Selected work

Representative Papers

GRACE: Boosting Video MLLMs with Grounded Action-Centric Evidence for Viewer Sentiment Prediction

Jun 15, 2026

This work addresses the challenge of predicting viewer sentiment in video advertisements, where fine-grained emotion-relevant behaviors and visual cues are difficult to capture from full-frame inputs. The authors propose an action-centric, structured evidence enhancement framework that extracts temporal subject-predicate-object triplets, crops visual patches of participating entities, and integrates visible text to construct explicit, spatially localizable multimodal reasoning cues. This approach uniquely combines action triplets with entity-specific visual crops to guide interpretable sentiment reasoning in multimodal large language models (Qwen2.5-VL/Qwen3-VL). Evaluated on the Pitts dataset, the method significantly outperforms baseline approaches, and transfer experiments on AdsQA and TVQA subsets demonstrate its strong generalization capability.

0 citationsRead paper

EvolveSignal: A Large Language Model Powered Coding Agent for Discovering Traffic Signal Control Algorithms

Sep 03, 2025

Conventional fixed-time traffic signal control relies heavily on manual formulae (e.g., Webster’s method) and empirical parameter tuning, exhibiting poor adaptability and limited effectiveness in heterogeneous congestion scenarios. Method: This paper proposes a large language model (LLM)-driven programming agent framework that uniquely integrates LLMs, program synthesis, and evolutionary search for autonomous design of traffic signal control algorithms. It represents policies as executable Python functions and iteratively optimizes them via closed-loop simulation feedback—requiring no human intervention. Contribution/Results: The framework automatically discovers novel, interpretable, and deployable control algorithms. In single-intersection experiments, the synthesized algorithm reduces average vehicle delay by 20.1% and stops per vehicle by 47.1% compared to Webster’s method, demonstrating substantial performance gains and yielding structured, actionable insights into optimization mechanisms.

0 citationsRead paper

SAMST: A Transformer framework based on SAM pseudo label filtering for remote sensing semi-supervised semantic segmentation

Jul 16, 2025

Remote sensing semantic segmentation faces two critical challenges: scarcity of annotated data and poor generalizability of public datasets—stemming from inter-dataset variations in spatial resolution and inconsistent class definitions. To address these, we propose SAMST, a semi-supervised framework that synergistically integrates the zero-shot boundary-aware capability of the Segment Anything Model (SAM) with the efficient trainability of lightweight segmentation networks. Methodologically, SAMST introduces three key components: threshold-based filtering, learnable prompt generation, and label stitching, enabling iterative refinement of pseudo-labels. Technically, it unifies a Transformer-based backbone, SAM-guided pseudo-label self-training, prompt learning, and connected-component optimization. Extensive evaluation on the Potsdam dataset demonstrates that SAMST significantly enhances pseudo-label quality and segmentation accuracy, effectively alleviating the annotation bottleneck in remote sensing imagery.

0 citationsRead paper

Chat2SPaT: A Large Language Model Based Tool for Automating Traffic Signal Control Plan Management

Jul 04, 2025

Traffic signal timing plan development relies heavily on manual configuration, leading to error-prone and inefficient processes—especially in multi-period/multi-day scheduling scenarios. To address this, we propose the first end-to-end natural language (NL)–to–SPaT (Signal Phase and Timing) generation framework for traffic signal control, leveraging large language models (LLMs) with structured prompt engineering to accurately parse semi-structured, ambiguous user descriptions into standardized phase sequences and timing parameters (output as JSON). The framework supports both ring-and-barrier and phase-based plan generation, iterative conversational editing, and precise intra-cycle phase positioning. We further introduce the first NL understanding benchmark for traffic signal control, releasing open-source code, prompt templates, and a bilingual test set of 300+ instances. Experiments demonstrate >94% plan generation accuracy, establishing a reusable, LLM-driven automation paradigm for intelligent transportation systems.

0 citationsRead paper

Birdie: Natural Language-Driven Table Discovery Using Differentiable Search Index

Apr 30, 2025

To address error accumulation and insufficient semantic alignment in conventional dense retrieval paradigms for natural language–driven table discovery, this paper proposes an end-to-end differentiable search indexing framework that unifies indexing and retrieval within an encoder-decoder language model. Its key contributions are: (1) prefix-aware table identifier embedding, enhancing structured semantic modeling; (2) large-model-driven synthetic query generation, improving training data quality and generalization; and (3) a parameter-isolated dynamic index update mechanism, substantially mitigating catastrophic forgetting in continual learning. Experiments demonstrate that the method achieves a 16.8% absolute accuracy gain over state-of-the-art dense retrieval models on table discovery tasks and reduces forgetting rate by over 90% in continual indexing scenarios.

0 citationsRead paper

Recent publications

Latest Papers

GRACE: Boosting Video MLLMs with Grounded Action-Centric Evidence for Viewer Sentiment Prediction

Jun 15, 2026

This work addresses the challenge of predicting viewer sentiment in video advertisements, where fine-grained emotion-relevant behaviors and visual cues are difficult to capture from full-frame inputs. The authors propose an action-centric, structured evidence enhancement framework that extracts temporal subject-predicate-object triplets, crops visual patches of participating entities, and integrates visible text to construct explicit, spatially localizable multimodal reasoning cues. This approach uniquely combines action triplets with entity-specific visual crops to guide interpretable sentiment reasoning in multimodal large language models (Qwen2.5-VL/Qwen3-VL). Evaluated on the Pitts dataset, the method significantly outperforms baseline approaches, and transfer experiments on AdsQA and TVQA subsets demonstrate its strong generalization capability.

0 citationsRead paper

EvolveSignal: A Large Language Model Powered Coding Agent for Discovering Traffic Signal Control Algorithms

Sep 03, 2025

Conventional fixed-time traffic signal control relies heavily on manual formulae (e.g., Webster’s method) and empirical parameter tuning, exhibiting poor adaptability and limited effectiveness in heterogeneous congestion scenarios. Method: This paper proposes a large language model (LLM)-driven programming agent framework that uniquely integrates LLMs, program synthesis, and evolutionary search for autonomous design of traffic signal control algorithms. It represents policies as executable Python functions and iteratively optimizes them via closed-loop simulation feedback—requiring no human intervention. Contribution/Results: The framework automatically discovers novel, interpretable, and deployable control algorithms. In single-intersection experiments, the synthesized algorithm reduces average vehicle delay by 20.1% and stops per vehicle by 47.1% compared to Webster’s method, demonstrating substantial performance gains and yielding structured, actionable insights into optimization mechanisms.

0 citationsRead paper

SAMST: A Transformer framework based on SAM pseudo label filtering for remote sensing semi-supervised semantic segmentation

Jul 16, 2025

Remote sensing semantic segmentation faces two critical challenges: scarcity of annotated data and poor generalizability of public datasets—stemming from inter-dataset variations in spatial resolution and inconsistent class definitions. To address these, we propose SAMST, a semi-supervised framework that synergistically integrates the zero-shot boundary-aware capability of the Segment Anything Model (SAM) with the efficient trainability of lightweight segmentation networks. Methodologically, SAMST introduces three key components: threshold-based filtering, learnable prompt generation, and label stitching, enabling iterative refinement of pseudo-labels. Technically, it unifies a Transformer-based backbone, SAM-guided pseudo-label self-training, prompt learning, and connected-component optimization. Extensive evaluation on the Potsdam dataset demonstrates that SAMST significantly enhances pseudo-label quality and segmentation accuracy, effectively alleviating the annotation bottleneck in remote sensing imagery.

0 citationsRead paper

Chat2SPaT: A Large Language Model Based Tool for Automating Traffic Signal Control Plan Management

Jul 04, 2025

Traffic signal timing plan development relies heavily on manual configuration, leading to error-prone and inefficient processes—especially in multi-period/multi-day scheduling scenarios. To address this, we propose the first end-to-end natural language (NL)–to–SPaT (Signal Phase and Timing) generation framework for traffic signal control, leveraging large language models (LLMs) with structured prompt engineering to accurately parse semi-structured, ambiguous user descriptions into standardized phase sequences and timing parameters (output as JSON). The framework supports both ring-and-barrier and phase-based plan generation, iterative conversational editing, and precise intra-cycle phase positioning. We further introduce the first NL understanding benchmark for traffic signal control, releasing open-source code, prompt templates, and a bilingual test set of 300+ instances. Experiments demonstrate >94% plan generation accuracy, establishing a reusable, LLM-driven automation paradigm for intelligent transportation systems.

0 citationsRead paper

Birdie: Natural Language-Driven Table Discovery Using Differentiable Search Index

Apr 30, 2025

To address error accumulation and insufficient semantic alignment in conventional dense retrieval paradigms for natural language–driven table discovery, this paper proposes an end-to-end differentiable search indexing framework that unifies indexing and retrieval within an encoder-decoder language model. Its key contributions are: (1) prefix-aware table identifier embedding, enhancing structured semantic modeling; (2) large-model-driven synthetic query generation, improving training data quality and generalization; and (3) a parameter-isolated dynamic index update mechanism, substantially mitigating catastrophic forgetting in continual learning. Experiments demonstrate that the method achieves a 16.8% absolute accuracy gain over state-of-the-art dense retrieval models on table discovery tasks and reduces forgetting rate by over 90% in continual indexing scenarios.

0 citationsRead paper