Papers by Fan Cheng
QUITO-X: A New Perspective on Context Compression from the Information Bottleneck Theory (2025.findings-emnlp)
Copied to clipboard
Yihang Wang, Xu Huang, Bowen Tian, Yueyang Su, Lei Yu, Huaming Liao, Yixing Fan, Jiafeng Guo, Xueqi Cheng
| Challenge: | Existing methods for compressing context by removing redundant tokens are inconsistent with the objective of retaining the most important tokens when conditioning on a given query. |
| Approach: | They propose a method that uses information bottleneck theory to compress context . they propose to remove redundant tokens using metrics such as self-information or perplexity . |
| Outcome: | The proposed method achieves a 25% increase in compression rate compared to the state-of-the-art . |
Schoenfeld’s Anatomy of Mathematical Reasoning by Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models expose reasoning traces, yet their underlying cognitive structure and steps remain difficult to identify and analyze beyond surface-level statistics. |
| Approach: | They propose a framework that explicitly abstracts reasoning traces into functional reasoning steps such as Analysis, Explore, Implement, Verify, etc. |
| Outcome: | The proposed framework reveals reproducible thinking dynamics and structural differences between reasoning and non-reasoning models, which are not apparent from token-level views. |
Dropping Experts, Recombining Neurons: Retraining-Free Pruning for Sparse Mixture-of-Experts LLMs (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Sparse Mixture-of-Experts (SMoE) architectures require loading all expert parameters . previous work focused on expert pruning and merging but focused on neuron-level structure . |
| Approach: | They propose a task-agnostic framework for expert pruning and reconstruction . it prunes redundant experts using router statistics, then decomposes them into neuron-level expert segments . |
| Outcome: | The proposed framework reduces the number of experts and memory usage, making it easier to deploy. |
SDiaReward: Modeling and Benchmarking Spoken Dialogue Rewards with Modality and Colloquialness (2026.acl-long)
Copied to clipboard
Jingyu Lu, Yuhan Wang, Fan Zhuo, Xize Cheng, Changhao Pan, Xueyi Pu, Yifu Chen, Chenyuhao Wen, Tianle Liang, Zhou Zhao
| Challenge: | SDiaReward is an end-to-end spoken dialogue system that integrates paralinguistic nuances and spontaneous nature of human conversation. |
| Approach: | They propose an end-to-end multi-turn reward model trained on SDiaReward-Dataset . it is a collection of episode-level preference pairs targeting modality and colloquiality gaps . |
| Outcome: | The proposed model outperforms general-purpose audio LLMs in episode-level evaluation. |
Controlling Risk of Retrieval-augmented Generation: A Counterfactual Prompting Framework (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies on retrieval-augmented generation (RAG) rarely address the issue of predictive uncertainty, i.e., how likely it is that a RAG model’s prediction is incorrect. |
| Approach: | They propose a framework that induces RAG models to alter latent factors and analyzes the effect on their answers. |
| Outcome: | The proposed framework identifies two critical factors affecting RAG models' confidence in their answers and analyzes the effect on their answers. |
PreP-OCR: A Complete Pipeline for Document Image Restoration and Enhanced OCR Accuracy (2025.acl-long)
Copied to clipboard
| Challenge: | Existing pipelines that combine document image restoration with semantic-aware post-OCR correction can improve text extraction from degraded images. |
| Approach: | They propose a two-stage pipeline that combines document image restoration with semantic-aware post-OCR correction to enhance both visual clarity and textual consistency. |
| Outcome: | The proposed pipeline reduces character error rates by 63.9-70.3% on 13,831 pages of real historical documents in English, French, and Spanish compared to OCR on raw images. |
WESR: A Benchmark and Strong Baseline for Word-level Event-Speech Recognition (2026.findings-acl)
Copied to clipboard
Chenchen Yang, Kexin Huang, Liwei Fan, Qian Tu, Botian Jiang, Dong Zhang, Linqi Yin, Shimin Li, Zhaoye Fei, Qinyuan Cheng, Xipeng Qiu
| Challenge: | aaron carroll: the precise localization of non-verbal vocal events remains a critical yet under-explored challenge. carroll says current methods suffer from insufficient task definitions with limited category coverage. carrol: knowing exactly where an event occurred is not enough; knowing exactly what it happened is. |
| Approach: | They propose a taxonomy of 21 vocal events with a new categorization into discrete versus continuous types. |
| Outcome: | The proposed model disentangles ASR errors from event detection while maintaining ASR quality. |
One Refiner to Unlock Them All: Inference-Time Reasoning Elicitation via Reinforcement Query Refinement (2026.acl-long)
Copied to clipboard
| Challenge: | Existing alignment methods for Large Language Models (LLMs) are expensive and lack the flexibility to fully activate their latent reasoning capabilities. |
| Approach: | They propose a modular framework that treats reasoning elicitation as an inference-time alignment task. |
| Outcome: | The proposed framework outperforms baselines by 2.1% on average across diverse architectures and benchmarks. |
Compete to Complete: Co-opetition Adversarial Learning for Retrieval-Augmented Generation (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to reduce hallucination in large language models lack a robust mechanism for generating a generative model. |
| Approach: | They propose a framework that formulates retriever–generator training in RAG as a minimax game. |
| Outcome: | The proposed framework improves retrieval-augmented generation performance on seven benchmark datasets. |
From Relevance to Utility: Evidence Retrieval with Feedback for Fact Verification (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing evidence retrieval models are based on probability ranking principle . existing models do not align with retrieval-enhanced verification frameworks . |
| Approach: | They propose a feedback-based evidence retriever that optimizes the evidence retrieval process by incorporating feedback from the claim verifier. |
| Outcome: | Empirical studies show that the proposed method is superior to baseline methods. |
Contextual Fine-to-Coarse Distillation for Coarse-grained Response Selection in Open-Domain Conversations (2022.acl-long)
Copied to clipboard
Wei Chen, Yeyun Gong, Can Xu, Huang Hu, Bolun Yao, Zhongyu Wei, Zhihao Fan, Xiaowu Hu, Bartuer Zhou, Biao Cheng, Daxin Jiang, Nan Duan
| Challenge: | Existing studies focus on coarse-grained response selection in retrieval-based dialogue systems. |
| Approach: | They propose a Contextual Fine-to-Coarse (CFC) distilled model for coarse-grained response selection in open-domain conversations. |
| Outcome: | The proposed model improves over baseline methods on two datasets based on the Reddit comments dump and Twitter corpus compared with baseline methods. |
From Implicit Graph Encoding to Explicit Evidence: A Training-Free LLM Framework for Temporal Knowledge Graph Reasoning (2026.findings-acl)
Copied to clipboard
Guo Tang, Ke Cheng, Huiming Fan, Heng Chang, Wenxiang Zheng, Xianhao Ou, Junjia Xiang, Ming Liu, Yujun Zhou, Li Lanyu, Bing Qin
| Challenge: | Existing Large Language Models (LLMs) struggle with implicit modality alignment and suboptimal graph linearization. |
| Approach: | They propose a training-free, test-time adaptive framework that reframes TKG prediction as explicit evidence-driven reasoning. |
| Outcome: | ExE-LLM outperforms fully trained graph neural networks on four benchmarks . it achieves SOTA performance in inductive settings, significantly outperforming fully trained neural networks . |
SPIDE: Serial and Parallel Intertwined Speculative Decoding (2026.findings-acl)
Copied to clipboard
Wenru Xu, Peixuan Xu, Ziqi Yang, Ming Hu, Zihui Wang, Jianzhong Qi, Rongshan Yu, Xiaoliang Fan, Cheng Wang
| Challenge: | Speculative decoding (SD) is a training-free SD framework that orchestrates dynamic alternation combining serial dynamic drafting with parallel draft verification. |
| Approach: | They propose a serial and parallel intertwined speculative DEcoding framework that orchestrates dynamic alternation combining serial dynamic drafting and parallel draft verification. |
| Outcome: | The proposed framework accelerates inference while reducing the LLM usage costs. |
Fewer is More: Boosting Math Reasoning with Reinforced Context Pruning (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have shown impressive capabilities, yet they struggle with math reasoning. |
| Approach: | They propose a coarse-to-fine pruner that prunes unimportant tokens to fit the context window. |
| Outcome: | The proposed approach outperforms prompting baselines across various LLMs and 5 math datasets and achieves 4.55% absolute improvements without any fine-tuning. |
NegotiationToM: A Benchmark for Stress-testing Machine Theory of Mind on Negotiation Surrounding (2024.findings-emnlp)
Copied to clipboard
Chunkit Chan, Cheng Jiayang, Yauwai Yim, Zheye Deng, Wei Fan, Haoran Li, Xin Liu, Hongming Zhang, Weiqi Wang, Yangqiu Song
| Challenge: | Theory of mind evaluations currently focus on testing models using machine-generated data or game settings prone to shortcuts and spurious correlations. |
| Approach: | They propose a benchmark to stress-test machine ToM in real-world negotiation surrounding covered multi-dimensional mental states. |
| Outcome: | The proposed benchmark builds upon the Belief-Desire-Intention theory and conducts the necessary empirical experiments to evaluate large language models. |
Just Adjust One Prompt: Enhancing In-Context Dialogue Scoring via Constructing the Optimal Subgraph of Demonstrations and Prompts (2023.emnlp-main)
Copied to clipboard
| Challenge: | Using large language models as chatbots can cause hallucinations and lack of empathy, authors report . a dimension-agnostic scoring method is proposed to improve the performance of chatbot performance . |
| Approach: | They propose a dimension-agnostic scoring method that leverages in-context learning . they propose to automatically generate prompts and then request the LLM multiple times . |
| Outcome: | The proposed method outperforms baselines on five datasets. |
Prompt Tuning with Contradictory Intentions for Sarcasm Recognition (2023.eacl-main)
Copied to clipboard
| Challenge: | Recent advances have shown that Pre-trained Language Models (PLMs) can achieve promising performance in many downstream Natural Language Processing (NLP) tasks. |
| Approach: | They propose to incorporate prior knowledge about contradictory intentions into prompt tuning for sarcasm recognition by mimicking the actual intention by verbalizer engineering. |
| Outcome: | The proposed model mimics the actual intention by prompt construction and indicates whether the actual intent contradicts the literal content by verbalizer engineering. |
TaCube: Pre-computing Data Cubes for Answering Numerical-Reasoning Questions over Tabular Data (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing auto-regressive pre-trained language models are challenged by recent emerging numerical reasoning datasets due to the error-prone implicit calculation. |
| Approach: | They propose a pre-computation tool to pre-compute aggregation/arithmetic results for the table in advance, so they are handy and readily available for PLMs to answer numerical reasoning questions. |
| Outcome: | The proposed model improves on TAT-QA and T5 and BART-large on multiple benchmarks. |
A Generative Framework for Personalized Sticker Retrieval (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing relevance-based generative retrieval methods lack personalization, leading to a mismatch between diverse user expectations and the retrieved results. |
| Approach: | They propose a representation learning model that learns discriminative user representations to encode user-specific sticker preferences. |
| Outcome: | The proposed framework outperforms state-of-the-art methods in generating relevant stickers for queries. |
CapArena: Benchmarking and Analyzing Detailed Image Captioning in the LLM Era (2025.findings-acl)
Copied to clipboard
Kanzhi Cheng, Wenpo Song, Jiaxin Fan, Zheng Ma, Qiushi Sun, Fangzhi Xu, Chenyang Yan, Nuo Chen, Jianbing Zhang, Jiajun Chen
| Challenge: | Image captioning has been a challenge for vision-language researchers for decades . current VLMs focus on tasks like visual question answering (YA) but image captioning is not as advanced as expected. |
| Approach: | They evaluate VLMs' performance on image captioning using human annotations . they find that some metrics show high caption-level agreement with humans . |
| Outcome: | The proposed model outperforms open-source models on image captioning . it achieves 93.4% correlation with human rankings at $4 per test . |
XToM: Exploring the Multilingual Theory of Mind for Large Language Models (2026.acl-long)
Copied to clipboard
Chunkit Chan, Yauwai Yim, Hongchuan Zeng, Zhiying Zou, Xinyuan Cheng, Zhifan Sun, Zheye Deng, Kawai Chung, Yuzhuo Ao, Fan Yixiang, Cheng Jiayang, Ercong Nie, Ginny Wong, Helmut Schmid, Hinrich Schuetze, Simon See, Yangqiu Song
| Challenge: | Existing evaluations of ToM in LLMs are limited to English, neglecting the linguistic diversity that shapes human cognition. |
| Approach: | They propose a multilingual benchmark that evaluates ToM across five languages . they find that models excel in multilingual language understanding, but their ToM performance varies across languages. |
| Outcome: | The proposed benchmark evaluates LLMs across five languages and incorporates diverse task scenarios. |
Empowering Reliable Visual-Centric Instruction Following in MLLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing benchmarks for evaluating instruction-following capabilities focus on verbal instructions in the textual modality. |
| Approach: | They propose to incorporate vision-dependent constraints into instruction design to enable a more rigorous assessment of how well MLLMs align their outputs with both visual input and textual instructions. |
| Outcome: | The proposed benchmark incorporates vision-dependent constraints into instruction design, enabling a more rigorous and fine-grained assessment of how well MLLMs align their outputs with both visual input and textual instructions. |
Stop Hardening Everything: A Training-Free Neuron-Level Defense for Neural Ranking Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing defenses for neural ranking models are data-centric and require retraining and adversarial data generation. |
| Approach: | They propose a model-centric defense that addresses vulnerability at its architectural source without costly retraining or adversarial data generation. |
| Outcome: | The proposed approach outperforms state-of-the-art models on MS MARCO and TREC 19 while maintaining strong performance on clean data. |
Learning to Control the Specificity in Neural Response Generation (P18-1)
Copied to clipboard
| Challenge: | Existing generative conversational models tend to favor general and trivial responses which appear frequently. |
| Approach: | They propose a controlled response generation mechanism to handle different utterance-response relationships in terms of specificity. |
| Outcome: | The proposed model outperforms state-of-the-art models under automatic and human evaluations. |
LCDS: A Logic-Controlled Discharge Summary Generation System Supporting Source Attribution and Expert Review (2025.acl-demo)
Copied to clipboard
| Challenge: | Large language models (LLMs) are capable of generating inaccurate discharge summary content or fabricating information without valid sources. |
| Approach: | They propose a tool for empowering LLMs with Logic-Controlled Discharge Summary generation. |
| Outcome: | The proposed tool identifies the writing logic of discharge summaries and integrates it with EMRs to generate silver discharge summararies. |
Minimize Exposure Bias of Seq2Seq Models in Joint Entity and Relation Extraction (2020.findings-emnlp)
Copied to clipboard
Ranran Haoran Zhang, Qianying Liu, Aysa Xuemo Fan, Heng Ji, Daojian Zeng, Fei Cheng, Daisuke Kawahara, Sadao Kurohashi
| Challenge: | Existing methods to extract relation triplets from plain text introduce exposure bias . prior work has focused on pipeline methods that ignore intrinsic interactions between subtasks and propagate classification errors through the tasks. |
| Approach: | They propose a model that reduces the decoding length to three within a triplet and removes the order among triplets. |
| Outcome: | The proposed model overfits to both datasets while showing better generalization. |
Your LLM Agents are Temporally Blind: The Misalignment Between Tool Use Decisions and Human Time Perception (2026.findings-acl)
Copied to clipboard
Yize Cheng, Arshia Soltani Moakhar, Chenrui Fan, Parsa Hosseini, Kazem Faghih, Zahra Sodagar, Wenxiao Wang, Soheil Feizi
| Challenge: | Large language model agents assume a stationary context, failing to account for real-world time elapsed between messages. |
| Approach: | They construct a dataset of multi-turn user–agent message trajectories across 76 scenarios . they collect human preferences between "calling a tool" and "directly answering" they also examine whether existing models lack human temporal perception . |
| Outcome: | The results show that existing models display poor alignment with human temporal perception . the findings provide insights to foster the development of more time-aware and human-aligned agents. |
Visual Named Entity Linking: A New Dataset and A Baseline (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing tasks in Visual Entity Linking (VEL) rely on textual data to complement multi-modal linking or only link objects with general entities. |
| Approach: | They propose a task to link regions of images with corresponding entities in Knowledge Bases . they propose three sub-tasks, based on a human-annotated visual person dataset . |
| Outcome: | The proposed task is based on a human-annotated visual person linking dataset . the proposed sub-tasks are validated on the WIKIPerson dataset based upon the proposed methods . |
TLUE: A Tibetan Language Understanding Evaluation Benchmark (2025.emnlp-main)
Copied to clipboard
Fan Gao, Cheng Huang, Yutong Liu, Nyima Tashi, Xiangxiang Wang, Thupten Tsering, Ban Ma-bao, Renzeng Duojie, Gadeng Luosang, Rinchen Dongrub, Dorje Tashi, Xiao Feng Cd, Yongbin Yu, Hao Wang
| Challenge: | Low-resource languages, like Tibetan, remain underrepresented in large language models' evaluations. |
| Approach: | They propose a Tibetan Language Understanding Evaluation Benchmark to assess LLMs' proficiency in Tibetan . they use a multi-task understanding benchmark and a safety benchmark to evaluate models . |
| Outcome: | The proposed benchmark shows that most large language models perform below the random baseline, especially in Tibetan language processing. |
Bootstrapped Pre-training with Dynamic Identifier Prediction for Generative Retrieval (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for document retrieval rely on static document identifiers . experimental results show that generative retrieval is outperforms dense retrieval in document retrievals. |
| Approach: | They propose a bootstrapped pre-training method that dynamically adjusts document identifiers during pre-train to accommodate the continuing memorization of the corpus. |
| Outcome: | The proposed method significantly outperforms existing pre-training generative retrieval baselines and performs well even in zero-shot settings. |
RoCEL: Advancing Table Entity Linking through Distinctive Row and Column Contexts (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for table entity linking ignore row and column contexts . existing methods for TEL focus on understanding sequential text contexts, making it difficult to adapt to the row and columns structure of tables. |
| Approach: | They propose to leverage row and column contexts to enhance the semantics of mentions in entity disambiguation. |
| Outcome: | The proposed method outperforms the state-of-the-art (SOTA) baseline by 1.5% on the in-domain dataset and 3.7% on average across three out-of domain datasets. |
PILE: Pairwise Iterative Logits Ensemble for Multi-Teacher Labeled Distillation (2022.emnlp-industry)
Copied to clipboard
Lianshang Cai, Linhao Zhang, Dehong Ma, Jun Fan, Daiting Shi, Yi Wu, Zhicong Cheng, Simiu Gu, Dawei Yin
| Challenge: | Pre-trained language models have been a key part of ranking systems . knowledge distillation is widely used to maintain high performance while keeping efficient computations. |
| Approach: | They propose an algorithm to combine knowledge from multi-teachers and label information to achieve competitive performance in offline and online experiments. |
| Outcome: | The proposed method has been deployed in a real-world commercial search system. |
OpenCoder: The Open Cookbook for Top-Tier Code Large Language Models (2025.acl-long)
Copied to clipboard
Siming Huang, Tianhao Cheng, Jason Klein Liu, Weidi Xu, Jiaran Hao, Liuyihan Song, Yang Xu, Jian Yang, Jiaheng Liu, Chenchen Zhang, Linzheng Chai, Ruifeng Yuan, Xianzhen Luo, Qiufeng Wang, YuanTao Fan, Qingfu Zhu, Zhaoxiang Zhang, Yang Gao, Jie Fu, Qian Liu, Houyi Li, Ge Zhang, Yuan Qi, Xu Yinghui, Wei Chu, Zili Wang
| Challenge: | Code LLMs lack reproducible data pipelines and training protocols for reproducible advancements in code intelligence. |
| Approach: | They propose a top-tier code LLM that releases model weights and inference code . reproducible data pipelines, rigorous experimental ablation results and training protocols are included . |
| Outcome: | The proposed model achieves comparable performance to leading models and serves as an "open cookbook" reproducible training data, rigorous experimental ablation results, and detailed training protocols are also included in the model. |
UEGP: Unified Expert-Guided Pre-training for Knowledge Rekindle (2024.findings-naacl)
Copied to clipboard
Yutao Mou, Kexiang Wang, Jianhe Lin, Dehong Ma, Jun Fan, Daiting Shi, Zhicong Cheng, Gu Simiu, Dawei Yin, Weiran Xu
| Challenge: | Existing paradigms for pre-training and fine-tuning have limitations . knowledge rekindle aims to break through performance upper bounds of experts without introducing additional annotated data. |
| Approach: | They propose a new paradigm for pre-training and fine-tuning that aims to re-incorporate the fine- tuned expert model into the training cycle and break through performance upper bounds of experts. |
| Outcome: | The proposed model breaks through performance upper bounds of experts without additional annotated data. |
Privacy Checklist: Privacy Violation Detection Grounding on Contextual Integrity Theory (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing privacy studies focus on sub-fields, but they focus on a few sub-domains. |
| Approach: | They propose to use the Health Insurance Portability and Accountability Act of 1996 as an example to develop a checklist that covers social identities, private attributes, and existing privacy regulations. |
| Outcome: | The proposed checklist covers social identities, private attributes, and existing privacy regulations. |
FORTAP: Using Formulas for Numerical-Reasoning-Aware Table Pretraining (2022.acl-long)
Copied to clipboard
| Challenge: | Tables store rich numerical data, but numerical reasoning over tables is still a challenge. |
| Approach: | They propose a spreadsheet formula is a valuable supervision for numerical reasoning in tables. |
| Outcome: | The proposed method outperforms state-of-the-art methods on three representative datasets of formula prediction, question answering, and cell type classification. |
Pretraining Data Detection for Large Language Models: A Divergence-based Calibration Method (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to detect text in training corpus are limited due to their low token probabilities. |
| Approach: | They propose a method to calibrate token probabilities for pretraining data detection by using a divergence-based calibration method. |
| Outcome: | The proposed method significantly outperforms existing methods on Chinese text on English-language benchmarks and patents. |
Enhancing Distractor Generation for Multiple-Choice Questions with Retrieval Augmented Pretraining and Knowledge Graph Integration (2024.findings-acl)
Copied to clipboard
Han Cheng Yu, Yu An Shih, Kin Man Law, KaiYu Hsieh, Yu Chen Cheng, Hsin Chih Ho, Zih An Lin, Wen-Chuan Hsu, Yao-Chung Fan
| Challenge: | Existing LMs undergo task-agnostic pertaining, but task-specific pretraining has gained prominence. |
| Approach: | They propose retrieval augmented pretraining and task-specific pretraining for DG . they propose to refine language model pretraining to align it more closely with downstream task . |
| Outcome: | The proposed method improves the performance of multiple-choice questions by integrating knowledge graphs and language models. |