Papers by Di Hu

26 papers
Mechanistic Unveiling of Transformer Circuits: Self-Influence as a Key to Model Reasoning (2025.findings-naacl)

Copied to clipboard

Challenge: Existing studies have shown that large language models implicitly embed reasoning trees, but their internal mechanisms remain largely opaque due to the complexity of non-linear interactions and high-dimensional operations.
Approach: They propose to use circuit analysis and self-influence functions to map the reasoning process of large models.
Outcome: The proposed model is able to map human-interpretable reasoning paths and a model's underlying circuits reveal human-mediated reasoning processes.
Flow Matching for Conditional Text Generation in a Few Sampling Steps (2024.eacl-short)

Copied to clipboard

Challenge: Current diffusion models face multiple drawbacks including slow sampling, noise schedule sensitivity, and misalignment between training and sampling stages.
Approach: They propose a method which leverages flow matching for conditional text generation.
Outcome: The proposed method can generate text in a few steps by training with a novel anchor loss, alleviating the need for expensive hyperparameter optimization of the noise schedule prevalent in diffusion models.
Flattery in Motion: Benchmarking and Analyzing Sycophancy in Video-LLMs (2026.acl-long)

Copied to clipboard

Challenge: Current sycophancy research has largely overlooked its specific manifestations in the video-language domain.
Approach: They propose a video-LLM sycophancy benchmarking and evaluation to evaluate scophancies in video-LLMs.
Outcome: The proposed benchmark evaluates sycophantic behavior in state-of-the-art Video-LLMs across diverse question formats, prompt biases, and visual reasoning tasks.
Texar: A Modularized, Versatile, and Extensible Toolkit for Text Generation (P19-3)

Copied to clipboard

Challenge: Texar is an open-source text generation toolkit that supports a broad set of text generation tasks.
Approach: They introduce Texar, an open-source text generation toolkit that supports text generation tasks.
Outcome: Texar supports machine translation, summarization, dialog, content manipulation, and more.
Understanding How Value Neurons Shape the Generation of Specified Values in LLMs (2025.findings-emnlp)

Copied to clipboard

Challenge: Current approaches to interpret value representations are limited by superficial judgments over mechanistic analysis.
Approach: They propose a mechanistic interpretability framework that uses the Schwartz Values Survey to interpret value . they use a dataset that operationalizes four dimensions of universal value through behavioral contexts .
Outcome: The proposed method bridges psychological value frameworks with neuron analysis in large language models.
QiMeng-PRepair: Precise Code Repair via Edit-Aware Reward Optimization (2026.acl-long)

Copied to clipboard

Challenge: Existing approaches to program repair are based on correctness alone.
Approach: They propose a framework that mitigates over-editing and improves repair accuracy by generating buggy programs and re-edits.
Outcome: The proposed framework improves repair precision by 31.4% under fix1@1, a metric that considers repair correctness and extent, and significantly increases decoding throughput when combined with speculative editing.
Visual Self-Fulfilling Alignment: Shaping Safety-Oriented Personas via Threat-Related Images (2026.acl-long)

Copied to clipboard

Challenge: Existing methods require explicit safety labels or contrastive data, yet visual inputs enable harmful outputs.
Approach: They propose a visual self-fulfilling alignment mechanism that fine-tunes vision-language models on neutral VQA tasks without any safety labels.
Outcome: The proposed approach reduces attack success rate, improves response quality, and mitigates over-refusal while preserving general capabilities.
Private Language Models via Truncated Laplacian Mechanism (2024.emnlp-main)

Copied to clipboard

Challenge: Existing methods for word embedding are prone to privacy leakage, resulting in weaker relaxations of DP that are inferior to the canonical DP in terms of privacy strength.
Approach: They propose a method for private word embedding that uses a non-trivial extension of the truncated Laplacian mechanism and propose to test its effectiveness.
Outcome: The proposed method has lower variance compared to the previous methods.
AutoMonitor-Bench: Evaluating the Reliability of LLM-Based Misbehavior Monitor (2026.findings-acl)

Copied to clipboard

Challenge: AutoMonitor-Bench evaluates the reliability of LLM-based misbehavior monitors across diverse tasks and failure modes.
Approach: They introduce AutoMonitor-Bench, a benchmark designed to evaluate misbehavior monitors across diverse tasks and failure modes.
Outcome: The new benchmark evaluates the reliability of LLM-based misbehavior monitors across tasks and failure modes.
COMPKE: Complex Question Answering under Knowledge Editing (2025.findings-acl)

Copied to clipboard

Challenge: Existing benchmarks for knowledge editing do not accurately evaluate how well models apply knowledge in real-life situations.
Approach: They propose a benchmark to evaluate how well updated models apply new knowledge in real-life situations.
Outcome: The proposed method achieves 39.47 accuracy on GPT-4o-mini but drops significantly to 3.83 on Qwen2.5-3B.
Revisiting Entropy Regularization: Adaptive Coefficient Unlocks Its Potential for LLM Reinforcement Learning (2026.findings-acl)

Copied to clipboard

Challenge: Reasoning ability is a defining capability of Large Language Models (LLMs), but RLVR training suffers from policy entropy collapse, hindering exploration and limiting reasoning performance.
Approach: They propose a framework that dynamically balances exploration and exploitation via three components: difficulty-aware coefficient allocation, initial-anchored target entropy, and dynamic global coefficient adjustment.
Outcome: The proposed framework outperforms baselines on multiple mathematical reasoning benchmarks.
CPsyExam: A Chinese Benchmark for Evaluating Psychology using Examinations (2025.coling-main)

Copied to clipboard

Challenge: CPsyExam prioritizes psychological knowledge and case analysis separately, recognizing the significance of applying psychological knowledge to real-world scenarios.
Approach: They propose a psychological benchmark, CPsyExam, constructed from questions from Chinese examination systems.
Outcome: The proposed benchmark prioritizes psychological knowledge and case analysis separately, recognizing the significance of applying psychological knowledge to real-world scenarios.
CODEMENV: Benchmarking Large Language Models on Code Migration (2025.findings-acl)

Copied to clipboard

Challenge: Large language models (LLMs) have demonstrated remarkable proficiency in handling a wide range of tasks within the software engineering domain, but their ability to perform code migration—adapting code to different environments—remains underexplored.
Approach: They propose a benchmark to evaluate large language models’ performance in handling code migration tasks.
Outcome: The proposed benchmark comprises 922 data points across 19 Python and Java packages and offers three tasks to systematically evaluate code migration: identifying version-incompatible functions, determining function changes, and adapting code to target environments.
CPsyCoun: A Report-based Multi-turn Dialogue Reconstruction and Evaluation Framework for Chinese Psychological Counseling (2024.findings-acl)

Copied to clipboard

Challenge: Existing datasets lack consulting knowledge, resulting in LLMs lacking professional consulting competence.
Approach: They propose a report-based multi-turn dialogue reconstruction framework for Chinese psychological counseling that uses large language models to assist counseling.
Outcome: The proposed framework is open-source and can be used in future research.
Fraud-R1 : A Multi-Round Benchmark for Assessing the Robustness of LLM Against Augmented Fraud and Phishing Inducements (2025.findings-acl)

Copied to clipboard

Challenge: Existing fraud detection benchmarks focus on single-turn classification tasks, failing to capture dynamic nature of real-world fraud attempts.
Approach: They propose a bilingual benchmark to assess LLMs' ability to resist fraud and phishing attacks across five key fraud categories: Fraudulent Services, Impersonation, Phishing Scams, Fake Job Postings, and Online Relationships.
Outcome: The proposed model improves in role-play settings and in e-commerce and recommendation systems.
Differentially Private Natural Language Models: Recent Advances and Future Directions (2024.findings-eacl)

Copied to clipboard

Challenge: Recent advances in deep learning have led to great success in various natural language processing tasks.
Approach: They propose a systematic review of recent advances in DP deep learning models . they discuss some differences and additional challenges of DP-NLP .
Outcome: The proposed method can prevent reconstruction attacks and protect against potential side knowledge while maintaining the privacy of sensitive data.
Towards Effective and Efficient Continual Pre-training of Large Language Models (2025.acl-long)

Copied to clipboard

Challenge: Continual pre-training (CPT) has been an important approach for adapting language models to specific domains or tasks.
Approach: They propose a Continual pre-training method that can greatly improve Chinese language ability and scientific reasoning ability of LLMs.
Outcome: The proposed method can greatly improve Chinese language ability and scientific reasoning ability of LLMs.
Dissecting Fine-Tuning Unlearning in Large Language Models (2024.emnlp-main)

Copied to clipboard

Challenge: Existing methods for fine-tuning-based unlearning are ineffective at completely erasing model-embedded knowledge, but their true effectiveness remains unclear.
Approach: They propose to use activation patching and parameter restoration experiments to examine the limitations of fine-tuning-based unlearning methods for erasing harmful, sensitive, or copyrighted information within large language models.
Outcome: The proposed methods alter the model’s knowledge retrieval process rather than genuinely erasing the problematic knowledge embedded in the model parameters.
GRI: Graph-based Relative Isomorphism of Word Embedding Spaces (2023.findings-emnlp)

Copied to clipboard

Challenge: Existing attempts to control relative isomorphism of different spaces fail to consider lexical variations of semantically similar words . Existing methods for building bilingual dictionaries rely on geometric similarity of individual spaces .
Approach: They propose a method that incorporates the impact of lexical variations of semantically similar words into the training objective.
Outcome: The proposed method outperforms existing research by improving the average P@1 by 63.6%.
Antonym vs Synonym Distinction using InterlaCed Encoder NETworks (ICE-NET) (2024.findings-eacl)

Copied to clipboard

Challenge: Existing research on antonym-synonym distinction is limited by the sparsity of the feature space.
Approach: They propose to capture and model relation-specific properties of antonyms and synonyms pairs . ICE-NET outperforms existing research by a relative score of upto 1.8% in F1-measure .
Outcome: The proposed model outperforms existing models by 1.8% in the F1-measure.
Understanding the Repeat Curse in Large Language Models from a Feature Perspective (2025.findings-acl)

Copied to clipboard

Challenge: Large language models suffer from repetitive text generation, a phenomenon we refer to as the ”Repeat Curse”.
Approach: They propose a method to induce and analyze the Repeat Curse in large language models by using mechanistic interpretability.
Outcome: The proposed method induces and analyzes the Repeat Curse in large language models using mechanistic interpretability.
Breaking the Stage Barrier: A Novel Single-Stage Approach to Long Context Extension for Large Language Models (2025.coling-main)

Copied to clipboard

Challenge: Recent studies show that Large language models struggle with handling long token sequences due to limited training context size.
Approach: They propose a single-stage continual pretraining method to equip LLMs with long context modeling capabilities.
Outcome: The proposed method outperforms existing methods on 4 language modeling benchmarks.
DSMoE: Matrix-Partitioned Experts with Dynamic Routing for Computation-Efficient Dense LLMs (2025.emnlp-main)

Copied to clipboard

Challenge: Existing sparsification methods like pruning can lose model knowledge through parameter removal.
Approach: They propose a novel approach that achieves sparsification by partitioning pre-trained FFN layers into computational blocks.
Outcome: The proposed approach achieves superior performance across language modeling and downstream tasks under equivalent computational constraints.
Can Large Language Models Identify Implicit Suicidal Ideation? An Empirical Evaluation (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing data on suicidal ideation in private conversations are limited . a new dataset of 1,200 test cases is presented to address this gap .
Approach: They propose a dataset of 1,200 test cases simulating implicit suicidal ideation in private contexts.
Outcome: The proposed dataset includes 1,200 test cases simulating implicit suicidal ideation in dialogue scenarios.
SPPD: Self-training with Process Preference Learning Using Dynamic Value Margin (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing approaches to improve numerical and logical reasoning of Large Language Models are limited . existing approaches rely on prompt engineering and pretrained knowledge to ensure correctness .
Approach: They propose to train LLMs with process-based reasoning using a dynamic value margin . they use the Bellman optimality equation to derive a value margin for step-level preference optimization .
Outcome: The proposed method is equivalent to on-policy policy gradient methods under constrained reward functions.
Rhombus: Incentivizing Coordination in Parallel Thinking through Reinforcement Learning (2026.findings-acl)

Copied to clipboard

Challenge: Parallel thinking is a promising avenue for scaling test-time compute in Large Language Models . however, coordinating the exploration and aggregation stages remains challenging .
Approach: They propose a parallel thinking framework that explicitly incentivizes coordination between components via end-to-end reinforcement learning.
Outcome: The proposed framework improves accuracy by 6.0% over long chain-of-thought baselines while reducing wall-clock latency by 39.4% under matched token budgets.

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