Papers by Yicheng Zhang

34 papers
Capability Salience Vector: Fine-grained Alignment of Loss and Capabilities for Downstream Task Scaling Law (2025.acl-long)

Copied to clipboard

Challenge: Large language models have demonstrated impressive performance across a wide range of tasks, but this achievement comes with the trade-off of significant computational demands.
Approach: They propose a scaling law that decomposes the overall validation loss and assigns different importance weights to tokens to assess a specific meta-capability.
Outcome: The proposed model can predict the loss trending of models across different levels of computation without a gap between validation loss and model's downstream capabilities.
Efficient Inference for Large Vision-Language Models: Bottlenecks, Techniques, and Prospects (2026.findings-acl)

Copied to clipboard

Challenge: Large Vision-Language Models are hindered by a systemic efficiency barrier known as visual token dominance.
Approach: They propose a systematic taxonomy of efficiency techniques structured around the inference lifecycle . they examine visual encoding, prefilling, and decoding to understand bottlenecks .
Outcome: The proposed techniques reveal how upstream decisions dictate downstream bottlenecks . the proposed techniques include hybrid compression and modality-aware decoding .
MINER: Improving Out-of-Vocabulary Named Entity Recognition from an Information Theoretic Perspective (2022.acl-long)

Copied to clipboard

Challenge: Named Entity Recognition models are feature-engineering and machine learning based.
Approach: They propose a new NER learning framework that uses entity mentions to improve model performance.
Outcome: The proposed model achieves better performance on OOV entities on various settings and datasets.
A Lexicon-Based Graph Neural Network for Chinese NER (D19-1)

Copied to clipboard

Challenge: Chinese named entity recognition models are vulnerable to word ambiguities due to the lack of global semantics and chain structure.
Approach: They propose a lexicon-based graph neural network with global semantics to solve word ambiguities in Chinese named entity recognition (NER) Lexicons are used to construct the graph and provide word-level features.
Outcome: The proposed model improves on four NER datasets on Chinese characters, potential words, and the whole-sentence semantics.
Are LLM-based Evaluators Confusing NLG Quality Criteria? (2024.acl-long)

Copied to clipboard

Challenge: Existing studies show that LLMs confuse evaluation criteria, which reduces their reliability.
Approach: They propose a hierarchical classification system for 11 common aspects with corresponding different evaluation criteria.
Outcome: The proposed system is based on 11 common aspects with different evaluation criteria.
AgentGym: Evaluating and Training Large Language Model-based Agents across Diverse Environments (2025.acl-long)

Copied to clipboard

Challenge: Large language models (LLMs) are promising foundations to build generally-capable agents . however, the community lacks a unified interactive framework that covers diverse environments for comprehensive evaluation of agents.
Approach: They propose a framework that features 7 real-world scenarios, 14 environments, and 89 tasks for unified, real-time, and concurrent agent interaction.
Outcome: The proposed framework features 7 real-world scenarios, 14 environments, and 89 tasks for unified, real-time, and concurrent agent interaction.
AdapterDistillation: Non-Destructive Task Composition with Knowledge Distillation (2023.emnlp-industry)

Copied to clipboard

Challenge: Recent work on learning from multiple tasks has shown that adding an extra fusion layer to implement knowledge composition is non-scalable for some applications.
Approach: They propose a two-stage knowledge distillation algorithm to extract task specific knowledge by using local data to train a student adapter.
Outcome: Experiments on frequently asked question retrieval in task-oriented dialog systems validate the efficiency of AdapterDistillation.
AI Agents for the Science of Science: A Survey of Tasks, Architectures, Evaluations, and Challenges (2026.findings-acl)

Copied to clipboard

Challenge: The Science of Science (SciSc) examines how scientific knowledge is produced, evaluated, and transformed by utilizing large-scale scholarly and bibliometric data.
Approach: They propose a task-centered taxonomy for AI agents that model citations, collaborations, and community dynamics.
Outcome: The proposed taxonomy distinguishes agents as simulations from tools for empirical analysis and scientific workflows.
RealBench: A Chinese Multi-image Understanding Benchmark Close to Real-world Scenarios (2025.findings-emnlp)

Copied to clipboard

Challenge: RealBench is the first Chinese multimodal multi-image dataset . the dataset contains 9393 samples and 69910 images .
Approach: They propose to create a Chinese multimodal multi-image dataset using 21 models . they use closed-source models that support multi-inputs as well as open-source visual and video models a .
Outcome: The first Chinese multimodal multi-image dataset contains 9393 samples and 69910 images.
Towards Understanding Omission in Dialogue Summarization (2023.acl-long)

Copied to clipboard

Challenge: Existing methods for dialogue summarization are far from satisfactory . omission is a major factor in affecting the quality of summarizing, but few studies have explored the problem .
Approach: They propose a dataset that provides high-quality omission labels for dialogue summarization . they propose to use this dataset to detect omitted dialogue utterances .
Outcome: The proposed dataset improves summarization quality by providing ground-truth omission labels . the proposed dataset and codes are publicly available .
Making Parameter-efficient Tuning More Efficient: A Unified Framework for Classification Tasks (2022.coling-1)

Copied to clipboard

Challenge: Large pre-trained language models (PLMs) have demonstrated superior performance in industrial applications.
Approach: They propose a framework that re-uses existing parameter-efficient methods with a unified classifier.
Outcome: The proposed framework improves the efficiency of existing parameter-efficient methods with a unified classifier.
Beyond Single-Value Metrics: Evaluating and Enhancing LLM Unlearning with Cognitive Diagnosis (2025.findings-acl)

Copied to clipboard

Challenge: Large Language Models (LLMs) have been used to remove harmful knowledge and undesirable capabilities.
Approach: They propose a framework that leverages Cognitive Diagnosis Modeling to evaluate LLM unlearning.
Outcome: The proposed framework enhances evaluation and facilitates removal of harmful abilities.
SpecVLM: Enhancing Speculative Decoding of Video LLMs via Verifier-Guided Token Pruning (2025.emnlp-main)

Copied to clipboard

Challenge: Video large language models (Vid-LLMs) rely on dense video token representations and require substantial memory and computational overhead in both prefilling and decoding.
Approach: They propose a training-free speculative decoding framework that prunes up to 90% of video tokens to enable efficient speculation without sacrificing accuracy.
Outcome: The proposed framework achieves 2.68 speedup on LLaVA-OneVision-72B and 2.11 speed up on Qwen2.5-VL-32B.
Teaching Small Language Models Reasoning through Counterfactual Distillation (2024.emnlp-main)

Copied to clipboard

Challenge: Large language models (LLMs) have demonstrated remarkable performance in a wide range of downstream tasks.
Approach: They propose a counterfactual distillation framework that leverages LLMs to generate high-quality counterfacts and utilizes multi-view CoT to enhance the diversity of reasoning samples.
Outcome: The proposed framework enhances reasoning capabilities of large language models and is more robust to OOD data.
Rethinking Multiple-Choice Questions for RLVR: Unlocking Potential via Distractor Design (2026.findings-acl)

Copied to clipboard

Challenge: Existing approaches to RLVR use multiple-choice questions as verifiable rewards . however, not all tasks provide reliable verification .
Approach: They propose a framework that actively constructs high-quality distractors to block elimination shortcuts and promote deep reasoning.
Outcome: The proposed method significantly improves reasoning capabilities of Large Language Models.
Low-Resource Dialogue Summarization with Domain-Agnostic Multi-Source Pretraining (2021.emnlp-main)

Copied to clipboard

Challenge: Existing methods for low-resource dialogue summarization neglect the difference between dialogues and conventional articles.
Approach: They propose a multi-source pretraining paradigm to leverage external summary data . they exploit large-scale in-domain non-summary data to separate dialogue encoder and summary decoder .
Outcome: The proposed model can be used to better leverage external summary data.
Reverse Chain: A Generic-Rule for LLMs to Master Multi-API Planning (2024.findings-naacl)

Copied to clipboard

Challenge: Existing approaches to enable large language models to implement function calling are limited in their tool-use capabilities.
Approach: They propose a controllable, target-driven approach to empower LLMs to operate external APIs only via prompts.
Outcome: The proposed approach limits LLMs to executing simple tasks, e.g., API Selection and Argument Completion.
Optimizing Instruction Synthesis: Effective Exploration of Evolutionary Space with Tree Search (2024.findings-emnlp)

Copied to clipboard

Challenge: Extensive research has highlighted the quality of instruction data is essential for the success of this alignment.
Approach: They propose a framework for iteratively improving existing instruction data by using Monte Carlo tree search to find suitable prompts that align the language model to effectively learn multiple skills.
Outcome: The proposed framework improves the evaluation scores of seed instruction data, raising the average evaluation scores from 2.19 to 3.81.
Which Reasoning Trajectories Teach Students to Reason Better? A Simple Metric of Informative Alignment (2026.acl-long)

Copied to clipboard

Challenge: Existing methods assess suitability primarily through student likelihood, favoring trajectories that align closely with the student model’s current behavior but overlooking more informative ones.
Approach: They propose a Rank–Surprisal Ratio metric that captures both alignment and informativeness to assess the suitability of a reasoning trajectory.
Outcome: The proposed metric captures both alignment and informativeness to assess the suitability of a reasoning trajectory.
Thinking Clearly, Talking Fast: Concept-Guided Non-Autoregressive Generation for Open-Domain Dialogue Systems (2021.emnlp-main)

Copied to clipboard

Challenge: Existing models with seq2seq framework lack ability to effectively manage concept transitions . lack of concept management strategies might lead to incoherent dialogue due to loosely connected concepts .
Approach: They propose a concept-guided non-autoregressive model for open-domain dialogue generation that learns to identify multiple associated concepts from a conceptual graph and a customized Insertion Transformer to perform concept-directed generation to complete a response.
Outcome: The proposed model outperforms state-of-the-art models in automatic and human evaluations with substantially faster inference speed.
Would LLMs be Good Historical Linguists and Chinese Dialect Learners? (2026.acl-long)

Copied to clipboard

Challenge: Large language models struggle with low-resource Chinese dialects due to substantial phonological divergence.
Approach: They propose to incorporate Middle Chinese, the common historical ancestor of modern Chinese dialects, into LLMs to improve dialectal pronunciation modeling.
Outcome: The proposed approach improves on standard Chinese but struggles with low-resource Chinese dialects . the proposed model improves over baselines while revealing variation across dialects.
Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling and Collective Failure in Open-Ended Idea Generation (2026.findings-acl)

Copied to clipboard

Challenge: Multi-agent systems (MAS) are increasingly used for open-ended idea generation . when and why collective interaction expands the solution space remains unclear .
Approach: They propose to study diversity in multi-agent systems across three bottom-up levels: model intelligence, agent cognition, and system dynamics.
Outcome: The proposed model yields diminishing diversity despite higher quality . the proposed model fails to expand diversity and causes it to collapse .
FANNO: Augmenting High-Quality Instruction Data with Open-Sourced LLMs Only (2025.findings-acl)

Copied to clipboard

Challenge: Recent studies explore approaches to synthesize instruction data with open-sourced LLMs but require high-quality human-crafted seed data.
Approach: They propose an end-to-end framework to synthesize high-quality instruction data with open-sourced LLMs and sampled unlabeled documents, eliminating the need for seed data.
Outcome: The proposed framework synthesizes high-quality instruction data with open-sourced LLMs and sampled unlabeled documents, eliminating the need for seed data.
PwnGPT: Automatic Exploit Generation Based on Large Language Models (2025.acl-long)

Copied to clipboard

Challenge: Automated exploit generation (AEG) is the automatic discovery and exploitation of vulnerabilities against unknown targets.
Approach: They propose an automatic exploit generation framework that automatically solves pwn challenges by using large language models.
Outcome: The proposed framework improves the completion rate of exploits on the openAI o1-preview model and the GPT-4o model.
Increasing Visual Awareness in Multimodal Neural Machine Translation from an Information Theoretic Perspective (2022.emnlp-main)

Copied to clipboard

Challenge: Existing studies focus on extracting multi-granularity visual features for integration or designing model architectures for better message passing across various modalities.
Approach: They propose to decompose the informative visual signals into two parts: source-specific information and target-specific info.
Outcome: The proposed method can enhance the visual awareness of MMT models against strong baselines.
Learning Implicit Sentiment in Aspect-based Sentiment Analysis with Supervised Contrastive Pre-Training (2021.emnlp-main)

Copied to clipboard

Challenge: Recent studies have focused on identifying the sentiment polarity of aspects in product reviews.
Approach: They propose to use supervised Contrastive Pre-Training to learn implicit sentiment . they propose to train large-scale sentiment-annotated corpora from in-domain language resources .
Outcome: The proposed model achieves state-of-the-art performance on SemEval2014 benchmarks and comprehensively validates its effectiveness on learning implicit sentiment.
Towards Autonomous Tool Utilization in Language Models: A Unified, Efficient and Scalable Framework (2024.lrec-main)

Copied to clipboard

Challenge: Recent advances in tool learning for large language models have led to a new trend to allow LLMs to leverage external tools.
Approach: They propose a framework for fine-tuning language models that categorizes queries into three different types . they also introduce an "instruct, execute, and reformat" strategy specifically designed for efficient data annotation .
Outcome: The proposed framework surpasses open-source language models and GPT-3.5/4 on multiple evaluation metrics.
BalanceSFT: Improving LLM Function Calling with Balanced Training Signals and Data Hardness (2026.findings-acl)

Copied to clipboard

Challenge: Currently, Supervised Fine-Tuning (SFT) is the prevailing method for equipping Large Language Models (LLMs) with function calling capabilities, but its effectiveness is often compromised by two challenges: 1) lengthy Chain-of-Thought (CoT) reasoning tokens dominate training signals over concise function calls in the learning objective; 2) scarcity of hard training examples.
Approach: They propose a framework that uses a self-adjusted signal balancing loss and a hard data re-sampling strategy to selectively generate new, high-quality complex data guided by model errors.
Outcome: The proposed framework surpasses state-of-the-art models like GPT-5 in function calling performance.
Mixed Distillation Helps Smaller Language Models Reason Better (2024.findings-emnlp)

Copied to clipboard

Challenge: Recent large language models (LLMs) have demonstrated impressive multiple step-by-step reasoning capabilities in recent NLP reasoning tasks.
Approach: They propose a mixed distillation framework that distills multiple step-by-step reasoning abilities into smaller language models (SLMs) they leverage LLMs to generate multiple step by step reasoning rationales by sampling automatically.
Outcome: The proposed framework outperforms existing models on SVAMP, GSM8K and ASDIV, while a single model generated by MD exceeds the comprehensive performance of two individual CoT and PoT distilled models.
Divide and Conquer: Text Semantic Matching with Disentangled Keywords and Intents (2022.findings-acl)

Copied to clipboard

Challenge: Existing text semantic matching models do not provide granularity for text comparison.
Approach: They propose a simple yet effective training strategy for text semantic matching by disentangling keywords from intents.
Outcome: The proposed approach achieves stable performance improvements against a wide range of models on three benchmarks.
A Lexicon-Based Supervised Attention Model for Neural Sentiment Analysis (C18-1)

Copied to clipboard

Challenge: Existing attention models do not take full advantage of sentiment lexicons, which provide rich sentiment information and play a critical role in sentiment analysis.
Approach: They propose a lexicon-based supervised attention model which allows a neural network to focus on the sentiment content, thus generating sentiment-informative representations.
Outcome: The proposed model outperforms existing models on three large-scale sentiment classification datasets.
EgoNormia: Benchmarking Physical-Social Norm Understanding (2025.findings-acl)

Copied to clipboard

Challenge: Existing VLMs lack robust grounded norm understanding, a new study finds . current VLM models lack robust grounding, despite a high score for safety and privacy .
Approach: They propose a pipeline to generate grounded MCQs from ego-centric videos of human interactions.
Outcome: The proposed pipeline can generate grounded MCQs from egocentric video . it shows that current VLMs lack robust grounded norm understanding .
TextFlint: Unified Multilingual Robustness Evaluation Toolkit for Natural Language Processing (2021.acl-demo)

Copied to clipboard

Challenge: Existing approaches to textual robustness evaluation focus on slightly modifying the input data, which maintains the original meaning and results in a different prediction.
Approach: They propose a multilingual robustness evaluation toolkit for NLP that integrates universal text transformations, task-specific transformations and adversarial attack.
Outcome: The toolkit includes universal text transformation, task-specific transformation, adversarial attack, subpopulation, and their combinations to provide comprehensive robustness analyses.
See the Forest for the Trees: Loosely Speculative Decoding via Visual-Semantic Guidance for Efficient Inference of Video LLMs (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for video understanding suffer from autoregressive generation of tokens.
Approach: They propose a training-free loosely SD framework for Video-LLMs that uses visual-relevant tokens to accurately pinpoint the latter.
Outcome: The proposed framework boosts the accepted length and speedup ratio by 136% and 35% compared to SOTA training-free SD methods for Video-LLMs.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations