Institution profile

ShanghaiTech University

Academic institutionasia · cn
Official website
Research library627linked papers
Opportunities0open roles
Selected work

Representative Papers

Interpreting and Controlling LLM Reasoning through Integrated Policy Gradient

Feb 02, 2026

Existing methods struggle to precisely identify the internal mechanisms within large language models that support complex reasoning and fail to effectively model the sequential influence from internal components to final outputs. This work proposes the Integrated Policy Gradient (IPG) framework, which introduces— for the first time—the policy gradient concept from reinforcement learning into the interpretability research of large language models. By backpropagating composite signals such as reasoning outcomes and incorporating retrospective analysis of reasoning trajectories, IPG identifies and modulates neurons or modules that cumulatively contribute to long-range reasoning. Experiments demonstrate that IPG achieves more accurate mechanistic localization across multiple reasoning models and effectively tunes both the capability and intensity of reasoning, thereby validating its efficacy and generalizability.

1 citationsRead paper

FinCARDS: Card-Based Analyst Reranking for Financial Document Question Answering

Jan 11, 2026arXiv.org

This work addresses the limitations of existing reranking methods for large language models in financial long-document question answering, which rely solely on semantic relevance and thus fail to enforce strict constraints on entities, financial metrics, fiscal years, and numerical values—leading to unstable and uninterpretable rankings. To overcome this, the authors propose FinCARDS, a structured reranking framework that formulates evidence selection as a constraint satisfaction problem guided by financial-aware patterns. By leveraging field-aligned card representations, deterministic field matching, multi-stage tournament reranking, and stability-aware aggregation, FinCARDS achieves auditable and highly stable evidence ranking without requiring model fine-tuning or additional inference overhead. Experiments demonstrate that the method significantly outperforms lexical and LLM-based reranking baselines on two financial QA benchmarks, substantially improving early retrieval performance and effectively reducing ranking variance.

1 citationsRead paper

KEENHash: Hashing Programs into Function-Aware Embeddings for Large-Scale Binary Code Similarity Analysis

Jun 13, 2025

To address the poor scalability of function-level matching in large-scale binary code similarity analysis (BCSA), this paper proposes KEENHash—the first end-to-end hashing framework designed for program-level similarity retrieval. KEENHash jointly leverages LLM-generated function-level semantic embeddings, K-Means–guided locality-sensitive hashing, and feature compression to produce fixed-length, function-aware program-level hash codes. Unlike conventional per-function matching approaches, KEENHash reduces time complexity to sublinear, enabling real-time program similarity retrieval at billion-scale. Evaluated on 5.3 billion similarity comparisons, KEENHash completes inference in just 395.83 seconds—accelerating over state-of-the-art methods by at least 215×—while improving malware detection accuracy by ≥23.16%.

1 citationsRead paper

MCPDepth: Omnidirectional Depth Estimation via Stereo Matching from Multi-Cylindrical Panoramas

Aug 03, 2024arXiv.org

Omnidirectional depth estimation suffers from inherent distortions in panoramic images—particularly vertical stretching—which existing methods inadequately model due to insufficient consideration of projection geometry. This paper proposes a two-stage framework: (1) stereo matching across multi-cylindrical projections of the same panoramic image, followed by (2) cross-view weighted fusion of resulting depth maps, augmented with a novel ring-shaped attention module that explicitly rectifies vertical distortion. We provide the first systematic empirical validation that cylindrical projection significantly outperforms spherical projection for stereo matching in omnidirectional settings. The entire architecture employs only standard CNN components—no custom operators—greatly enhancing feasibility for embedded deployment. On Deep360 and 3D60 benchmarks, our method reduces depth MAE by 18.8% and 19.9%, respectively, achieving state-of-the-art performance in omnidirectional depth estimation.

1 citationsRead paper
Recent publications

Latest Papers