Papers by Zhaoxuan Tan
MTMCS-Bench: Evaluating Contextual Safety of Multimodal Large Language Models in Multi-Turn Dialogues (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing contextual safety benchmarks are mostly single-turn and miss how malicious intent can emerge gradually or how the same scene can support both benign and exploitative goals. |
| Approach: | They propose a benchmark that evaluates contextual safety in multimodal large language models . they observe persistent trade-offs between contextual safety and utility . |
| Outcome: | The proposed model combines multi-turn and multi-switch scenarios to evaluate safety in multimodal large language models. |
Aligning Large Language Models with Implicit Preferences from User-Generated Content (2025.acl-long)
Copied to clipboard
Zhaoxuan Tan, Zheng Li, Tianyi Liu, Haodong Wang, Hyokun Yun, Ming Zeng, Pei Chen, Zhihan Zhang, Yifan Gao, Ruijie Wang, Priyanka Nigam, Bing Yin, Meng Jiang
| Challenge: | Existing preference learning methods rely heavily on curated data from humans or advanced LLMs, which is costly and difficult to scale. |
| Approach: | They propose a framework that leverages implicit preferences in unlabeled user-generated content to generate preference data. |
| Outcome: | The proposed framework transforms user-generated content into user queries and generates responses from the policy model. |
Personalized Pieces: Efficient Personalized Large Language Models through Collaborative Efforts (2024.emnlp-main)
Copied to clipboard
| Challenge: | Experimental results show Per-Pcs outperforms non-personalized and PEFT retrieval baselines, offering performance comparable to OPPU with significantly lower resource use across six tasks. |
| Approach: | They propose a framework that allows users to safely share and assemble personalized large language models using their history data. |
| Outcome: | Experimental results show that Per-Pcs outperforms non-personalized and PEFT retrieval baselines with significantly lower resource use across six tasks. |
CodeTaxo: Enhancing Taxonomy Expansion with Limited Examples via Code Language Prompts (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing taxonomies are mainly constructed by experts or through crowd-sourcing, making the process time-consuming, labor-intensive, and restricted in coverage. |
| Approach: | They propose a method that leverages large language models to capture taxonomic structure . existing taxonomies are mainly constructed by experts or through crowd-sourcing . |
| Outcome: | Experiments on five real-world domains show that CodeTaxo outperforms state-of-the-art methods. |
Detecting Spoilers in Movie Reviews with External Movie Knowledge and User Networks (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing models focus on the textual content of the review, while spoiler detection requires putting the review into the context of facts and knowledge regarding movies. |
| Approach: | They propose a network-based spoiler detection model that takes into account external knowledge about movies and user activities on movie review platforms. |
| Outcome: | The proposed model takes into account external knowledge about movies and user activities on movie review platforms while incorporating user networks. |
IHEval: Evaluating Language Models on Following the Instruction Hierarchy (2025.naacl-long)
Copied to clipboard
Zhihan Zhang, Shiyang Li, Zixuan Zhang, Xin Liu, Haoming Jiang, Xianfeng Tang, Yifan Gao, Zheng Li, Haodong Wang, Zhaoxuan Tan, Yichuan Li, Qingyu Yin, Bing Yin, Meng Jiang
| Challenge: | Instruction-tuned language models (LMs) are increasingly deployed as interactive services across various applications. |
| Approach: | They propose a benchmark to evaluate models' ability to follow the instruction hierarchy by comparing their models to a set of benchmarks. |
| Outcome: | The proposed benchmark covers 3,538 examples across nine tasks covering cases where instructions in different priorities either align or conflict. |
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. |
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. |
Towards Safer Large Language Models through Machine Unlearning (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing work attempted to implement a gradient ascent based approach to prevent LLMs from producing harmful output when faced with problematic prompts. |
| Approach: | They propose a gradient ascent based approach to prevent LLMs from producing harmful output when faced with problematic prompts. |
| Outcome: | The proposed approach eliminates harmful knowledge while preserving utility on normal prompts. |
Enhancing Mathematical Reasoning in LLMs by Stepwise Correction (2025.acl-long)
Copied to clipboard
| Challenge: | Existing Best-of-N decoding methods often lead to incorrect solutions . a novel method is proposed to help large language models identify and revise incorrect steps in their generated reasoning paths. |
| Approach: | They propose a method that helps large language models identify and revise incorrect steps in their generated reasoning paths. |
| Outcome: | The proposed method outperforms the state-of-the-art Best-ofN decoding method by +2.4 and reduces token consumption by 77.8%. |
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. |
Protecting Privacy in Multimodal Large Language Models with MLLMU-Bench (2025.naacl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) and Multimodal Large Language models (MLLMs) trained on vast web corpora can memorize and disclose individuals’ confidential and private data, raising legal and ethical concerns. |
| Approach: | They propose a benchmark to assess unlearning algorithms from multiple perspectives and provide a baseline for existing generative models. |
| Outcome: | The proposed benchmark consists of 500 fictitious profiles and 153 profiles for public celebrities, evaluated from both multimodal (image+text) and unimodal (text) perspectives. |
Agentic Conversational Search with Contextualized Reasoning via Reinforcement Learning (2026.findings-acl)
Copied to clipboard
Fengran Mo, Yifan Gao, Sha Li, Hansi Zeng, Xin Liu, Zhaoxuan Tan, Xian Li, Jianshu Chen, Dakuo Wang, Meng Jiang
| Challenge: | Existing studies focus on single-turn scenarios, which might lack the ability to handle multi-turn interactions. |
| Approach: | They propose a conversational agent that interleaves search and reasoning across turns and provides tailored rewards towards evolving user goals. |
| Outcome: | The proposed agent interleaves search and reasoning across turns, enabling exploratory and adaptive behaviors learned through reinforcement learning (RL) training with tailored rewards towards evolving user goals. |
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. |
From Personal to Collective: On the Role of Local and Global Knowledge in LLM Personalization (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are the backbone of modern AI systems supporting a wide range of applications. |
| Approach: | They propose a Local–Global knowledge framework that augments user-specific signals with a global knowledge encoding collective behavioral trends. |
| Outcome: | The proposed framework outperforms existing methods on five personalization benchmarks. |
Can LLM Graph Reasoning Generalize beyond Pattern Memorization? (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies seek to enhance the graph reasoning capabilities of Large Language Models (LLMs) by specialized instruction tuning. |
| Approach: | They propose to evaluate LLM graph reasoning generalization using in-distribution settings . they propose to use three strategies to improve LLM generalization . |
| Outcome: | The proposed benchmark evaluates LLM graph reasoning generalization with in-distribution settings only . it shows that LLMs struggle to generalize across reasoning and real-world patterns . |
KALM: Knowledge-Aware Integration of Local, Document, and Global Contexts for Long Document Understanding (2023.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to infuse knowledge graphs with pre-trained LMs are limited by the input sequence length. |
| Approach: | They propose a language model that leverages knowledge in local, document-level, and global contexts for long document understanding. |
| Outcome: | The proposed model achieves state-of-the-art on three long document understanding tasks across 6 datasets/settings. |
Large Language Models Can Self-Correct with Key Condition Verification (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to correct reasoning without external feedback have not been used in large language models. |
| Approach: | They propose an iterative verify-then-correct framework to progressively identify and correct (probably) false responses, named ProCo. |
| Outcome: | The proposed method improves the accuracy of LLMs on three reasoning tasks. |
Democratizing Large Language Models via Personalized Parameter-Efficient Fine-tuning (2024.emnlp-main)
Copied to clipboard
| Challenge: | Experimental results demonstrate that OPPU significantly outperforms existing prompt-based methods across seven diverse tasks in the LaMP benchmark. |
| Approach: | They propose to integrate parametric user knowledge into the personal PEFT parameters and non-parametric knowledge from retrieval and profiles, adapting LLMs to user behavior shifts. |
| Outcome: | The proposed method outperforms existing prompt-based methods across seven diverse tasks in the LaMP benchmark. |
PAR: Political Actor Representation Learning with Social Context and Expert Knowledge (2022.emnlp-main)
Copied to clipboard
Shangbin Feng, Zhaoxuan Tan, Zilong Chen, Ningnan Wang, Peisheng Yu, Qinghua Zheng, Xiaojun Chang, Minnan Luo
| Challenge: | Existing approaches focus on textual data and voting records to induce political actors' stances. |
| Approach: | They propose a Political Actor Representation learning framework that leverages social context and expert knowledge to model ideological stances. |
| Outcome: | The proposed framework improves political text understanding and improves roll call vote prediction and political perspective detection. |
Generalizable LLM Learning of Graph Synthetic Data with Post-training Alignment (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing research has focused on enhancing graph reasoning capabilities of LLMs by supervised fine-tuning on synthetic graph data. |
| Approach: | They propose to unlock generalizable learning of graph with post-training alignment with synthetic graph data by aligning off-the-shelf LLMs and LLM fine-tuned on synthetic graphs. |
| Outcome: | The proposed algorithm improves on synthetic graph problems and out-of-domain tasks with implicit graph structures. |
Knowledge Crosswords: Geometric Knowledge Reasoning with Large Language Models (2024.findings-acl)
Copied to clipboard
Wenxuan Ding, Shangbin Feng, Yuhan Liu, Zhaoxuan Tan, Vidhisha Balachandran, Tianxing He, Yulia Tsvetkov
| Challenge: | Existing tasks and datasets assess LLM knowledge abilities mostly by focusing on atomic (e.g., open-domain QA) or linear (e-hop QA). |
| Approach: | They propose a geometric knowledge reasoning benchmark consisting of incomplete knowledge networks bounded by structured factual constraints where LLMs are tasked with inferring the missing facts to meet all constraints. |
| Outcome: | The proposed methods outperform baseline methods and are more robust towards problems in the hard subset. |