Papers by Feng Wan
CE-RM: A Pointwise Generative Reward Model Optimized via Two-Stage Rollout and Unified Criteria (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing studies have shown that rule-based evaluation methods are ineffective for open-ended natural language generation. |
| Approach: | They propose a pointwise generative reward model with a dedicated two-stage rollout method and unified query-based criteria that can be trained with 5.7K high-quality data. |
| Outcome: | The proposed model achieves superior performance on diverse reward model benchmarks, especially in Best-of-N scenarios, and delivers more effective improvements in downstream RL practice. |
Learning Bilingual Sentiment-Specific Word Embeddings without Cross-lingual Supervision (N19-1)
Copied to clipboard
| Challenge: | Unsupervised BWE methods are evaluated on word translation or word similarity tasks. |
| Approach: | They propose a method that learns sentiment-specific word representations for two languages in a common space without cross-lingual supervision. |
| Outcome: | The proposed method outperforms previous unsupervised BWE methods and even supervised Bwe methods on three language pairs for cross-lingual sentiment analysis. |
Empowering Math Problem Generation and Reasoning for Large Language Model via Synthetic Data based Continual Learning Framework (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing learning frameworks for large language models (LLMs) for math problem generation are limited and lack quality data. |
| Approach: | They propose a synthetic data based continual learning framework to improve LLMs ability for MPG and math reasoning. |
| Outcome: | The proposed framework improves performance on large language models and math reasoning using supervised fine-tuning, data synthesis and direct preference optimization. |
Add-One-In: Incremental Sample Selection for Large Language Models via a Choice-Based Greedy Paradigm (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies focus on individual quality and do not assess the value of training data. |
| Approach: | They propose a choice-based sample selection framework that evaluates sample quality . they use LLMs to evaluate the value of each option during the selection process . |
| Outcome: | The proposed model outperforms the full dataset and recent studies on a larger medical dataset. |
MultiDoc2Dial: Modeling Dialogues Grounded in Multiple Documents (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing work treats document-grounded dialogue modeling as a machine reading comprehension task based on a single document or passage. |
| Approach: | They propose a task and dataset for modeling goal-oriented dialogues grounded in multiple documents. |
| Outcome: | The proposed task and dataset address realistic scenarios where goal-oriented dialogues involve multiple topics and hence are grounded on different documents. |
ToolBeHonest: A Multi-level Hallucination Diagnostic Benchmark for Tool-Augmented Large Language Models (2024.emnlp-main)
Copied to clipboard
Yuxiang Zhang, Jing Chen, Junjie Wang, Yaxin Liu, Cheng Yang, Chufan Shi, Xinyu Zhu, Zihao Lin, Hanwen Wan, Yujiu Yang, Tetsuya Sakai, Tian Feng, Hayato Yamana
| Challenge: | Currently, tool-augmented large language models (LLMs) only achieve total scores of 45.3 and 37.0, respectively, on a scale of 100. |
| Approach: | They propose a multi-level diagnostic process to assess the LLM's hallucinations through two perspectives: depth and breadth. |
| Outcome: | The proposed diagnostic process assesses the hallucinations of large language models through two perspectives: depth and breadth. |
doc2dial: A Goal-Oriented Document-Grounded Dialogue Dataset (2020.emnlp-main)
Copied to clipboard
| Challenge: | doc2dial dataset is a goal-oriented document-grounded dialogue model . it is based on how the authors compose documents for guiding end users . |
| Approach: | They propose a dataset of goal-oriented dialogues grounded in documents . they use annotated conversations with an average of 14 turns to generate conversational utterances . |
| Outcome: | The proposed dataset includes over 4500 annotated conversations with an average of 14 turns grounded in over 450 documents from four domains. |
BotPercent: Estimating Bot Populations in Twitter Communities (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing approaches to bot detection are agnostic to social environments the bots operate in . however, standard approaches are not a good fit for the social environments they operate in. |
| Approach: | They propose a method that estimates the percentage of Twitter bots given a community . they use Twitter bot detection datasets and feature-, text-, and graph-based models adjusted to a particular community based on Twitter . |
| Outcome: | The proposed method achieves state-of-the-art in community-level Twitter bot detection across balanced and imbalanced class distribution settings. |
HuatuoGPT, Towards Taming Language Model to Be a Doctor (2023.findings-emnlp)
Copied to clipboard
Hongbo Zhang, Junying Chen, Feng Jiang, Fei Yu, Zhihong Chen, Guiming Chen, Jianquan Li, Xiangbo Wu, Zhang Zhiyi, Qingying Xiao, Xiang Wan, Benyou Wang, Haizhou Li
| Challenge: | Experimental results show that the distilled language model outperforms its teacher model (ChatGPT) in most cases. |
| Approach: | They propose a Large Language Model (LLM) that leverages both distilled data from **ChatGPT** and real-world data from**doctors** in the supervised fine-tuning stage. |
| Outcome: | The proposed model outperforms the teacher model in most cases by using additional real-world data and RLMF to align the language model with the merits of both sources. |
Unveiling Language-Specific Features in Large Language Models via Sparse Autoencoders (2025.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) exhibit impressive abilities in various domains such as text generation, instruction following, and reasoning. |
| Approach: | They propose a method to decompose the activations of Large Language Models into a sparse linear combination of SAE features. |
| Outcome: | The proposed method shows that some features are strongly related to specific languages, while others are unaffected by ablating them. |
AgentOCR: Reimagining Agent History via Optical Self-Compression (2026.acl-long)
Copied to clipboard
| Challenge: | Recent advances in large language models (LLMs) have enabled agentic systems trained with reinforcement learning over multi-turn interaction, but practical deployment is bottlenecked by rapidly growing textual histories that inflate token and memory costs. |
| Approach: | They propose a framework that represents the accumulated observation-action history as a compact rendered image. |
| Outcome: | The proposed framework preserves over 95% of text-based agent performance while significantly reducing token consumption (>50%), yielding consistent token and memory efficiency. |
ConvLab-3: A Flexible Dialogue System Toolkit Based on a Unified Data Format (2023.emnlp-demo)
Copied to clipboard
Qi Zhu, Christian Geishauser, Hsien-chin Lin, Carel van Niekerk, Baolin Peng, Zheng Zhang, Shutong Feng, Michael Heck, Nurul Lubis, Dazhen Wan, Xiaochen Zhu, Jianfeng Gao, Milica Gasic, Minlie Huang
| Challenge: | Existing tools for building TOD systems often lack a user-friendly interface . a toolkit with advanced, easily integrable modules is needed to bridge this gap . |
| Approach: | They propose a multifaceted dialogue system toolkit that integrates diverse datasets and models with a streamlined training process and in-depth evaluation tools. |
| Outcome: | The proposed toolkit combines RL and transfer learning to support the rapid development and evaluation of robust dialogue policies. |
PlatoLM: Teaching LLMs in Multi-Round Dialogue via a User Simulator (2024.acl-long)
Copied to clipboard
| Challenge: | Recent efforts to democratize ChatGPT have focused on leveraging real user and ChatGPP dialogues, but the most direct human needs are often ignored. |
| Approach: | They propose a method to simulate human behavior better by using real human-like questions extracted from real human conversations as a learning goal and a user simulator called ‘Socratic’. |
| Outcome: | The proposed model achieves SoTA performance among LLaMA-based 7B models in MT-Bench. |
DELL: Generating Reactions and Explanations for LLM-Based Misinformation Detection (2024.findings-acl)
Copied to clipboard
| Challenge: | Large language models are limited by challenges in factuality and hallucinations to be directly employed off-the-shelf for judging the veracity of news articles. |
| Approach: | They propose to integrate large language models into the news pipeline by generating news reactions and generating proxy tasks. |
| Outcome: | The proposed model outperforms state-of-the-art baselines by 16.8% in macro f1-score on seven datasets with three LLMs. |
What Does the Bot Say? Opportunities and Risks of Large Language Models in Social Media Bot Detection (2024.acl-long)
Copied to clipboard
| Challenge: | Social media bot detection has always been an arms race between advancements in machine learning and adversarial bot strategies to evade detection. |
| Approach: | They propose a mixture-of-heterogeneous-experts framework to divide and conquer diverse user information modalities and propose LLM-guided manipulation of user textual and structured information to evade detection. |
| Outcome: | The proposed framework outperforms state-of-the-art baselines on 1,000 annotated examples while bringing down existing detectors by 29.6% and harming calibration and reliability of bot detection systems. |
BIC: Twitter Bot Detection with Text-Graph Interaction and Semantic Consistency (2023.acl-long)
Copied to clipboard
Zhenyu Lei, Herun Wan, Wenqian Zhang, Shangbin Feng, Zilong Chen, Jundong Li, Qinghua Zheng, Minnan Luo
| Challenge: | Existing methods to identify bots rely on text or networks alone . text-graph interactions and semantic consistency are essential improvements to combat bot evolution. |
| Approach: | They propose to combine text-graph interaction and semantic Consistency to model Twitter bots' behavior based on attention weights and a text-graphic interaction module to enable information exchange across modalities in the learning process. |
| Outcome: | The proposed framework outperforms state-of-the-art methods on two widely adopted datasets and the results are consistent with previous work. |
Does Structure Matter? Encoding Documents for Machine Reading Comprehension (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing Transformer-based models for machine reading comprehension treat documents as flat sequences. |
| Approach: | They propose a Transformer-based method that reads a document as tree slices and jointly trains and consults the modules at inference time. |
| Outcome: | The proposed method outperforms several baseline approaches on two datasets from varied domains. |
Correcting Language Model Bias for Text Classification in True Zero-Shot Learning (2024.lrec-main)
Copied to clipboard
| Challenge: | Experimental results show that pre-trained language models outperform standard prompt learning in zero-shot settings. |
| Approach: | They propose a pipeline for annotating and filtering examples from unlabeled examples . they propose 'model bias validation' method that utilizes unlabed examples as validation set . |
| Outcome: | The proposed approach outperforms standard prompt learning on six text classification tasks. |
SEA: Supervised Embedding Alignment for Token-Level Visual-Textual Integration in MLLMs (2025.emnlp-main)
Copied to clipboard
Yuanyang Yin, Yaqi Zhao, Yajie Zhang, Yuanxing Zhang, Ke Lin, Jiahao Wang, Xin Tao, Pengfei Wan, Wentao Zhang, Feng Zhao
| Challenge: | Multimodal Large Language Models (MLLMs) integrate visual and textual inputs, yet modality alignment remains one of the most challenging aspects. |
| Approach: | They propose a token-level supervision alignment method that enables more precise visual-text alignment during pretraining. |
| Outcome: | The proposed method improves performance across various model sizes, with smaller models benefiting the most. |
CMB: A Comprehensive Medical Benchmark in Chinese (2024.naacl-long)
Copied to clipboard
Xidong Wang, Guiming Chen, Song Dingjie, Zhang Zhiyi, Zhihong Chen, Qingying Xiao, Junying Chen, Feng Jiang, Jianquan Li, Xiang Wan, Benyou Wang, Haizhou Li
| Challenge: | Large Language Models (LLMs) provide a great breakthrough in medicine, says a new study . existing studies on LLMs leverage subjective evaluation, but evaluation in medicine is professional . |
| Approach: | They propose a localized medical benchmark in Chinese rooted in native Chinese . they propose to use traditional Chinese medicine to evaluate large-scale LLMs . |
| Outcome: | a new benchmark is developed to evaluate large-scale LLMs in china . the proposed model is rooted in the native Chinese linguistic and cultural framework . |
Chain-of-Procedure: Hierarchical Visual-Language Reasoning for Procedural QA (2026.findings-acl)
Copied to clipboard
Guanhua Chen, Yutong Yao, Shenghe Sun, Ci-jun Gao, Shudong Liu, Lidia S. Chao, Feng Wan, Derek F. Wong
| Challenge: | Recent advances in vision-language models (VLMs) have achieved impressive results on standard image-text tasks, yet their capability in visual procedure question answering (VP-QA) remains largely unexplored. |
| Approach: | They propose a multimodal benchmark specifically designed for visual procedural reasoning that synergizes cross-modal procedure retrieval, context-aware step decomposition, and the next step prediction. |
| Outcome: | The proposed framework significantly outperforms baselines on visual procedure question answering (VP-QA) Experiments on six VLMs show that it performs better than baselines. |
APLOT: Robust Reward Modeling via Adaptive Preference Learning with Optimal Transport (2025.emnlp-main)
Copied to clipboard
| Challenge: | Experimental results show that RLHF improves performance of Large Language Models . BT-based RMs struggle to distinguish between similar preference responses . |
| Approach: | They propose to enhance BT-based reward models by using an adaptive margin mechanism . they use semantic similarity and reward-predicted reward differences to adjust focus . |
| Outcome: | Experimental results show that the proposed method outperforms existing methods in both in-distribution and OOD settings. |
Explaining Neural Network Predictions on Sentence Pairs via Learning Word-Group Masks (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing methods to explain neural network models are computationally inefficient for text inputs. |
| Approach: | They propose a method to implicitly detect word correlations by grouping correlated words from input text pairs together and measuring their contribution to corresponding NLP tasks. |
| Outcome: | The proposed method is evaluated with two different model architectures across four datasets. |
CGBridge: Bridging Code Graphs and Large Language Models for Better Structure-Aware Code Understanding (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing structure-aware approaches treat structure as serialized text prompts or auxiliary training objectives, failing to provide explicit guidance during inference. |
| Approach: | They propose a plug-and-play method that enhances Large Language Models with Code Graph information through an external, trainable Bridge module. |
| Outcome: | The proposed method decouples structural reasoning from textual generation without updating the backbone. |