Papers by Mengnan Du
Feature Extraction and Steering for Enhanced Chain-of-Thought Reasoning in Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) can solve reasoning and mathematical problems using the Chain-of-Thought technique, but require costly and long CoT data and fine-tuning. |
| Approach: | They propose a method that uses Sparse Autoencoders to extract interpretable features from vanilla CoT and use them to steer the LLM's internal states. |
| Outcome: | The proposed method uses Sparse Autoencoders (SAEs) to extract interpretable features from vanilla CoT and steer the LLM's internal states during generation. |
The Impact of Reasoning Step Length on Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Long reasoning steps in LLMs improve reasoning abilities, but the correlation between their effectiveness and the length of reasoning steps remains largely unknown. |
| Approach: | They conducted experiments that expand and compress the rationale reasoning steps within CoT demonstrations while keeping all other factors constant. |
| Outcome: | The results show that lengthening the reasoning steps in prompts significantly enhances LLMs’ reasoning abilities across multiple datasets. |
Unveiling Project-Specific Bias in Neural Code Models (2024.lrec-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) based neural code models struggle to generalize effectively to real-world inter-project out-of-distribution data. |
| Approach: | They propose a Cond-Idf measurement to measure the relatedness of a token with a label and its project-specificness. |
| Outcome: | The proposed framework improves both inter-project OOD generalization and adversarial robustness while not sacrificing accuracy on intra-project IID data. |
Robustness Challenges in Model Distillation and Pruning for Natural Language Understanding (2023.eacl-main)
Copied to clipboard
| Challenge: | Recent studies have focused on compressing pre-trained language models (PLMs) however, few studies have examined the impact of compression on generalizability and robustness of compressed models for out-of-distribution data. |
| Approach: | They propose to use knowledge distillation and pruning to reduce model generalization and generalization on out-of-distribution data. |
| Outcome: | The proposed compression techniques overfit on shortcut samples and generalize poorly on hard ones. |
LLM Agents in Law: Taxonomy, Applications, and Challenges (2026.acl-long)
Copied to clipboard
Shuang Liu, Ruijia Zhang, Ruoyun Ma, Yujia Deng, Lanyi Zhu, Jiayu Li, Zelong Li, Zhibin Shen, Mengnan Du
| Challenge: | Large language models (LLMs) have improved the legal domain, but deployment of standalone models faces significant limitations regarding hallucination, outdated information, and verifiability. |
| Approach: | They present a survey of LLM agents for legal tasks and analyze their architectures . they analyze the transition from standard legal LLMs to legal agents . |
| Outcome: | The proposed architectures bridge the gap between technical capabilities and domain-specific needs. |
FinCall-Surprise: A Large Scale Multi-modal Benchmark for Earning Surprise Prediction (2026.acl-long)
Copied to clipboard
| Challenge: | Existing models for earnings surprise prediction rely on expensive, proprietary data. |
| Approach: | They propose to use textual transcripts and audio recordings to build a dataset for earnings surprise prediction. |
| Outcome: | The proposed dataset includes 2,688 unique conference calls from 2019 to 2021. |
Mitigating Shortcuts in Language Models with Soft Label Encoding (2024.lrec-main)
Copied to clipboard
| Challenge: | Recent studies have shown that large language models rely on spurious correlations in the data for natural language understanding (NLU) tasks. |
| Approach: | They propose a framework for debiasing shortcuts and a dummy class to encode shortcuts into a model and use it to generate soft labels. |
| Outcome: | The proposed framework significantly improves out-of-distribution generalization while maintaining satisfactory in-district accuracy. |
A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Sparse Autoencoders (SAEs) can disentangle complex features into more interpretable components. |
| Approach: | They propose to use Sparse Autoencoders to disentangle LLM features into more interpretable components. |
| Outcome: | The proposed method disentangles complex features into more interpretable components. |
Towards Interpreting and Mitigating Shortcut Learning Behavior of NLU models (2021.naacl-main)
Copied to clipboard
Mengnan Du, Varun Manjunatha, Rajiv Jain, Ruchi Deshpande, Franck Dernoncourt, Jiuxiang Gu, Tong Sun, Xia Hu
| Challenge: | Recent studies indicate that NLU models are prone to rely on shortcut features for prediction, without achieving true language understanding. |
| Approach: | They propose a shortcut mitigation framework to suppress NLU models from making overconfident predictions for samples with large shortcut degree. |
| Outcome: | The proposed framework suppresses the model from making overconfident predictions for samples with large shortcut degree. |
AdaptiveK: Complexity-Driven Sparse Autoencoders for Interpretable Language Model Representations (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to decomposing model activations into interpretable features fail to account for input complexity. |
| Approach: | They propose a framework that dynamically adjusts sparsity levels based on the semantic complexity of each input. |
| Outcome: | The proposed framework outperforms fixed-sparsity approaches on reconstruction fidelity, explained variance, cosine similarity and interpretability metrics while eliminating the burden of extensive hyperparameter tuning. |
FaithLM: Towards Faithful Explanations for Large Language Models (2026.eacl-long)
Copied to clipboard
Yu-Neng Chuang, Guanchu Wang, Chia-Yuan Chang, Ruixiang Tang, Shaochen Zhong, Fan Yang, Andrew Wen, Mengnan Du, Xuanting Cai, Vladimir Braverman, Xia Hu
| Challenge: | Large language models (LLMs) produce natural language explanations, but they lack faithfulness and do not reflect the evidence the model uses to decide. |
| Approach: | They propose a model-agnostic framework that evaluates and improves the faithfulness of LLM explanations without token masking or task-specific heuristics. |
| Outcome: | The proposed framework improves faithfulness of large language models without masking or heuristics. |
Data-Centric Explainable Debiasing for Improving Fairness in Pre-trained Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing data-centric debiasing strategies mainly leverage explicit bias words for counterfactual data augmentation to balance the training data. |
| Approach: | They propose a method which uses an explainability method to search for implicit bias words to assist in debiasing PLMs. |
| Outcome: | Extensive results show that the proposed method achieves state-of-the-art debiasing performance and strong generalization while maintaining predictive abilities. |
Large Vision-Language Model Alignment and Misalignment: A Survey Through the Lens of Explainability (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large Vision-Language Models have demonstrated remarkable capabilities in processing both visual and textual information. |
| Approach: | They examine the challenge of alignment and misalignment in LVLMs through an explainability lens. |
| Outcome: | The findings highlight the need for standardized evaluation protocols and in-depth explainability studies. |
Exploring Concept Depth: How Large Language Models Acquire Knowledge and Concept at Different Layers? (2025.coling-main)
Copied to clipboard
Mingyu Jin, Qinkai Yu, Jingyuan Huang, Qingcheng Zeng, Zhenting Wang, Wenyue Hua, Haiyan Zhao, Kai Mei, Yanda Meng, Kaize Ding, Fan Yang, Mengnan Du, Yongfeng Zhang
| Challenge: | Large language models have shown remarkable performances across a wide range of tasks, but mechanisms by which they encode tasks of varying complexity remain poorly understood. |
| Approach: | They propose to explore the possibility that LLMs process concepts in different layers . they propose to categorize concepts based on their level of abstraction . |
| Outcome: | The proposed model can process complex concepts in shallow layers, the authors show . the proposed model could be used to prob complex tasks in shallow ones . |
Data-centric NLP Backdoor Defense from the Lens of Memorization (2025.findings-naacl)
Copied to clipboard
| Challenge: | Backdoor attacks pose a severe threat to the trustworthiness of DNN-based language models. |
| Approach: | They propose a data-centric defense that extends memorization definitions to fine-grained sentences . they find that duplicated sentence elements are necessary for successful backdoor attacks . |
| Outcome: | The proposed defense outperforms state-of-the-art defenses against backdoor attacks. |
SAE-FiRE: Enhancing Earnings Surprise Predictions Through Sparse Autoencoder Feature Selection (2026.findings-acl)
Copied to clipboard
| Challenge: | Conference call transcripts contain significant redundancy and industry-specific terminology that creates obstacles for language models. |
| Approach: | They propose a Sparse Autoencoder for Financial Representation Enhancement framework to extract key information from earnings conference call transcripts and eliminate redundancy. |
| Outcome: | The proposed method outperforms baselines in analyzing earnings conference call transcripts. |
Prompt Tuning Pushes Farther, Contrastive Learning Pulls Closer: A Two-Stage Approach to Mitigate Social Biases (2023.acl-long)
Copied to clipboard
| Challenge: | Existing debiasing techniques use Counterfactual Data Augmentation (CDA) to balance the training corpus, but this technique slightly modifies the original corpus limiting the representation distance between different demographic groups. |
| Approach: | They propose a two-stage debiasing model using Contrastive learning with Continuous Prompt Augmentation to mitigate social biases in PLMs’ encoding. |
| Outcome: | The proposed model outperforms baselines in terms of debiasing performance while maintaining the language modeling capability of PLMs. |
AdaJudge: Adaptive Multi-Perspective Judging for Reward Modeling (2026.acl-long)
Copied to clipboard
| Challenge: | Existing reward models rely on a static pooling strategy to condense sequences into scalar scores, which is ill-suited for fine-grained discrimination. |
| Approach: | They propose a framework that jointly adapts representation and aggregation to address these limitations by integrating a static inductive bias with a representational mismatch. |
| Outcome: | Experiments on RM-Bench and JudgeBench show that AdaJudge outperforms strong off-the-shelf reward models and traditional pooling baselines. |
Strategic Demonstration Selection for Improved Fairness in LLM In-Context Learning (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies highlight the effectiveness of using in-context learning (ICL) to steer large language models in processing tabular data. |
| Approach: | They propose a method that uses clustering and evolutionary strategies to curate a representative sample set from training data. |
| Outcome: | The proposed method significantly improves fairness across various metrics, showing its efficacy in real-world scenarios. |
SAGE: An Agentic Explainer Framework for Interpreting SAE Features in Language Models (2026.eacl-industry)
Copied to clipboard
| Challenge: | Large language models (LLMs) have achieved remarkable progress, yet their internal mechanisms remain largely opaque. |
| Approach: | They propose an agent-based framework that recasts feature interpretation from a passive, single-pass generation task into an explanation-driven process. |
| Outcome: | The proposed framework produces explanations with significantly higher generative and predictive accuracy compared to state-of-the-art baselines. |
DeepSieve: Information Sieving via LLM-as-a-Knowledge-Router (2026.findings-eacl)
Copied to clipboard
Minghao Guo, Qingcheng Zeng, Xujiang Zhao, Yanchi Liu, Wenchao Yu, Mengnan Du, Haifeng Chen, Wei Cheng
| Challenge: | Existing RAG methods lack fine-grained control over query and source sides, resulting in noisy retrieval and shallow reasoning. |
| Approach: | They propose an agentic RAG framework that integrates information sieving via LLM-as-a-knowledge-router. |
| Outcome: | Experiments on multi-hop QA tasks across heterogeneous sources demonstrate improved reasoning depth, retrieval precision, and interpretability over conventional approaches. |
Beyond Input Activations: Identifying Influential Latents by Gradient Sparse Autoencoders (2025.emnlp-main)
Copied to clipboard
| Challenge: | Sparse Autoencoders (SAEs) have recently emerged as powerful tools for interpreting and steering the internal representations of large language models (LLMs). |
| Approach: | They propose a method that identifies the most influential latents by incorporating output-side gradient information. |
| Outcome: | The proposed method identifies the most influential latents by incorporating output-side gradient information. |
Secure Your Model: An Effective Key Prompt Protection Mechanism for Large Language Models (2024.findings-naacl)
Copied to clipboard
| Challenge: | Recent years have seen an unprecedented surge in the development and application of large language models (LLMs) however, the development of LLMs is a complex endeavor, requiring substantial investments in terms of financial and computational resources. |
| Approach: | They propose a mechanism wherein a unique key prompt is embedded within the LLM to protect it from unauthorized access and potential theft. |
| Outcome: | The proposed protection can protect the model without significantly impacting its original function. |
FinChart-Bench: Benchmarking Financial Chart Comprehension in Vision-Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | FinChart-Bench is the first benchmark specifically focused on real-world financial charts. |
| Approach: | They propose a benchmark specifically focused on real-world financial charts. |
| Outcome: | The proposed benchmark evaluates 26 state-of-the-art LVLMs on FinChart-Bench. |
Improving LLM Reasoning through Interpretable Role-Playing Steering (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for role-playing rely on prompt engineering, which lacks stability and interpretability. |
| Approach: | They propose a framework that extracts latent representations from role-play prompts and constructs a steering vector that can be injected into the model's residual stream with controllable intensity. |
| Outcome: | The proposed framework extracts latent representations from role-play prompts, selects the most relevant features based on activation patterns, and constructs a steering vector that can be injected into the model’s residual stream with controllable intensity. |
SAE-SSV: Supervised Steering in Sparse Representation Spaces for Reliable Control of Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) have impressive capabilities in natural language understanding and generation, but controlling their behavior remains a challenge. |
| Approach: | They propose a supervised steering approach that operates in sparse, interpretable representation spaces. |
| Outcome: | The proposed approach achieves higher success rates with minimal degradation in generation quality compared to existing methods. |
Denoising Concept Vectors with Sparse Autoencoders for Improved Language Model Steering (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for steering concept vectors suffer from noisy features in diverse datasets that undermine steering robustness. |
| Approach: | They propose a Sparse Autoencoder-Denoised Concept Vector (SDCV) which selectively keeps the most discriminative SAE latents while reconstructing hidden representations. |
| Outcome: | The proposed method improves steering success rates by 4-16% across six challenging concepts while maintaining topic relevance. |