Papers by Difan Zou
SWE-Fixer: Training Open-Source LLMs for Effective and Efficient GitHub Issue Resolution (2025.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models excel in code generation benchmarks, but these benchmarks focus on single-file scenarios with constrained context scope. |
| Approach: | They propose an open-source framework to effectively resolve GitHub issues using a code file retrieval module and a model-based code editing module. |
| Outcome: | The proposed approach achieves state-of-the-art performance on two GitHub benchmarks. |
Deciphering Cultural Representations in Large Language Models via Sparse Autoencoders (2026.findings-acl)
Copied to clipboard
| Challenge: | Prior work has identified so-called cultural neurons, but individual neurons are often polysemous, conflating abstract cultural knowledge with surface-level lexical cues due to superposition. |
| Approach: | They apply Sparse Autoencoders to decompose LLM activations into sparse, interpretable feature representations that disentangle culturally selective features. |
| Outcome: | The proposed model disentangles culturally selective features from paraphrasing and task formats, indicating abstraction beyond lexical correlations. |
Model Unlearning via Sparse Autoencoder Subspace Guided Projections (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing unlearning strategies lack interpretability or fail to provide robust defense against adversarial prompts. |
| Approach: | They propose a framework that leverages SAE features to drive targeted updates in the model’s parameter space. |
| Outcome: | The proposed framework reduces harmful knowledge accuracy by 3.22% compared to baselines and improves adversarial robustness under jailbreak prompts. |
Breaking Contextual Inertia: Reinforcement Learning with Single-Turn Anchors for Stable Multi-Turn Interaction (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated remarkable capabilities when provided with full information in a single turn, yet they exhibit substantial vulnerability in multi-turn interactions. |
| Approach: | They propose a generalizable training approach to stabilize multi-turn interactions by leveraging the model's intrinsic single-turn capabilities as stable internal anchors. |
| Outcome: | The proposed approach outperforms fine-tuning and abstention-based methods and exhibits strong cross-domain generalization. |