Papers by Zheyuan Zhang
Do LLMs Catch Their Own Mistakes? A Comprehensive Benchmark for Reflective Tool Use LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing benchmarks primarily evaluate planning and execution success, overlooking the self-reflective dimension of tool use. |
| Approach: | They propose a benchmark to assess LLMs’ self-reflective reasoning in tool-augmented multi-turn dialogues. |
| Outcome: | The proposed benchmark covers 10 domains with 88 distinct APIs and 968 annotated dialogues, systematically injecting diverse error types arising from both user and assistant behavior. |
Interpretable Graph-Language Modeling for Detecting Youth Illicit Drug Use (2026.findings-eacl)
Copied to clipboard
Yiyang Li, Zehong Wang, Zhengqing Yuan, Zheyuan Zhang, Keerthiram Murugesan, Chuxu Zhang, Yanfang Ye
| Challenge: | Illicit drug use among teens and young adults remains a public health concern . existing models ignore latent and interconnected structures among survey variables . |
| Approach: | They propose a joint graph-language modeling framework to detect illicit drug use among TYAs . they use large-scale surveys such as the Youth Risk Behavior Survey and the National Survey on Drug Use and Health to analyze data . |
| Outcome: | The proposed framework outperforms baseline models on YRBS and NSDUH datasets in predictive accuracy. |
From Heuristic to Analytic: Cognitively Motivated Strategies for Coherent Physical Commonsense Reasoning (2023.emnlp-main)
Copied to clipboard
| Challenge: | Pre-trained language models have shown impressive performance in various language tasks, but are prone to spurious correlations and illusory information. |
| Approach: | They propose to use pre-trained language models to justify decisions with formalized, coherent reasoning chains. |
| Outcome: | The proposed strategies improve coherence of rationalizations yielding state-of-the-art results on Tiered Reasoning for Intuitive Physics (TRIP). |
Exploring the Cognitive Knowledge Structure of Large Language Models: An Educational Diagnostic Assessment Approach (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies on LLMs evaluation with exams are lacking in cognitive research on their overall knowledge structure. |
| Approach: | They conduct an evaluation using a human test dataset based on Bloom Taxonomy to reveal the knowledge structures of Large Language Models and gain insights of their cognitive capabilities. |
| Outcome: | The proposed model can pass AP, SAT, and Leetcode exams, but lacks the cognitive power to perform on human exams. |
Instant Personalized Large Language Model Adaptation via Hypernetwork (2026.acl-long)
Copied to clipboard
Zhaoxuan Tan, Zixuan Zhang, Haoyang Wen, Zheng Li, Rongzhi Zhang, Pei Chen, Fengran Mo, Zheyuan Liu, Qingkai Zeng, Qingyu Yin, Meng Jiang
| Challenge: | Existing parameter-efficient fine-tuning methods require training a separate adapter for each user, making them computationally expensive and impractical for real-time updates. |
| Approach: | They propose a scalable framework that maps a user's profile directly to a full set of adapter parameters. |
| Outcome: | The proposed framework outperforms prompt-based personalization and OPPU while using substantially fewer computational resources at deployment. |
AgentRouter: A Knowledge-Graph-Guided LLM Router for Collaborative Multi-Agent Question Answering (2026.acl-long)
Copied to clipboard
Zheyuan Zhang, Kaiwen Shi, Zhengqing Yuan, Zehong Wang, Tianyi Ma, Keerthiram Murugesan, Vincent Galassi, Chuxu Zhang, Yanfang Ye
| Challenge: | Existing approaches to agent routing emphasize cost efficiency while overlooking the fine-grained contextual and relational structure inherent in QA tasks. |
| Approach: | They propose a framework that formulates multi-agent QA as a knowledge-graph-guided routing problem supervised by empirical performance signals. |
| Outcome: | The proposed framework outperforms single-agent and ensemble baselines while generalizing across benchmarks and LLM backbones. |
AutoRubric: Rubric-Based Generative Rewards for Faithful Multimodal Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Multimodal large language models (MLLMs) have advanced from perception tasks to complex multi-step reasoning. |
| Approach: | They propose a framework that integrates reinforcement learning with verifiable rewards with process-level supervision through automatically collected rubric-based generative rewards. |
| Outcome: | The proposed framework achieves state-of-the-art performance on six multimodal reasoning benchmarks and significantly improves reasoning faithfulness in dedicated evaluations. |
NGQA: A Nutritional Graph Question Answering Benchmark for Personalized Health-aware Nutritional Reasoning (2025.acl-long)
Copied to clipboard
Zheyuan Zhang, Yiyang Li, Nhi Ha Lan Le, Zehong Wang, Tianyi Ma, Vincent Galassi, Keerthiram Murugesan, Nuno Moniz, Werner Geyer, Nitesh V Chawla, Chuxu Zhang, Yanfang Ye
| Challenge: | Diet plays a critical role in human health, but tailoring dietary reasoning to individual health conditions remains a challenge. |
| Approach: | a new benchmark evaluates dietary reasoning using a national health survey data set. |
| Outcome: | The NGQA benchmark evaluates dietary reasoning across three tasks using a set of question complexity settings and baseline models. |
A Combinatorial Approach to Neural Emergent Communication (2025.coling-main)
Copied to clipboard
| Challenge: | Existing research on emergent communication uses the Lewis signaling game . however, the training data is limited and the messages are often ineffective . |
| Approach: | They propose a combinatorial algorithm to solve the symbolic complexity for classification, which is the minimum number of symbols in the message for successful communication. |
| Outcome: | The proposed algorithm increases the number of effective symbols in the emergent language. |
Knowing More, Acting Better: Hierarchical Representation for Embodied Decision-Making (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Modern embodied AI uses multimodal large language models as policy models, predicting actions from final-layer hidden states. |
| Approach: | They propose a hierarchical action probing method that aggregates representations from all layers, mirroring the brain's multi-level organization. |
| Outcome: | Experiments show that hierarchical probing improves on last-layer embodied models and achieves a 46.6% success rate and a 62.5% gain in spatial reasoning tasks. |
LLM-Empowered Class Imbalanced Graph Prompt Learning for Online Drug Trafficking Detection (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to combat illicit drug trafficking are impractical due to the scarcity of labeled samples and imbalance of classes. |
| Approach: | They propose a Large Language Model-empowered Heterogeneous Graph Prompt Learning framework for illicit drug trafficking detection that leverages LLM to facilitate heterogeneous graph neural networks to effectively identify minority classes. |
| Outcome: | The proposed framework is able to identify minority classes in class-imbalanced scenarios. |
Modality-Aware Neuron Pruning for Unlearning in Multimodal Large Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | Large Language Models and Multimodal Large Language Modells can memorize sensitive information, raising ethical and privacy concerns. |
| Approach: | They propose a novel unlearning framework that selectively clips neurons based on their relative importance to the targeted forget data. |
| Outcome: | The proposed framework selectively clips neurons based on their relative importance to the targeted forget data, curated for different modalities. |
Eliciting In-Context Learning in Vision-Language Models for Videos Through Curated Data Distributional Properties (2024.emnlp-main)
Copied to clipboard
| Challenge: | Emergent In-context Learning on Videos induces in-contact learning over video and text . eILeV-trained models outperform other off-the-shelf VLMs in few-shot video narration for novel, rare actions. |
| Approach: | They implement Emergent In-context Learning on Videos (EILeV) that induces in-contact learning over video and text by capturing key properties of pre-training data. |
| Outcome: | The proposed training paradigm outperforms off-the-shelf VLMs in few-shot video narration for novel, rare actions. |
MAPRO: Recasting Multi-Agent Prompt Optimization as Maximum a Posteriori Inference (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have demonstrated remarkable capabilities across diverse tasks. |
| Approach: | They propose a framework that optimizes MAS prompts as a maximum a posteriori problem and then iteratively updates agent prompts. |
| Outcome: | The proposed framework surpasses manual and automated benchmarks in multiple tasks and provides general guidelines for building more reliable and principled multi-agent systems in the future. |
Addressing Overthinking in Large Vision-Language Models via Gated Perception-Reasoning Optimization (2026.findings-acl)
Copied to clipboard
Xingjian Diao, Zheyuan Liu, Chunhui Zhang, Weiyi Wu, Keyi Kong, Lin Shi, Kaize Ding, Soroush Vosoughi, Jiang Gui
| Challenge: | Prior work has attempted to mitigate this issue by using adaptive reasoning strategies, but these methods overlook a fundamental bottleneck: visual perception failures. |
| Approach: | They propose a meta-reasoning controller that dynamically routes computation among three decision paths at each generation step. |
| Outcome: | The proposed method outperforms slow-thinking methods while producing shorter responses. |
Superficial Self-Improved Reasoners Benefit from Model Merging (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) rely heavily on large-scale reasoning data, but as data becomes scarce, model self-improvement offers a promising alternative. |
| Approach: | They propose to merge the weights of original and self-improved LLMs to mitigate model collapse and improve generalized reasoning capability. |
| Outcome: | The proposed model merge mitigates model collapse and improves generalized reasoning capability. |
EmoBench: Evaluating the Emotional Intelligence of Large Language Models (2024.acl-long)
Copied to clipboard
Sahand Sabour, Siyang Liu, Zheyuan Zhang, June Liu, Jinfeng Zhou, Alvionna Sunaryo, Tatia Lee, Rada Mihalcea, Minlie Huang
| Challenge: | Existing benchmarks for Emotional Intelligence (EI) focus on emotion recognition, neglecting essential EI capabilities. |
| Approach: | They propose a benchmark that proposes a comprehensive definition for machine EI . they propose 400 hand-crafted questions in English and Chinese to evaluate EI. |
| Outcome: | The proposed benchmarks focus on emotion recognition, neglecting EI capabilities . they are constructed from existing datasets, which include frequent patterns and errors . the proposed benchmark includes questions in English and Chinese that require thorough reasoning and understanding . |
NG-Router: Graph-Supervised Multi-Agent Collaboration for Nutrition Question Answering (2026.eacl-long)
Copied to clipboard
Kaiwen Shi, Zheyuan Zhang, Zhengqing Yuan, Keerthiram Murugesan, Vincent Galassi, Chuxu Zhang, Yanfang Ye
| Challenge: | Existing methods for nutrition question answering face limited reasoning capacity and contextual overload . poor dietary patterns are associated with more than 11 million deaths in 2017 . |
| Approach: | They propose a framework that enables supervised multi-agent collaboration for nutritional QA. |
| Outcome: | The proposed framework outperforms single-agent and ensemble baselines in multi-agency reasoning tasks. |
Can LLMs Convert Graphs to Text-Attributed Graphs? (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing approaches to model graph-structured data are limited by the availability of text-attributed graph data. |
| Approach: | They propose a method to convert existing graphs into text-attributed graphs using large language models. |
| Outcome: | The proposed method outperforms existing approaches that manually design node features on text-free graphs. |
Simulating Classroom Education with LLM-Empowered Agents (2025.naacl-long)
Copied to clipboard
Zheyuan Zhang, Daniel Zhang-Li, Jifan Yu, Linlu Gong, Jinchang Zhou, Zhanxin Hao, Jianxiao Jiang, Jie Cao, Huiqin Liu, Zhiyuan Liu, Lei Hou, Juanzi Li
| Challenge: | Initial studies have focused on task-specific, independent LLM-empowered agents, but the potential of LLMs within a multi-agent collaborative framework for classroom simulation with real user participation remains unexplored. |
| Approach: | They propose a multi-agent classroom simulation teaching framework that recognizes representative class roles and introduces a novel class control mechanism for automatic classroom teaching. |
| Outcome: | The proposed framework can simulate dynamic learning environment for users with active teacher-student and student-studente interactions. |
Behavior Knowledge Merge in Reinforced Agentic Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for supervised fine-tuning (SFT) are suboptimal to preserve task-specific capabilities on RL-trained agentic models. |
| Approach: | They propose a distribution-aware merging framework specifically designed for RL-trained agentic models that disentangles shared and task-specific unique parameter updates while selectively preserving and rescaling unique ones. |
| Outcome: | Experiments across multiple agent domains and model architectures show that the proposed framework surpasses baselines and unlocks synergistic potential among agents. |
Transparent and Coherent Procedural Mistake Detection (2025.emnlp-main)
Copied to clipboard
| Challenge: | Procedural mistake detection (PMD) is a problem of classifying whether a human user has successfully executed a task. |
| Approach: | They extend PMD to require generating visual self-dialog rationales to inform decisions . they leverage a natural language inference model to formulate two automated metrics for coherence of generated rationale. |
| Outcome: | The proposed model improves on a reframed task with a natural language inference model and a multi-faceted metrics visualization of common outcomes. |
Explaining Length Bias in LLM-Based Preference Evaluations (2025.findings-emnlp)
Copied to clipboard
Zhengyu Hu, Linxin Song, Jieyu Zhang, Zheyuan Xiao, Tianfu Wang, Zhengyu Chen, Nicholas Jing Yuan, Jianxun Lian, Kaize Ding, Hui Xiong
| Challenge: | a preference evaluation metric is often biased towards longer responses, revealing a reliability problem . a decomposition of the preference evaluation into two components is needed to understand this bias. |
| Approach: | They propose to decompose the preference evaluation metric into two key components . the first component is length-dependent and related to trustworthiness . |
| Outcome: | The proposed evaluation metric is based on two components: desirability and information mass. |