Papers by Tianyu Du

19 papers
Probing the Geometry of Truth: Consistency and Generalization of Truth Directions in LLMs Across Logical Transformations and Question Answering Tasks (2025.findings-acl)

Copied to clipboard

Challenge: Large language models (LLMs) are trained on vast corpora that contain substantial knowledge but their outputs often contain confidently stated inaccuracies.
Approach: They propose to encode truthfulness as a distinct linear feature, termed the "truth direction", which can classify truthfulness reliably.
Outcome: The proposed model can generalize to logical transformations, question-answering tasks, in-context learning, and external knowledge sources.
VideoEraser: Concept Erasure in Text-to-Video Diffusion Models (2025.emnlp-main)

Copied to clipboard

Challenge: Experimental results show that VideoEraser outperforms prior methods regarding efficacy, integrity, fidelity, robustness, and generalizability.
Approach: They propose a training-free framework that prevents T2V diffusion models from generating videos with undesirable concepts even when explicitly prompted with those concepts.
Outcome: The proposed framework outperforms existing methods in erasure, celebrity erasion, and explicit content erasing tasks.
KARPA: A Training-free Method of Adapting Knowledge Graph as References for Large Language Model’s Reasoning Path Aggregation (2025.findings-acl)

Copied to clipboard

Challenge: Existing methods for large language models (LLMs) are limited by step-by-step decision-making on KGs, or require fine-tuning or pre-training on specific KG.
Approach: They propose a framework that harnesses the global planning abilities of large language models (LLMs) for efficient and accurate KG reasoning.
Outcome: Extensive experiments show that the proposed framework achieves state-of-the-art performance in KGQA tasks, delivering both high efficiency and accuracy.
IW-Bench: Evaluating Large Multimodal Models for Converting Image-to-Web (2025.findings-acl)

Copied to clipboard

Challenge: Existing models have been introduced to improve image comprehension, but there is no robust benchmark for imagetoweb conversion.
Approach: They propose a benchmark to assess imagetoweb conversion proficiency of large multimodal models . they propose to measure layout information of web pages by parsing the Document Object Model tree .
Outcome: The proposed benchmark measures the layout information of web pages—i.e., the positional relationships between elements—which has been overlooked by prior work.
LongSpec: Long-Context Lossless Speculative Decoding with Efficient Drafting and Verification (2026.acl-long)

Copied to clipboard

Challenge: Large Language Models (LLMs) can process extremely long contexts, requiring efficient inference over extended inputs.
Approach: They propose a model that uses a constant-sized key-value cache to train long-context models.
Outcome: Experimental results show that LongSpec achieves 3.26x speedup over strong Flash Attention baselines and 2.34x wall clock time on four math reasoning tasks.
RA-ISF: Learning to Answer and Understand from Retrieval Augmentation via Iterative Self-Feedback (2024.findings-acl)

Copied to clipboard

Challenge: Large language models (LLMs) have demonstrated excellent performance in numerous tasks but the parameterized knowledge stored within LLMs may be incomplete and hard to incorporate up-to-date knowledge.
Approach: They propose a framework that iteratively decomposes tasks and processes them in three submodules to enhance the model’s problem-solving capabilities.
Outcome: The proposed method outperforms existing benchmarks on GPT3.5, Llama2 and other large language models significantly enhancing factual reasoning capabilities and reducing hallucinations.
A Fast and High-quality Text-to-Speech Method with Compressed Auxiliary Corpus and Limited Target Speaker Corpus (2024.lrec-main)

Copied to clipboard

Challenge: Existing methods to generate high-quality speech with limited target speaker corpus require extensive training data.
Approach: They propose an auxiliary corpus compression algorithm that reduces the training cost while the naturalness of synthesized speech is not significantly degraded.
Outcome: The proposed method significantly reduces training costs while maintaining the naturalness of synthesized speech.
“I See What You Did There”: Can Large Vision-Language Models Understand Multimodal Puns? (2026.acl-long)

Copied to clipboard

Challenge: Puns are a common form of rhetorical wordplay that exploits polysemy and phonetic similarity to create humor.
Approach: They propose a multimodal pun generation pipeline and a model to evaluate their understanding of puns.
Outcome: The proposed benchmark improves the understanding of multimodal puns by 16.5% in the F1 test.
CLMTracing: Black-box User-level Watermarking for Code Language Model Tracing (2025.emnlp-main)

Copied to clipboard

Challenge: Open-source code language models (code LMs) are a growing threat for intellectual property protection.
Approach: They propose a black-box code LM watermarking framework that uses rule-based watermarks and utility-preserving injection method for user-level model tracing.
Outcome: The proposed framework shows that it performs well across multiple state-of-the-art code LMs and is harmless compared to existing baselines.
ACIArena: Toward Unified Evaluation for Agent Cascading Injection (2026.acl-long)

Copied to clipboard

Challenge: Existing studies consider only limited attack strategies and simplified MAS settings, limiting their generalizability and comprehensive evaluation.
Approach: They propose a framework to evaluate the robustness of Multi-Agent Systems (MAS) they propose unified evaluation suites spanning attack surfaces and attack objectives .
Outcome: ACIArena provides a benchmark of 1,356 test cases for evaluating MAS robustness . it covers six widely used MAS implementations and provides measurable results .
IPIGuard: A Novel Tool Dependency Graph-Based Defense Against Indirect Prompt Injection in LLM Agents (2025.emnlp-main)

Copied to clipboard

Challenge: Existing methods for detecting Indirect Prompt Injection (IPI) attacks rely on assumptions about the model's inherent security, which lacks structural constraints on agent behaviors.
Approach: They propose a novel task execution paradigm that models the agents’ task execution process as a traversal over a planned Tool Dependency Graph (TDG).
Outcome: The proposed model reduces unintended tool invocations triggered by injected instructions, enhancing robustness against IPI attacks.
Efficiently Computing Susceptibility to Context in Language Models (2024.findings-emnlp)

Copied to clipboard

Challenge: a current language model is able to incorporate information from a user-input context when answering queries, but it is not equally sensitive to subtle changes to that context.
Approach: They propose a metric to quantify the degree to which contexts can influence a model’s response to a query at a distributional level.
Outcome: The proposed method is comparable to Monte Carlo's estimated susceptibility across a diverse set of query domains despite being 70 faster.
COIG-P: A High-Quality and Large-Scale Chinese Preference Dataset for Alignment with Human Values (2026.findings-eacl)

Copied to clipboard

Challenge: Existing Chinese preference datasets suffer from limited scale, restricted domain coverage, and insufficiently rigorous data validation.
Approach: They propose an LLM-based data annotation pipeline with no human intervention to annotate Chinese preference datasets.
Outcome: The proposed pipeline outperforms existing Chinese preference datasets on AlignBench and Chinese Reward Benchmark.
Compiling Activation Steering into Weights via Null-Space Constraints for Stealthy Backdoors (2026.acl-long)

Copied to clipboard

Challenge: Existing methods to inject safety-aligned large language models rely on token-level mappings, which do not guarantee sustained harmful output.
Approach: They propose a method that directly modifies model weights to map a trigger to an attacker-specified response.
Outcome: The proposed method achieves high triggered attack success while maintaining non-triggered safety and general utility.
DROWN: Towards Tighter LiRPA-based Robustness Certification (2025.coling-main)

Copied to clipboard

Challenge: Existing methods for certifying the robustness of deep neural networks suffer from precision or scalability issues.
Approach: They propose a method to certify the robustness of deep neural networks . they propose to use two pairs of linear bounds to refine pre-activation bounds .
Outcome: The proposed method achieves higher certified robustness than the baseline on CNNs and 4.68 times larger certified radii than the Transformers.
PerMemSafe: Benchmarking Implicit Personalized Safety of Long Horizon Self-Evolving Agents (2026.findings-acl)

Copied to clipboard

Challenge: Existing self-evolving agents have a low safety rate in long-horizon interactions . however, this reliance on context-independent safety evaluations is insufficient .
Approach: They propose a framework that explicitly models personalized risk inference and memory evolution.
Outcome: The proposed framework improves implicit personalized safety by 23.8% over prior frameworks while maintaining helpfulness in long-horizon interactions.
COIG-CQIA: Quality is All You Need for Chinese Instruction Fine-tuning (2025.findings-naacl)

Copied to clipboard

Challenge: Existing datasets for Chinese instruction tuning are not well-aligned with Chinese users’ interaction patterns.
Approach: They propose to use Chinese instruction tuning datasets to improve instruction fine-tuning for Chinese users.
Outcome: The proposed dataset shows that Chinese models achieve competitive performance in diverse benchmarks.
SecCoder: Towards Generalizable and Robust Secure Code Generation (2024.emnlp-main)

Copied to clipboard

Challenge: Existing secure code generation methods have limited generalizability to unseen test cases and poor robustness against the attacked model, leading to safety failures in code generation.
Approach: They propose a generalizable and robust secure code generation method SecCoder by using in-context learning and the safe demonstration.
Outcome: The proposed method achieves a significant security improvement of 7.20% on unseen test cases and better robustness against the attacked model.
ERA-CoT: Improving Chain-of-Thought through Entity Relationship Analysis (2024.acl-long)

Copied to clipboard

Challenge: Large language models (LLMs) have demonstrated remarkable in-context learning capabilities in various natural language processing tasks.
Approach: They propose a novel approach ERA-CoT which aids LLMs in understanding context by capturing relationships between entities and supports the reasoning of diverse tasks through Chain-of-Thoughts (CoT).
Outcome: The proposed method improves on GPT3.5 and previous SOTA prompting methods by an average of 5.1% compared to previous prompting approaches.

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