Papers by Zhifang Sui

Bi-Drop: Enhancing Fine-tuning Generalization via Synchronous sub-net Estimation and Optimization (2023.findings-emnlp)

Challenge:	Existing evaluation frameworks that use large language models as referees are insufficient for accurately assessing their alignment with human intent.
Approach:	They propose a calibration framework to address positional bias in large language models as evaluators by manually annotating the “win/tie/lose” outcomes of responses from ChatGPT and Vicuna-13B in the Vicun A Benchmark’s question prompt.
Outcome:	The proposed framework alleviates evaluation bias, resulting in closer alignment with human judgments.

Copied to clipboard

Shoujie Tong, Heming Xia, Damai Dai, Runxin Xu, Tianyu Liu, Binghuai Lin, Yunbo Cao, Zhifang Sui

SG-FSM: A Self-Guiding Zero-Shot Prompting Paradigm for Multi-Hop Question Answering Based on Finite State Machine (2025.findings-naacl)

Challenge:	Pretrained language models can be fine-tuned on limited training data, which can overfit and thus diminish performance.
Approach:	They propose a fine-tuning strategy that selectively updates model parameters using gradients from various sub-nets dynamically generated by dropout.
Outcome:	The proposed method outperforms existing methods on the GLUE benchmark and exhibits excellent generalization ability and robustness for domain transfer, data imbalance, and low-resource scenarios.

Copied to clipboard

Xiaochen Wang, Junqing He, Liang Chen, Gholamreza Haffari, Yiru Wang, Zhe Yang, Xiangdi Meng, Kunhao Pan, Zhifang Sui

Towards Fine-grained Text Sentiment Transfer (P19-1)

Challenge:	Multi-hop Question Answering (MHQA) is a challenging task that requires models to answer multiple questions with multiple passages.
Approach:	They propose a self-guided prompting finite state machine to improve multi-hop reasoning abilities by iterating over multiple questions and correcting itself to improve accuracy.
Outcome:	The proposed approach outperforms baselines on Musique and other datasets.

Copied to clipboard

Fuli Luo, Peng Li, Pengcheng Yang, Jie Zhou, Yutong Tan, Baobao Chang, Zhifang Sui, Xu Sun

Sentiment Analysis

Challenge:	Existing methods for fine-grained text sentiment transfer only reverse the sentiment polarity of text, but they lack a robust and parallel learning algorithm.
Approach:	They propose a novel fine-grained text sentiment transfer task that revises a sequence to satisfy a given sentiment intensity while preserving the original semantic content.
Outcome:	The proposed model outperforms existing methods by a large margin in automatic evaluation and human evaluation.

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

Revisiting Distant Supervision for Relation Extraction (L18-1)

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.

Copied to clipboard

Tingsong Jiang, Jing Liu, Chin-Yew Lin, Zhifang Sui

Information Extraction & Text Mining

Challenge:	Existing approaches for relation extraction (RE) use supervised learning on relation-specific training data, which is expensive to acquire.
Approach:	They propose to use a new testing dataset to re-examine distant supervision approaches . they aim to draw new conclusions based on the new testing data .
Outcome:	The proposed method can generate training data without noise and bias issues . the proposed method is annotated by the researchers on Amzaon Mechanical Turk .

Denoising Bottleneck with Mutual Information Maximization for Video Multimodal Fusion (2023.acl-long)

Copied to clipboard

Shaoxiang Wu, Damai Dai, Ziwei Qin, Tianyu Liu, Binghuai Lin, Yunbo Cao, Zhifang Sui

Towards Comprehensive Description Generation from Factual Attribute-value Tables (P19-1)

Challenge:	Prior denoising methods suppress redundant and noisy information at risk of losing critical information.
Approach:	They propose a denoising bottleneck fusion model for fine-grained video multimodal fusion . they employ a bottleneck mechanism to filter out noise and redundancy with a restrained receptive field .
Outcome:	The proposed model improves on state-of-the-art video multimodal fusion benchmarks.

Copied to clipboard

Tianyu Liu, Fuli Luo, Pengcheng Yang, Wei Wu, Baobao Chang, Zhifang Sui

Can Large Language Models Always Solve Easy Problems if They Can Solve Harder Ones? (2024.emnlp-main)

Challenge:	Existing models for comprehensive descriptions for factual attribute-value tables might suffer from missing key attributes and groundless information problems.
Approach:	They propose a force attention method to encourage the generator to pay more attention to uncovered attributes to avoid potential key attributes missing.
Outcome:	The proposed model outperforms the state-of-the-art baselines on automatic and human evaluation.

Copied to clipboard

Zhe Yang, Yichang Zhang, Tianyu Liu, Jian Yang, Junyang Lin, Chang Zhou, Zhifang Sui

DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models (2024.acl-long)

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.

Copied to clipboard

Damai Dai, Chengqi Deng, Chenggang Zhao, R.x. Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y. Wu, Zhenda Xie, Y.k. Li, Panpan Huang, Fuli Luo, Chong Ruan, Zhifang Sui, Wenfeng Liang

ImageNetVC: Zero- and Few-Shot Visual Commonsense Evaluation on 1000 ImageNet Categories (2023.findings-emnlp)

Challenge:	Mixture-of-Experts (MoE) architectures face challenges in ensuring expert specialization . despite the promising performance, scaling language models to an extremely large scale is associated with exceedingly high computational costs.
Approach:	They propose an architecture that allows for ultimate expert specialization by segmenting experts into mN ones and activating mK from them.
Outcome:	The proposed architecture achieves comparable performance with GShard with 2B parameters and computation.

Copied to clipboard

Heming Xia, Qingxiu Dong, Lei Li, Jingjing Xu, Tianyu Liu, Ziwei Qin, Zhifang Sui

Learning Robust Representations for Continual Relation Extraction via Adversarial Class Augmentation (2022.emnlp-main)

Challenge:	Large Language Models (LLMs) are becoming general-purpose APIs, requiring visual knowledge to be understood.
Approach:	They propose to evaluate the visual capability of large-scale large-language models through visual commonsense evaluation using a human-annotated dataset.
Outcome:	The proposed dataset compares the visual commonsense knowledge of large-scale models with those of unimodal LLMs and visually augmented models.

Copied to clipboard

Peiyi Wang, Yifan Song, Tianyu Liu, Binghuai Lin, Yunbo Cao, Sujian Li, Zhifang Sui

Robustness in NLP

Challenge:	Existing studies attribute catastrophic forgetting to the corruption of the learned representations as new relations come . Continual relation extraction models suffer from catastrophic forgetting when learning new relations .
Approach:	They propose to use adversarial class augmentation mechanism to learn more precise representations . they propose to train the model on a sequence of tasks where two new relations are discovered .
Outcome:	The proposed model improves on two popular benchmarks.

HPT: Hierarchy-aware Prompt Tuning for Hierarchical Text Classification (2022.emnlp-main)

Copied to clipboard

Zihan Wang, Peiyi Wang, Tianyu Liu, Binghuai Lin, Yunbo Cao, Zhifang Sui, Houfeng Wang

Hierarchical Curriculum Learning for AMR Parsing (2022.acl-short)

Challenge:	Hierarchical text classification (HTC) is a multi-label classification problem with a complex label hierarchy.
Approach:	They propose a Hierarchy-aware Prompt Tuning method to handle HTC from a multi-label perspective using a dynamic virtual template and label words that take the form of soft prompts to fuse the label hierarchy knowledge.
Outcome:	The proposed method achieves state-of-the-art performance on 3 popular HTC datasets and is adept at handling imbalance and low resource situations.

Copied to clipboard

Peiyi Wang, Liang Chen, Tianyu Liu, Damai Dai, Yunbo Cao, Baobao Chang, Zhifang Sui

SelfBudgeter: Adaptive Token Allocation for Efficient LLM Reasoning (2026.findings-acl)

Challenge:	Abstract Meaning Representation (AMR) parsing aims to translate sentences to semantic representation with a hierarchical structure, but there is a gap between their flat training objective and the hierarchic structure, which limits the model generalization.
Approach:	They propose a Hierarchical Curriculum Learning framework with Structure-level (SC) and Instance-level curricula (IC) that aims to translate sentences to semantic representation with a hierarchical structure.
Outcome:	Experiments on AMR2.0, AMR3.0, structure-complex and out-of-distribution situations confirm the effectiveness of the proposed framework.

Copied to clipboard

Zheng Li, Qingxiu Dong, Jingyuan Ma, Di Zhang, Kai Jia, Zhifang Sui

Reasoning

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

Zhe Yang, Yichang Zhang, Yudong Wang, Ziyao Xu, Junyang Lin, Zhifang Sui

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.

Beyond Single Frames: Can LMMs Comprehend Implicit Narratives in Comic Strip? (2025.findings-emnlp)

Copied to clipboard

Xiaochen Wang, Heming Xia, Jialin Song, Longyu Guan, Qingxiu Dong, Rui Li, Yixin Yang, Yifan Pu, Weiyao Luo, Yiru Wang, Xiangdi Meng, Wenjie Li, Zhifang Sui

DialogUSR: Complex Dialogue Utterance Splitting and Reformulation for Multiple Intent Detection (2022.findings-emnlp)

Challenge:	Large Multimodal Models have demonstrated strong performance on vision-language benchmarks, yet current evaluations focus on single-image reasoning.
Approach:	STRIPCIPHER is a benchmark designed to evaluate model ability on understanding implicit narratives in silent comics.
Outcome:	STRIPCIPHER is a high-quality, human-annotated dataset featuring fine-grained annotations and comprehensive coverage of varying difficulty levels.

Copied to clipboard

Haoran Meng, Zheng Xin, Tianyu Liu, Zizhen Wang, He Feng, Binghuai Lin, Xuemin Zhao, Yunbo Cao, Zhifang Sui

Natural Language Interfaces

Challenge:	DialogUSR is a plug-in and domain-agnostic module that empowers multi-intent detection for chatbots . a single user query triggers inquiries on highspeed train ticket price and weather of destination.
Approach:	They propose a dialog utterance splitting and reformulation task that splits multi-intent user query into multiple single-intention sub-queries and recovers all coreferred and omitted information in the sub-questions.
Outcome:	The proposed model can be used to split multi-intent user queries into multiple sub-queries . it can be trained in two stages and perform in-depth analyses on the proposed models .

Knowledge Neurons in Pretrained Transformers (2022.acl-long)

Copied to clipboard

Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, Furu Wei

InfoCL: Alleviating Catastrophic Forgetting in Continual Text Classification from An Information Theoretic Perspective (2023.findings-emnlp)

Challenge:	Existing studies show that pretrained language models are good at recalling factual knowledge without fine-tuning.
Approach:	They propose a method to identify neurons that express factual knowledge in pretrained Transformers by filling-in-the-blank cloze queries.
Outcome:	The proposed method can be used to edit, erase, and update factual knowledge without fine-tuning.

Copied to clipboard

Yifan Song, Peiyi Wang, Weimin Xiong, Dawei Zhu, Tianyu Liu, Zhifang Sui, Sujian Li

Information Extraction & Text Mining

Challenge:	Recent studies have identified the severe performance decrease on analogous classes as a key factor for catastrophic forgetting.
Approach:	They propose a replay-based continual text classification method that uses fast-slow and current-past contrastive learning to perform mutual information maximization and better recover previously learned representations.
Outcome:	The proposed method achieves state-of-the-art on three text classification tasks.

Pun-GAN: Generative Adversarial Network for Pun Generation (D19-1)

Copied to clipboard

Fuli Luo, Shunyao Li, Pengcheng Yang, Lei Li, Baobao Chang, Zhifang Sui, Xu Sun

Robustness in NLP

Challenge:	Existing methods for generating pun sentences with word senses lack large-scale corpus for supervised learning . a pun is a clever and amusing use of a word with two meanings (word senses)
Approach:	They propose an adversarial generative network for pun generation with a generator and a discriminator to distinguish between generated pun sentences and real sentences with specific word senses.
Outcome:	The proposed network generates sentences that are more ambiguous and diverse in both automatic and human evaluation.

SenseJudge: Human-Centric Preference-Driven Judgment Framework (2026.findings-acl)

Copied to clipboard

Rui Li, Junfeng Liu, Xiangwen Kong, Zhifang Sui

Challenge:	Existing judgment approaches rely on trained judgers using fixed preference data . existing judgment approaches neglect diverse user preferences and struggle to adapt to real-world human-AI dialogue scenarios.
Approach:	They propose a customizable judgment framework driven by human preferences and a diverse instruction following benchmark derived from real-world multi-turn interactions.
Outcome:	The proposed framework surpasses other judgment methods and models in two tasks, and achieves model ranking that aligns with real human sense.

Discriminatively-Tuned Generative Classifiers for Robust Natural Language Inference (2020.emnlp-main)

Copied to clipboard

Xiaoan Ding, Tianyu Liu, Baobao Chang, Zhifang Sui, Kevin Gimpel

Text Classification

Challenge:	Recent work has shown advantages of generative classifiers in terms of data efficiency and robustness.
Approach:	They propose a generative classifier for natural language inference (NLI) they compare it to discriminative models and large-scale pretrained models like BERT .
Outcome:	The proposed classifier outperforms discriminative and pretrained baselines across several challenging NLI experimental settings, including small training sets, imbalanced label distributions, and label noise.

Premise-based Multimodal Reasoning: Conditional Inference on Joint Textual and Visual Clues (2022.acl-long)

Copied to clipboard

Qingxiu Dong, Ziwei Qin, Heming Xia, Tian Feng, Shoujie Tong, Haoran Meng, Lin Xu, Zhongyu Wei, Weidong Zhan, Baobao Chang, Sujian Li, Tianyu Liu, Zhifang Sui

Why Can GPT Learn In-Context? Language Models Secretly Perform Gradient Descent as Meta-Optimizers (2023.findings-acl)

Challenge:	Existing work in vision language cross-modal reasoning uses binary or multi-choice classification based on source image and textual query.
Approach:	They propose a task where a textual premise is the background presumption on each source image.
Outcome:	The proposed task is based on a dataset of 15,360 movie screenshots and human-curated premise templates from 6 pre-defined categories.

Copied to clipboard

Damai Dai, Yutao Sun, Li Dong, Yaru Hao, Shuming Ma, Zhifang Sui, Furu Wei

An Anchor-Based Automatic Evaluation Metric for Document Summarization (2020.coling-main)

Challenge:	Large pretrained language models have shown surprising in-context learning ability . despite the great success in performance, its working mechanism remains unclear .
Approach:	They explain language models as meta-optimizers and understand in-context learning as implicit finetuning . they find that Transformer attention has a dual form of gradient descent .
Outcome:	The proposed model can predict labels for unseen inputs without parameter updates . the proposed model outperforms smaller models with a single parameter update .

Copied to clipboard

Kexiang Wang, Tianyu Liu, Baobao Chang, Zhifang Sui

Text Generation

Challenge:	Existing reference-based evaluation metrics such as ROUGE have their own drawbacks.
Approach:	They propose a protocol for a reference-based automatic evaluation metric that requires the endorsement of source document.
Outcome:	The proposed metric is anchored on source document and has higher correlation with human judgments.

A Two-Stream AMR-enhanced Model for Document-level Event Argument Extraction (2022.naacl-main)

Copied to clipboard

Runxin Xu, Peiyi Wang, Tianyu Liu, Shuang Zeng, Baobao Chang, Zhifang Sui

HypoNLI: Exploring the Artificial Patterns of Hypothesis-only Bias in Natural Language Inference (2020.lrec-1)

Challenge:	Existing studies aim at extracting event arguments from a single sentence . document-level event extraction still remains under-explored .
Approach:	They propose a two-stream abstract meaning representation enhanced extraction model to extract event arguments from an entire document.
Outcome:	The proposed model outperforms state-of-the-art in extracting event arguments from documents by 2.54 F1 and 5.13 F1 on public RAMS and WikiEvents datasets.

Copied to clipboard

Tianyu Liu, Zheng Xin, Baobao Chang, Zhifang Sui

StableMoE: Stable Routing Strategy for Mixture of Experts (2022.acl-long)

Challenge:	Recent studies have shown that for models trained on datasets for natural language inference (NLI), it is possible to make correct predictions by looking at the hypothesis while completely ignoring the premise.
Approach:	They propose to derive adversarial examples in terms of the hypothesis-only bias and explore eligible ways to mitigate such bias.
Outcome:	The proposed models can be used to mitigate the hypothesis-only bias by using down-sampling and adversarial training.

Copied to clipboard

Damai Dai, Li Dong, Shuming Ma, Bo Zheng, Zhifang Sui, Baobao Chang, Furu Wei

Challenge:	Existing learning-to-route methods suffer from the routing fluctuation issue . with the model scale growing, training speed will go slower and memory requirements are heavy .
Approach:	They propose a Mixture-of-Experts technique that can scale up the model size of Transformers with an affordable computational overhead.
Outcome:	The proposed method outperforms existing learning-to-route methods on language modeling and multilingual machine translation.

A Spectral Method for Unsupervised Multi-Document Summarization (2020.emnlp-main)

Copied to clipboard

Kexiang Wang, Baobao Chang, Zhifang Sui

Text Generation

Challenge:	a spectral-based hypothesis is proposed for the unsupervised task of multi-document summarization.
Approach:	They propose a spectral-based hypothesis that a summary candidate's spectral impact is closely linked to its spectre.
Outcome:	The proposed method has a competitive result compared to state-of-the-art systems.

Calibrating Factual Knowledge in Pretrained Language Models (2022.findings-emnlp)

Copied to clipboard

Qingxiu Dong, Damai Dai, Yifan Song, Jingjing Xu, Zhifang Sui, Lei Li

A Survey on In-context Learning (2024.emnlp-main)

Challenge:	Existing studies show that Pretrained Language Models can store factual knowledge, but facts stored in PLMs are not always correct.
Approach:	They propose a lightweight method to calibrate factual knowledge in PLMs without re-training from scratch.
Outcome:	The proposed method can be used to calibrate factual knowledge in PLMs without re-training from scratch.

Copied to clipboard

Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Baobao Chang, Xu Sun, Lei Li, Zhifang Sui

Low-Resource NLP

Challenge:	In-context learning (ICL) is a new paradigm for natural language processing . large language models (LLMs) demonstrate the ability to learn from a few examples .
Approach:	They propose to explore ICL to evaluate and extrapolate the ability of large language models.
Outcome:	The proposed methods can be used to evaluate and extrapolate the ability of large language models.

Achilles-Bench: A Challenging Benchmark for Low-Resource Evaluation (2024.findings-acl)

Copied to clipboard

Yudong Wang, Chang Ma, Qingxiu Dong, Zhifang Sui, Lingpeng Kong, Jingjing Xu

A Token-level Reference-free Hallucination Detection Benchmark for Free-form Text Generation (2022.acl-long)

Challenge:	Existing low-resource datasets that challenge neural networks cause over-estimated performance, despite promising yet saturated results in high-res settings.
Approach:	They propose a benchmark Achilles-Bench to better evaluate the learning ability of neural networks in low-resource settings.
Outcome:	The proposed benchmarks show that even pre-trained language models show performance drops on NLP tasks.

Copied to clipboard

Tianyu Liu, Yizhe Zhang, Chris Brockett, Yi Mao, Zhifang Sui, Weizhu Chen, Bill Dolan

Reasoning

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 .

Can Large Multimodal Models Uncover Deep Semantics Behind Images? (2024.findings-acl)

Copied to clipboard

Yixin Yang, Zheng Li, Qingxiu Dong, Heming Xia, Zhifang Sui

HistLens: Mapping Idea Change across Concepts and Corpora (2026.acl-long)

Challenge:	Existing studies on visual deep semantics focus primarily on superficial description of images, revealing a notable deficiency in the systematic investigation of the inherent deep semantic.
Approach:	They propose a benchmark to assess Large Multimodal Models’ (LMMs) capacities of visual deep semantics.
Outcome:	The proposed benchmark demonstrates a substantial gap between the deep semantic comprehension capabilities of existing LMMs and humans.

Copied to clipboard

Yi Jing, Weiyun Qiu, Yihang Peng, Zhifang Sui

Explainability & Interpretability in NLP

Challenge:	Existing approaches to diachronic semantics and discourse analysis focus on a single concept or corpus, argues a new paper.
Approach:	They propose a framework for multi-concept, multi-corpus conceptual-history analysis that decomposes concept representations into interpretable features and tracks activation dynamics over time and across sources.
Outcome:	The proposed framework decomposes concept representations into interpretable features and tracks their activation dynamics over time and across sources.

Towards Harmonized Uncertainty Estimation for Large Language Models (2025.acl-long)

Copied to clipboard

Rui Li, Jing Long, Muge Qi, Heming Xia, Lei Sha, Peiyi Wang, Zhifang Sui

ATP: AMRize Then Parse! Enhancing AMR Parsing with PseudoAMRs (2022.findings-naacl)

Challenge:	Large language models (LLMs) have demonstrated exceptional capabilities in handling a wide range of downstream tasks.
Approach:	They propose a method that employs a lightweight model trained on data aligned with the target LLM’s performance to adjust uncertainty scores.
Outcome:	The proposed method achieves improvements of up to 60% over existing methods.

Copied to clipboard

Liang Chen, Peiyi Wang, Runxin Xu, Tianyu Liu, Zhifang Sui, Baobao Chang

A Probabilistic Inference Scaling Theory for LLM Self-Correction (2025.emnlp-main)

Challenge:	Abstract Meaning Representation (AMR) implicitly involves compound semantic annotations.
Approach:	They propose to use auxiliary tasks which are semantically or formally related to enhance AMR parsing.
Outcome:	The proposed method achieves state-of-the-art performance on benchmarks especially in topology-related scores.

Copied to clipboard

Zhe Yang, Yichang Zhang, Yudong Wang, Ziyao Xu, Junyang Lin, Zhifang Sui

Linguistic Theories

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

Di Zhang, Xun Wu, Shaohan Huang, Lingjie Jiang, Yaru Hao, Li Dong, Zewen Chi, Zhifang Sui, Furu Wei

Reasoning

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.

Not All Demonstration Examples are Equally Beneficial: Reweighting Demonstration Examples for In-Context Learning (2023.findings-emnlp)

Copied to clipboard

Zhe Yang, Damai Dai, Peiyi Wang, Zhifang Sui

An Enhanced Span-based Decomposition Method for Few-Shot Sequence Labeling (2022.naacl-main)

Challenge:	Large Language Models (LLMs) have recently gained the In-Context Learning ability . however, the quality of demonstration examples is usually uneven .
Approach:	They propose to determine optimal weights for demonstration examples and apply them during ICL.
Outcome:	The proposed approach outperforms conventional ICL on 8 classification tasks.

Copied to clipboard

Peiyi Wang, Runxin Xu, Tianyu Liu, Qingyu Zhou, Yunbo Cao, Baobao Chang, Zhifang Sui

Unlocking Efficiency in Large Language Model Inference: A Comprehensive Survey of Speculative Decoding (2024.findings-acl)

Challenge:	Existing approaches to tagging tasks are limited to predefined classes and require large-scale annotated data.
Approach:	They propose an Enhanced Span-based Decomposition method for Few-Shot Sequence Labeling to generalize on emerging, resource-scare domains.
Outcome:	The proposed method achieves state-of-the-art results on two popular FSSL benchmarks, FewNERD and SNIPS, and is more robust in noisy and nested tagging scenarios.

Copied to clipboard

Heming Xia, Zhe Yang, Qingxiu Dong, Peiyi Wang, Yongqi Li, Tao Ge, Tianyu Liu, Wenjie Li, Zhifang Sui

Guiding AMR Parsing with Reverse Graph Linearization (2023.findings-emnlp)

Challenge:	Large Language Models (LLMs) have a high inference latency stemming from autoregressive decoding.
Approach:	They propose a novel decoding paradigm that drafts multiple tokens and verifies them in parallel . they aim to provide a catalyst for further research on Speculative Decoding .
Outcome:	The proposed method drafts multiple tokens and verifies them in parallel . it can be used to accelerate inference in large language models.

Copied to clipboard

Bofei Gao, Liang Chen, Peiyi Wang, Zhifang Sui, Baobao Chang

Learning to Control the Fine-grained Sentiment for Story Ending Generation (P19-1)

Challenge:	Abstract Meaning Representation (AMR) parsing aims to extract an abstract semantic graph from a sentence.
Approach:	They propose a new framework that allows for reversed linearization of AMR graphs . they propose to combine sequence-to-sequence approaches with a linearized graph .
Outcome:	The proposed framework outperforms the best AMR parser by 0.8 and 0.5 Smatch scores on the AMR 2.0 and AMR 3.0 datasets.

Copied to clipboard

Fuli Luo, Damai Dai, Pengcheng Yang, Tianyu Liu, Baobao Chang, Zhifang Sui, Xu Sun

Sentiment Analysis

Challenge:	Existing studies focus on controlling the sentiment of story endings.
Approach:	They propose a generic and novel framework which controls fine-grained sentiment intensity for automatic story ending generation without manually annotating sentiment labels.
Outcome:	The proposed framework can generate story endings which meet the given sentiment intensity better.

Language Models Encode the Value of Numbers Linearly (2025.coling-main)

Copied to clipboard

Fangwei Zhu, Damai Dai, Zhifang Sui

EventWiki: A Knowledge Base of Major Events (L18-1)

Challenge:	Existing studies show that large language models encode the value of numbers linearly.
Approach:	They construct a large language model and use linear probes to read out input numbers from hidden states.
Outcome:	The proposed model encodes the value of numbers linearly, and can store the outputs via simple vector additions.

Copied to clipboard

Tao Ge, Lei Cui, Baobao Chang, Zhifang Sui, Furu Wei, Ming Zhou

Knowledge Representation

Challenge:	Existing knowledge bases focus on static entities such as people, locations and organizations.
Approach:	They propose a new knowledge base resource called EventWiki which concentrates on major events . they show that EventWiki is a very useful resource for information extraction regarding events in NLP .
Outcome:	The proposed resource is the first knowledge base resource of major events.

Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotations (2024.acl-long)

Copied to clipboard

Peiyi Wang, Lei Li, Zhihong Shao, Runxin Xu, Damai Dai, Yifei Li, Deli Chen, Yu Wu, Zhifang Sui

Numerical Reasoning

Challenge:	Existing methods for process-oriented math reward models rely on manual annotation.
Approach:	They propose a process-oriented math process reward model called Math-shepherd which assigns a reward score to each step of math problem solutions.
Outcome:	The proposed model breaks the bottleneck of manual supervision in two scenarios.

How Far are LLMs from Being Our Digital Twins? A Benchmark for Persona-Based Behavior Chain Simulation (2025.findings-acl)

Copied to clipboard

Rui Li, Heming Xia, Xinfeng Yuan, Qingxiu Dong, Lei Sha, Wenjie Li, Zhifang Sui

Challenge:	Recent studies have focused on dialogue simulation while overlooking human behavior simulation, which is crucial for digital twins.
Approach:	They propose to integrate persona metadata into LLMs and use it to iteratively infer contextually appropriate behaviors within dynamic scenarios.
Outcome:	The proposed model is based on 15,846 distinct behaviors across 1,001 unique personas and incorporates persona metadata to iteratively infer appropriate behaviors within dynamic scenarios.

Fine-grained Coordinated Cross-lingual Text Stream Alignment for Endless Language Knowledge Acquisition (D18-1)

Copied to clipboard

Tao Ge, Qing Dou, Heng Ji, Lei Cui, Baobao Chang, Zhifang Sui, Furu Wei, Ming Zhou

Multilinguality

Challenge:	a novel information network decipherment paradigm is proposed for fine-grained coordinated cross-lingual text stream alignment.
Approach:	They propose to use Burst Information Networks as media to represent text streams . they propose a simple yet effective information network decipherment algorithm with diverse clues .
Outcome:	The proposed approach outperforms existing approaches on bilingual lexicon extraction from coordinated text streams and can harvest high-quality alignments from large amounts of streaming data for endless language knowledge mining.

HAUNTATTACK: When Attack Follows Reasoning as a Shadow (2026.findings-acl)

Copied to clipboard

Jingyuan Ma, Rui Li, Zheng Li, Junfeng Liu, Heming Xia, Lei Sha, Zhifang Sui

Speculative Decoding: Exploiting Speculative Execution for Accelerating Seq2seq Generation (2023.findings-emnlp)

Challenge:	Emerging Large Reasoning Models (LRMs) excel in mathematical and reasoning tasks, showcasing remarkable capabilities.
Approach:	They propose a framework that embeds harmful instructions into reasoning questions . they evaluate 11 LRMs and observe an average attack success rate of over 70% .
Outcome:	The proposed framework improves reasoning models by 13 percentage points over baseline.

Copied to clipboard

Heming Xia, Tao Ge, Peiyi Wang, Si-Qing Chen, Furu Wei, Zhifang Sui

Green & Sustainable NLP

Challenge:	Experimental results show draft-then-verify paradigm can achieve around 5x speedup for the popular Transformer architectures with comparable generation quality to beam search decoding.
Approach:	They propose to use Spec-Drafter and Spec Verification to accelerate autoregressive (AR) decoding by combining a model optimized for efficient and accurate drafting and a reliable method for verifying the drafted tokens efficiently.
Outcome:	The proposed method achieves 5x speedup on seq2seq tasks with comparable generation quality to beam search decoding, refreshing the impression that draft-then-verify paradigm introduces only 1.4x2x speed up.

Incorporating Glosses into Neural Word Sense Disambiguation (P18-1)

Copied to clipboard

Fuli Luo, Tianyu Liu, Qiaolin Xia, Baobao Chang, Zhifang Sui

Be a Multitude to Itself: A Prompt Evolution Framework for Red Teaming (2024.findings-emnlp)

Challenge:	Existing neural networks for Word Sense Disambiguation rely on labeled data and lexical knowledge.
Approach:	They propose a gloss-augmented WSD neural network which integrates context and glosses of the target word into a unified framework.
Outcome:	The proposed model outperforms the state-of-the-art systems on several English all-words WSD datasets.

Copied to clipboard

Rui Li, Peiyi Wang, Jingyuan Ma, Di Zhang, Lei Sha, Zhifang Sui

CBLUE: A Chinese Biomedical Language Understanding Evaluation Benchmark (2022.acl-long)

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.

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.

Decompose, Fuse and Generate: A Formation-Informed Method for Chinese Definition Generation (2021.naacl-main)

Copied to clipboard

Hua Zheng, Damai Dai, Lei Li, Tianyu Liu, Zhifang Sui, Baobao Chang, Yang Liu

Challenge:	Existing definition generation methods take the source word as an indecomposable semantic unit, but in parataxis languages like Chinese, word meanings can be composed using the word formation process.
Approach:	They propose to use word formation features to enhance Definition Generation (DG) in Chinese to generate an explanatory text.
Outcome:	The proposed model enhances Definition Generation (DG) in Chinese by decomposing the word meaning into different semantic components.

FaGANet: An Evidence-Based Fact-Checking Model with Integrated Encoder Leveraging Contextual Information (2024.lrec-main)

Copied to clipboard

Weiyao Luo, Junfeng Ran, Zailong Tian, Sujian Li, Zhifang Sui

Leveraging Gloss Knowledge in Neural Word Sense Disambiguation by Hierarchical Co-Attention (D18-1)

Challenge:	Existing evidence-based fact-checking efforts are time-consuming and challenging . however, relying on surface patterns of claims makes it difficult to identify subtle connections between claims and evidence.
Approach:	They propose a model that leverages sentence-level attention and graph attention network to enhance accuracy and fusing claims and evidence information for accurate identification of even well-disguised data.
Outcome:	The proposed model improves accuracy and state-of-the-art in the evidence-based fact-checking task.

Copied to clipboard

Fuli Luo, Tianyu Liu, Zexue He, Qiaolin Xia, Zhifang Sui, Baobao Chang

Learn to Not Link: Exploring NIL Prediction in Entity Linking (2023.findings-acl)

Challenge:	Existing models for Word Sense Disambiguation use labeled data, but lack gloss knowledge.
Approach:	They propose a co-attention mechanism to generate co-dependent representations for context and gloss . they propose to incorporate gloss knowledge into neural networks for Word Sense Disambiguation .
Outcome:	The proposed model achieves state-of-the-art results on standard English all-words WSD datasets.

Copied to clipboard

Fangwei Zhu, Jifan Yu, Hailong Jin, Lei Hou, Juanzi Li, Zhifang Sui

Taking a Deep Breath: Enhancing Language Modeling of Large Language Models with Sentinel Tokens (2024.findings-emnlp)

Challenge:	Entity linking models have been successful in capturing semantic features, but the NIL prediction problem has not been addressed.
Approach:	They propose an entity linking dataset that categorizes mentions linking to NIL into Missing Entity and Non-Entity Phrases.
Outcome:	The proposed dataset categorizes mentions linking to NIL into Missing Entity and Non-Entity Phrase categories and ensures the presence of mentions by human annotation and entity masking.

Copied to clipboard

Weiyao Luo, Suncong Zheng, Heming Xia, Weikang Wang, Yan Lei, Tianyu Liu, Shuang Chen, Zhifang Sui