Papers by Du Su

18 papers
Steering Away from Refusal: A Black-box Jailbreak Method Based on First-Token Distribution (2026.findings-acl)

Copied to clipboard

Challenge: Existing methods to analyze black-box jailbreaks lack direct optimization signals to refine adversarial prompts.
Approach: They propose a distribution-jailbreak attack method that selects effective jailbreak templates and iteratively optimizes adversarial suffixes by maximizing the KL divergence from the standard refusal distribution.
Outcome: The proposed method achieves state-of-the-art Attack Success Rate (ASR) on all tested open-source models and delivers over 94% ASR on GPT-4.1.
LLMs Assist NLP Researchers: Critique Paper (Meta-)Reviewing (2024.emnlp-main)

Copied to clipboard

Challenge: a comparative analysis of paper (meta-)reviews by large language models (LLMs) aims to identify and distinguish LLMs from human activities .
Approach: They present a comparative analysis to identify and distinguish LLM activities from human activities.
Outcome: The proposed analysis aims to improve recognition of instances when someone implicitly uses LLMs for reviewing activities.
GIFT: Guided Fine-Tuning and Transfer for Enhancing Instruction-Tuned Language Models (2026.acl-long)

Copied to clipboard

Challenge: Existing adapter-based transfer methods treat instruction-tuned models as passive targets . direct fine-tuning can disrupt this delicate balance and lead to instability or performance degradation.
Approach: They propose a framework that incorporates instruction-level guidance into task adaptation.
Outcome: The proposed framework outperforms direct fine-tuning and representative transfer-based baselines while maintaining robust generalization and favorable test-time scaling behavior.
DetectRL-X: Towards Reliable Multilingual and Real-World LLM-Generated Text Detection (2026.acl-long)

Copied to clipboard

Challenge: Existing detectors are limited in their ability to detect large language models generated content in multilingual environments.
Approach: They propose a multilingual benchmark to evaluate advanced detectors across 8 dimensions to better align with real-world applications.
Outcome: The proposed benchmark encompasses 8 languages commonly used in commercial contexts and collects human-written texts from 6 domains highly susceptible to LLM misuse.
Distilling Large Embeddings via Hyperspherical Householder Quantization (2026.acl-long)

Copied to clipboard

Challenge: Existing methods for quantizing large embeddings rely on Euclidean quantization, which is poorly aligned with the angular geometry induced by contrastive embeddment training.
Approach: They propose a geometry-aware distillation method that compresses large embeddings into short discrete representations via iterative Householder transformations on the unit hypersphere.
Outcome: The proposed method reduces decoding cost and maintains strong semantic retrieval accuracy.
PRDetect: Perturbation-Robust LLM-generated Text Detection Based on Syntax Tree (2025.findings-naacl)

Copied to clipboard

Challenge: Recent methods for detecting LLM-generated text have shown impressive performance, but in real-world scenarios, users often introduce perturbations to the text.
Approach: They propose a method that detects syntactic trees that are minimally affected by perturbations and exhibit distinct differences between human-written and LLM-generated text.
Outcome: The proposed method shows that it is significantly better against perturbations on the HC3 and GPT-3.5-mixed datasets and also has the shortest time expenditure.
Adopting the Word-Pair-Dependency-Triplets with Individual Comparison for Natural Language Inference (C18-1)

Copied to clipboard

Challenge: Existing approaches to perform natural language inference ignore syntactic dependency among words or use tree-LSTM to generate sentence representation with irrelevant information.
Approach: They propose to perform natural language inference with Word-Pair-Dependency-Triplets . they propose to compare the triplets of a given passage-pair to make judgement more interpretable .
Outcome: The proposed approach is better than most of the approaches that use tree structures and comparable to other state-of-the-art approaches.
Training LLMs to be Better Text Embedders through Bidirectional Reconstruction (2025.emnlp-main)

Copied to clipboard

Challenge: Existing text embedding approaches often leverage the embeddment of the final token, typically a reserved special token such as ‘[EOS]‘.
Approach: They propose to add a new training stage before contrastive learning to enrich the semantics of the final token embedding.
Outcome: The proposed training stage improves performance on the Massive Text Embedding Benchmark (MTEB), achieving new state-of-the-art results across different LLM base models and scales.
Jailbreak LLMs through Internal Stance Manipulation (2025.emnlp-main)

Copied to clipboard

Challenge: Existing approaches to exploit LLMs' inherent safety mechanism, including GCG and AutoDAN, are ineffective for certain malicious requests.
Approach: They propose a method that generates jailbreak prompts to suppress a refusal stance and induce affirmative responses by modifying adversarial prompts.
Outcome: The proposed method outperforms the best baseline approach in Llama-2-7b-chat and achieves a 92.2% success rate across all models.
TwinVoice: A Multi-dimensional Benchmark Towards Digital Twins via LLM Persona Simulation (2026.findings-acl)

Copied to clipboard

Challenge: Existing studies show that advanced LLMs produce text indistinguishable from human writing.
Approach: They propose a benchmark to assess persona simulation across diverse contexts by decomposing the evaluation into six fundamental capabilities including opinion consistency, memory recall, logical reasoning, persona tone, and syntactic style.
Outcome: The proposed model achieves moderate accuracy but falls short of the basic capabilities needed to simulate personas in real-world contexts.
Too Consistent to Detect: A Study of Self-Consistent Errors in LLMs (2025.emnlp-main)

Copied to clipboard

Challenge: Existing detection methods fail to account for **self-consistent error** . study identifies self-consistency errors and evaluates them .
Approach: They propose a method that fuses hidden state evidence from an external verifier LLM to detect self-consistent errors.
Outcome: The proposed method significantly enhances performance on self-consistent errors across three LLM families.
The Subtle Art of Defection: Understanding Uncooperative Behaviors in LLM based Multi-Agent Systems (2026.eacl-industry)

Copied to clipboard

Challenge: Existing literature on uncooperative behavior degrades collective outcomes and requires more resilient multi-agent systems.
Approach: They propose a game theory-based taxonomy of uncooperative agent behaviors and a structured, multi-stage simulation pipeline that dynamically generates and refines uncooperation behaviors as agents’ states evolve.
Outcome: The proposed framework achieves 96.7% accuracy in generating realistic uncooperative behaviors, validated by human evaluations.
WebOlympus: An Open Platform for Web Agents on Live Websites (2024.emnlp-demo)

Copied to clipboard

Challenge: Web agents are emerging as powerful tools for automating tasks in cyberspace . however, there is a lack of standardized and user-friendly tools for research and development .
Approach: They propose an open platform for web agents operating on live websites with a Chrome extension and a safety monitor module to ensure their trustworthiness.
Outcome: WebOlympus is an open platform for web agents operating on live websites.
Low-Entropy Watermark Detection via Bayes’ Rule Derived Detector (2025.findings-acl)

Copied to clipboard

Challenge: Existing methods for text watermarking ignore strong evidences embedded in low-entropy tokens, causing statistical measures to falsely indicate the absence of a watermark.
Approach: They propose a Bayes' Rule derived watermark Detector which exploits watermark information from every token by leveraging the posterior probability of watermark’s presence.
Outcome: The proposed method achieves 50% and 70% relative improvements over baselines in code generation and math problem-solving tasks.
Token-Level Policy Optimization: Linking Group-Level Rewards to Token-Level Aggregation via sequence-level likelihood (2026.acl-long)

Copied to clipboard

Challenge: Group Relative Policy Optimization (GRPO) has significantly advanced the reasoning ability of large language models (LLMs).
Approach: They propose a token-level framework that leverages sequence-level likelihood to link group-level rewards with individual tokens via token- level aggregation and introduces a KL-Divergence mask constraint that targets tokens with positive advantages and decreasing entropy to mitigate abrupt policy updates.
Outcome: Experiments show that TEPO achieves state-of-the-art performance on mathematical reasoning benchmarks and reduces convergence time by 50% compared with GRPO/DAPO.
Safety-Utility Conflicts Are Not Global: Surgical Alignment via Head-Level Diagnosis (2026.acl-long)

Copied to clipboard

Challenge: Existing mitigation strategies rely on global gradient geometry to resolve alignment conflicts . however, they overlook Modular Heterogeneity within Transformers, resulting in suboptimal trade-offs . Conflict-Aware Sparse Tuning (CAST) combines head-level diagnosis with sparse fine-tuning .
Approach: They propose a framework that integrates head-level diagnosis with sparse fine-tuning to address this limitation.
Outcome: The proposed framework integrates head-level diagnosis with sparse fine-tuning to reduce alignment conflicts in LLMs.
The Fall of ROME: Understanding the Collapse of LLMs in Model Editing (2024.findings-emnlp)

Copied to clipboard

Challenge: Recent studies have found that model editing methods can cause large language models to collapse with just a single edit.
Approach: They propose a method that uses prefixed keys and adds prefixes during testing to prevent model collapse.
Outcome: The proposed method prevents model collapse while maintaining effectiveness, the authors show . Rank-One Model Editing (ROME) has been found to cause model collapse with just a single edit .
BaseCal: Unsupervised Confidence Calibration via Base Model Signals (2026.acl-long)

Copied to clipboard

Challenge: Post-trained LLMs typically compromise reliability with severe overconfidence, resulting in inaccurate responses.
Approach: They propose a solution that feeds PoLLMs into the base LLM to get confidence.
Outcome: The proposed solution reduces expected calibration error (ECE) by 42.90% compared to the best unsupervised baselines.

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