Selected work

Representative Papers

Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models

Jan 20, 2026

This work proposes a practical, three-stage “Locate–Guide–Improve” framework that transforms mechanistic interpretability from a post-hoc diagnostic tool into an engineering-driven optimization methodology for large language models. By systematically integrating techniques for identifying critical neurons and pathways with targeted interventions—such as activation manipulation and module editing—the framework establishes a standardized protocol for model refinement while clearly distinguishing between localization and guidance mechanisms. Empirical results demonstrate significant improvements in model alignment, task performance, and reasoning efficiency, thereby advancing mechanistic interpretability toward real-world applicability.

4 citationsRead paper

Analysis of Long Range Dependency Understanding in State Space Models

Jan 19, 2026

This work addresses the limited interpretability of existing state space models regarding their long-range dependency mechanisms, particularly the unclear relationship between modeling capacity and architectural design in real-world tasks. Focusing on the S4D model, we present the first systematic analysis of its kernel behavior in the context of source code vulnerability detection. By integrating time-domain and frequency-domain analyses, we demonstrate that S4D can function as a low-pass, band-pass, or high-pass filter depending on its architectural configuration. This finding reveals that the model’s ability to capture long-range dependencies is profoundly influenced by its architecture, thereby offering both theoretical insights and concrete guidance for designing more effective state space models.

1 citations1 influentialRead paper

Self-Rationalization in the Wild: A Large Scale Out-of-Distribution Evaluation on NLI-related tasks

Feb 07, 2025

This study investigates the capability of large language models (LLMs) to generate trustworthy free-text explanations in zero-shot settings, with a focus on out-of-distribution (OOD) generalization. We introduce the first large-scale, cross-task benchmark for explanation generation, covering 19 OOD datasets across natural language inference, fact verification, and summary hallucination detection. Our method employs fine-tuned T5-Large and OLMo-7B models integrated with a few-shot selection strategy and a novel reference-free evaluation framework—including the proposed Acceptability score—assessing explanation faithfulness, coherence, and informativeness. Key findings include: (i) a small number of high-quality annotations substantially improves OOD explanation quality; (ii) explanation quality strongly correlates with prediction accuracy; (iii) the Acceptability score achieves a Pearson correlation of 0.82 with human judgments; and (iv) data source quality exerts a significantly greater influence on OOD performance than sampling strategy.

1 citations1 influentialRead paper

Depth-Recurrent Attention Mixtures: Giving Latent Reasoning the Attention it Deserves

Jan 29, 2026

Existing deep recurrent models are constrained by fixed hidden dimensions and rigid stacked architectures, and they lack baselines aligned across FLOPs, parameters, and memory, hindering efficient multi-step implicit reasoning. This work proposes Dreamer, a modular framework that, for the first time, integrates deep attention with a sparse mixture-of-experts (MoE) mechanism. By introducing attention along the depth dimension, Dreamer decouples model scaling factors and overcomes the limitation of constant hidden size. The approach substantially enhances expert selection diversity (by 2–11×) and knowledge utilization efficiency. On language reasoning benchmarks, Dreamer achieves comparable accuracy with 2–8× fewer training tokens and outperforms current state-of-the-art models of approximately twice its size under the same training budget.

1 citationsRead paper

SciCoQA: Quality Assurance for Scientific Paper--Code Alignment

Jan 19, 2026

Scientific papers and their code implementations often exhibit inconsistencies that threaten research reproducibility. This work presents the first systematic definition and annotation of paper–code discrepancy types, introducing SciCoQA—a dataset comprising 611 samples spanning domains such as artificial intelligence, physics, and quantitative biology. The dataset integrates real GitHub issues with synthetically augmented data to support a scalable framework for generating and detecting discrepancies. Evaluation using large language models reveals that even the most advanced model, GPT-5, identifies only 45.7% of real discrepancies, underscoring both the difficulty of the task and the necessity of this study.

1 citationsRead paper

Recent publications

Latest Papers

ForwardDLO: Model-Based Bimanual Shape Matching of Unconstrained Deformable Linear Objects

Sep 16, 2026

研究提出ForwardDLO模型，通过预测绳索各段位移解决机器人双臂操控非固定柔性线性物体形状匹配问题。

0 citationsRead paper

Beyond Measurement Metrics: A Human-Centered Framework for Semantic Validation of Network Traffic Classification

Sep 15, 2026

本文针对网络流量分类模型的语义验证问题，提出结合数据、机器学习模型、可解释性、可视化和专家推理的人类中心框架，以确保模型准确且可信。

0 citationsRead paper

Towards Digital Halftoning on Closed Manifolds--An Error Diffusion Scheme for the $2D$ Torus based on Sigma-Delta Quantization along the Rank-one Lattice

Sep 15, 2026

本文研究了二维环面上的数字半色调问题，通过沿秩一格子的Sigma-Delta量化方法进行误差扩散，减少了边界伪影，并提高了重构精度。

0 citationsRead paper

Reasoning with Image Generation

Sep 14, 2026

该研究通过将图像生成模型作为多模态大语言模型的灵活视觉推理机制，解决了需要直接操作视觉表示的任务问题。

0 citationsRead paper

Calibrated Ambiguity in Multimodal Language Models: Humans reach for cultural references, while models describe the picture

Sep 11, 2026

研究探讨了多模态语言模型在生成具有校准模糊性的线索时的问题，发现模型倾向于过度指定且缺乏文化引用，与人类生成的线索不同。

0 citationsRead paper