discard valuable reinforcement signals, they argue . |
| Outcome: | The proposed framework outperforms baselines in evaluating a robust RLVR solution. |
From log 𝜋 to 𝜋: Taming Divergence in Soft Clipping via Bilateral Decoupled Decay of Probability Gradient Weight (2026.acl-long)
Copied to clipboard
Xiaoliang Fu, Jiaye Lin, Yangyi Fang, Chaowen Hu, Cong Qin, Zekai Shao, Binbin Zheng, Lu Pan, Ke Zeng
| Challenge: | Standard algorithms for Large Language Models (LLMs) enforce stability via "hard clipping" but relying on log-probability gradient yields divergent weights as probabilities vanish, destabilizing LLM training. |
| Approach: | They propose a decoupled gradient policy optimization that uses a decay mechanism to decouple the probability of a boundary token. |
| Outcome: | The proposed algorithm outperforms baselines on various mathematical benchmarks. |
ChatMusician: Understanding and Generating Music Intrinsically with LLM (2024.findings-acl)
Copied to clipboard
Ruibin Yuan, Hanfeng Lin, Yi Wang, Zeyue Tian, Shangda Wu, Tianhao Shen, Ge Zhang, Yuhang Wu, Cong Liu, Ziya Zhou, Liumeng Xue, Ziyang Ma, Qin Liu, Tianyu Zheng, Yizhi Li, Yinghao Ma, Yiming Liang, Xiaowei Chi, Ruibo Liu, Zili Wang, Chenghua Lin, Qifeng Liu, Tao Jiang, Wenhao Huang, Wenhu Chen, Jie Fu, Emmanouil Benetos, Gus Xia, Roger Dannenberg, Wei Xue, Shiyin Kang, Yike Guo
| Challenge: | Despite LLMs' impressive capabilities in musical knowledge, music reasoning remains an unsolved task. |
| Approach: | They propose an open-source large language model (LLM) that integrates intrinsic musical abilities into LLaMA2 and GPT-3.5. |
| Outcome: | The proposed model can understand and generate music with a pure text tokenizer without external multi-modal neural structures or tokenizers. |
Collaborative Policy Learning for Open Knowledge Graph Reasoning (D19-1)
Copied to clipboard
| Challenge: | Existing models of knowledge graph reasoning suffer from limited performance when working on sparse and incomplete graphs due to the lack of evidential paths that can reach target entities. |
| Approach: | They propose a framework to train two collaborative agents to reason for missing facts over a graph augmented by a text corpus. |
| Outcome: | Experiments on two public datasets show the proposed approach is effective on a knowledge graph reasoning task. |
Contrastive Prompting Enhances Sentence Embeddings in LLMs through Inference-Time Steering (2025.acl-long)
Copied to clipboard
| Challenge: | Existing studies focus on prompt engineering to encode the full semantics of a sentence into the embedding of the last token. |
| Approach: | They propose a technique that introduces an extra auxiliary prompt to elicit better sentence embedding . they propose to use the hidden state of the token as the sentence embedded in LLMs . |
| Outcome: | The proposed technique can improve performance of existing prompt-based methods on STS tasks and downstream classification tasks. |
Adaptive Simultaneous Sign Language Translation with Confident Translation Length Estimation (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing non-simultaneous sign language translation methods suffer from inherent inference delays in real-time scenarios. |
| Approach: | They propose an adaptive policy for simultaneous sign language translation that progressively converts incrementally received sign video into its corresponding natural sentence. |
| Outcome: | The proposed policy excels in situations requiring extremely low latency. |
Distance between Relevant Information Pieces Causes Bias in Long-Context LLMs (2025.findings-acl)
Copied to clipboard
Runchu Tian, Yanghao Li, Yuepeng Fu, Siyang Deng, Qinyu Luo, Cheng Qian, Shuo Wang, Xin Cong, Zhong Zhang, Yesai Wu, Yankai Lin, Huadong Wang, Xiaojiang Liu
| Challenge: | Positional biases in large language models hinder their ability to process long inputs. |
| Approach: | They propose a benchmark to assess positional bias in large language models involving multiple pieces of relevant information. |
| Outcome: | The proposed benchmark assesses the performance of long-context language models by examining their models with different input lengths and tasks. |
From Cross-Task Examples to In-Task Prompts: A Graph-Based Pseudo-Labeling Framework for In-context Learning (2025.findings-emnlp)
Copied to clipboard
| Challenge: | In-context learning (ICL) enables large language models to perform novel tasks without parameter updates by conditioning on a few input-output examples. |
| Approach: | They propose a cost-efficient two-stage pipeline that reduces reliance on LLMs for data labeling. |
| Outcome: | The proposed pipeline reduces reliance on LLMs for data labeling . it leverages readily available cross-task examples to prompt an LLM and pseudo-label a small set of target task instances. |
How to Allocate, How to Learn? Dynamic Rollout Allocation and Advantage Modulation for Policy Optimization (2026.findings-acl)
Copied to clipboard
Yangyi Fang, Jiaye Lin, Xiaoliang Fu, Cong Qin, Haolin Shi, Chaowen Hu, Lu Pan, Ke Zeng, Xunliang Cai
| Challenge: | Existing methods for reinforcement learning with verifiable rewards are limited by the complexity of the problem and the complexity. |
| Approach: | They propose a theoretically-grounded dual-pronged optimization framework for reinforcement learning with verifiable rewards that compensates for gradient attenuation of high-confidence correct actions while utilizing entropy changes as computable indicators to stabilize excessive update magnitudes. |
| Outcome: | The proposed framework compensates for gradient attenuation of high-confidence correct actions while utilizing entropy changes as computable indicators to stabilize excessive update magnitudes. |
Proximity-Based Multi-Turn Optimization: Practical Credit Assignment for LLM Agent Training (2026.acl-industry)
Copied to clipboard
| Challenge: | Existing group-based policy optimization methods rely on statistical deviation within discrete batches, misallocating credit when task difficulty fluctuates. |
| Approach: | They propose a framework for multi-turn LLM agents that integrates global context . they propose GRPO, which integrates success-rate-aware modulation and proximity-based soft aggregation . |
| Outcome: | The proposed framework yields performance gains over existing baselines with negligible computational cost. |
Free-MAD: Consensus-Free Multi-Agent Debate (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing multi-agent debate methods rely on multiple rounds of interaction among agents to reach consensus, and the final output is decided by majority voting in the last round. |
| Approach: | They propose a multi-agent debate framework that eliminates the need for consensus among agents and reconstructs the debate phase by introducing anti-conformity. |
| Outcome: | Experiments on eight benchmark datasets show that Free-MAD significantly improves reasoning performance while requiring only a single-round debate and thus reducing token costs. |