Selected work

Representative Papers

Neural timescales from a computational perspective

Sep 04, 2024arXiv.org

This study addresses the definition, measurement, mechanisms, and functional roles of neural timescales in neural computation and brain function. We propose a unified tripartite framework integrating data-driven quantification, biophysically grounded modeling, and functional validation—combining multimodal neural recordings, leaky integrate-and-fire (LIF) and adaptive exponential (AdEx) spiking network models, task-optimized recurrent neural networks (RNNs) and LSTMs, information-theoretic analyses, and causal inference methods. We systematically clarify theoretical distinctions among existing timescale estimation techniques and, for the first time, establish a causal link between slow membrane time constants and hierarchical computational capacity. Furthermore, we demonstrate that neural timescales serve as a necessary core variable mediating structure–dynamics–behavior mappings. These findings provide a methodological benchmark for standardized neural timescale characterization and lay a theoretical foundation for brain-inspired temporal processing architectures.

2 citationsRead paper

FinRank: An Evidence-Grounded Benchmark for Financial Question Answering and Retrieval over SEC Filings

Aug 07, 2026

This work addresses the challenge of evidence provenance errors in financial question answering, which frequently arise due to similar disclosures across sections, reporting periods, and companies in SEC filings. The authors introduce the first evidence-oriented benchmark for 10-K/10-Q–based financial QA and retrieval, requiring systems to precisely align entities, reporting periods, and contextual cues to locate supporting evidence. They contribute a novel set of human-annotated hard negatives—including confounding passages—and define three distinct evaluation tasks: retrieval, re-ranking, and hard negative discrimination. Evaluated on 1,185 annotated instances, even the strongest 7B-scale model achieves only 44.8% Recall@10 using BM25, instruction-tuned embeddings, and pairwise re-ranking. Performance drops by 13.0–20.5 percentage points when hard negatives are introduced, underscoring the task’s difficulty.

0 citationsRead paper

MOT-SR: Multi-Objective Tool-Augmented Scientific Equation Discovery with Large Language Models

Jul 31, 2026

This work addresses critical limitations in existing large language model (LLM)-driven symbolic regression methods, which typically neglect variable dependency analysis and optimize solely for fitting error while disregarding structural complexity and generalization capability, often leading to suboptimal local solutions. To overcome these issues, the authors propose MOT-SR, a novel framework that introduces multi-objective optimization and tool-augmented mechanisms into symbolic regression for the first time. MOT-SR leverages external analytical tools to extract structural priors and employs dual LLM modules in a closed-loop iterative process to generate candidate equations. It jointly optimizes accuracy, complexity, and generalization, dynamically maintaining a Pareto front to enable co-evolution of search strategies and equation structures. Evaluated on 40 standard benchmarks, MOT-SR significantly outperforms current approaches and achieves the lowest trajectory integration error in modeling extreme mass-ratio inspiral orbits, demonstrating its effectiveness and reliability in scientific dynamical modeling.

0 citationsRead paper

Using Hierarchical Controlled Vocabularies to Understand CLIP Retrieval Failures in Historical Photo Collections

Jul 22, 2026

This study investigates the instability of CLIP models in retrieving historical photographs from GLAM institution collections, a failure mode not previously explained through the lens of controlled vocabulary structure. For the first time, it correlates the AAT thesaurus’s root facets and hierarchical depth with CLIP’s retrieval performance. Through visual embedding clustering, image–text similarity evaluation, and fine-tuning experiments across three historical photo datasets, the work systematically analyzes CLIP’s visual consistency and image–text alignment. The findings reveal that these two properties are nearly uncorrelated and jointly define distinct failure types: root facets significantly influence visual consistency; image–text alignment for shallow-level terms is more amenable to improvement via fine-tuning; and terms exhibiting tight visual clusters yet deviant semantic labels yield the poorest retrieval performance.

0 citationsRead paper

Learning to Detect Cross-Modal Negation: An Analysis of Latent Representations and an Attention-Based Solution

Jul 20, 2026

This study addresses the challenge that existing vision-language models struggle to detect high-level semantic concepts such as negation, primarily due to the non-separability of negation classes in their latent spaces. The work reveals, for the first time, that textual negation can be represented independently, whereas visual negation is inherently dependent on linguistic context. To tackle this issue, the authors propose a novel cross-attention architecture that explicitly models inter-modal dependencies. The approach integrates self-supervised video representations from JEPA2 with automatic annotations generated by Qwen2.5-VL, and is evaluated on a dataset of 3,222 political video–text pairs. Experimental results demonstrate that the method achieves up to a 7.03% improvement in F1 score over unimodal baselines, advancing multimodal representation learning for temporal negation modeling and semantic alignment.

0 citationsRead paper

Recent publications

Latest Papers

FinRank: An Evidence-Grounded Benchmark for Financial Question Answering and Retrieval over SEC Filings

Aug 07, 2026

This work addresses the challenge of evidence provenance errors in financial question answering, which frequently arise due to similar disclosures across sections, reporting periods, and companies in SEC filings. The authors introduce the first evidence-oriented benchmark for 10-K/10-Q–based financial QA and retrieval, requiring systems to precisely align entities, reporting periods, and contextual cues to locate supporting evidence. They contribute a novel set of human-annotated hard negatives—including confounding passages—and define three distinct evaluation tasks: retrieval, re-ranking, and hard negative discrimination. Evaluated on 1,185 annotated instances, even the strongest 7B-scale model achieves only 44.8% Recall@10 using BM25, instruction-tuned embeddings, and pairwise re-ranking. Performance drops by 13.0–20.5 percentage points when hard negatives are introduced, underscoring the task’s difficulty.

0 citationsRead paper

MOT-SR: Multi-Objective Tool-Augmented Scientific Equation Discovery with Large Language Models

Jul 31, 2026

This work addresses critical limitations in existing large language model (LLM)-driven symbolic regression methods, which typically neglect variable dependency analysis and optimize solely for fitting error while disregarding structural complexity and generalization capability, often leading to suboptimal local solutions. To overcome these issues, the authors propose MOT-SR, a novel framework that introduces multi-objective optimization and tool-augmented mechanisms into symbolic regression for the first time. MOT-SR leverages external analytical tools to extract structural priors and employs dual LLM modules in a closed-loop iterative process to generate candidate equations. It jointly optimizes accuracy, complexity, and generalization, dynamically maintaining a Pareto front to enable co-evolution of search strategies and equation structures. Evaluated on 40 standard benchmarks, MOT-SR significantly outperforms current approaches and achieves the lowest trajectory integration error in modeling extreme mass-ratio inspiral orbits, demonstrating its effectiveness and reliability in scientific dynamical modeling.

0 citationsRead paper

Using Hierarchical Controlled Vocabularies to Understand CLIP Retrieval Failures in Historical Photo Collections

Jul 22, 2026

This study investigates the instability of CLIP models in retrieving historical photographs from GLAM institution collections, a failure mode not previously explained through the lens of controlled vocabulary structure. For the first time, it correlates the AAT thesaurus’s root facets and hierarchical depth with CLIP’s retrieval performance. Through visual embedding clustering, image–text similarity evaluation, and fine-tuning experiments across three historical photo datasets, the work systematically analyzes CLIP’s visual consistency and image–text alignment. The findings reveal that these two properties are nearly uncorrelated and jointly define distinct failure types: root facets significantly influence visual consistency; image–text alignment for shallow-level terms is more amenable to improvement via fine-tuning; and terms exhibiting tight visual clusters yet deviant semantic labels yield the poorest retrieval performance.

0 citationsRead paper

Learning to Detect Cross-Modal Negation: An Analysis of Latent Representations and an Attention-Based Solution

Jul 20, 2026

This study addresses the challenge that existing vision-language models struggle to detect high-level semantic concepts such as negation, primarily due to the non-separability of negation classes in their latent spaces. The work reveals, for the first time, that textual negation can be represented independently, whereas visual negation is inherently dependent on linguistic context. To tackle this issue, the authors propose a novel cross-attention architecture that explicitly models inter-modal dependencies. The approach integrates self-supervised video representations from JEPA2 with automatic annotations generated by Qwen2.5-VL, and is evaluated on a dataset of 3,222 political video–text pairs. Experimental results demonstrate that the method achieves up to a 7.03% improvement in F1 score over unimodal baselines, advancing multimodal representation learning for temporal negation modeling and semantic alignment.

0 citationsRead paper

Faster Algorithms for Deciding the Unbiased Maker-Breaker Triangle Game on General Graphs

Jul 20, 2026

This work addresses the computational complexity of determining the winner in the unbiased Maker-Breaker triangle game on general graphs. Introducing, for the first time, an edge-triangle incidence graph model combined with structural characterizations and monotone strategies, the paper establishes necessary and sufficient conditions for Maker’s victory. Leveraging this framework, it designs polynomial-time algorithms that reduce the decision complexity from $O(n^{16})$ to $O(n^7)$ for arbitrary graphs; achieve $O(n^{\omega+1})$ when the graph contains a $K_4$ and its incidence graph is connected; and attain $O(n^3)$ for $K_4$-free cactus-like incidence graphs. Furthermore, the study presents a linear-time reduction from triangle detection to game outcome determination, significantly enhancing computational efficiency and broadening the class of applicable graphs.

0 citationsRead paper