Selected work

Representative Papers

Proof-of-Perception: Certified Tool-Using Multimodal Reasoning with Compositional Conformal Guarantees

Feb 27, 2026

This work addresses the unreliability of multimodal reasoning, which often leads to error propagation and hallucination due to insufficient uncertainty calibration. The authors propose a novel executable reasoning graph framework that models perceptual and logical operations as nodes producing conformal prediction sets, thereby providing calibrated, stepwise uncertainty guarantees. A lightweight controller dynamically schedules tool invocations based on available computational budget. This approach establishes, for the first time, a compositional conformal guarantee mechanism that enables verifiable, evidence-backed reasoning, suppresses error accumulation, and allows controllable trade-offs between computation and accuracy. Experiments demonstrate consistent superiority over strong baselines across document, chart, and multi-image question-answering benchmarks in terms of performance, reliability, and computational efficiency.

0 citationsRead paper

SCOUT: A Lightweight Framework for Scenario Coverage Assessment in Autonomous Driving

Oct 28, 2025

Current autonomous driving scenario coverage evaluation relies either on labor-intensive manual annotation or computationally expensive large vision-language models (LVLMs), hindering scalable deployment. This paper proposes SCOUT, a lightweight framework that enables efficient coverage assessment for the first time using precomputed sensor-derived implicit representations. SCOUT employs knowledge distillation to transfer semantic labels—originally generated by LVLMs—into a compact surrogate network, which directly predicts coverage scores from perception features, thereby eliminating both manual annotation and real-time LVLM inference. The method integrates implicit representation extraction, an optimized neural architecture, and rigorous validation on large-scale real-world driving data. Experiments demonstrate that SCOUT maintains high accuracy while accelerating inference by two orders of magnitude and substantially reducing computational cost. As a result, SCOUT establishes a scalable, low-cost, automated paradigm for evaluating scenario coverage in large-scale autonomous driving systems.

0 citationsRead paper

Learning Causal Structure Distributions for Robust Planning

Aug 08, 2025

This work addresses the insufficient planning robustness and excessive computational overhead in robotic dynamic modeling caused by uncertainty in causal structure. We propose a lightweight dynamic learning method based on probabilistic causal structure modeling. Our approach treats the graph structure of structural causal models as a random variable and learns its posterior distribution—rather than assuming a fixed topology. We design an encoder–multi-decoder probabilistic framework that jointly optimizes structure distribution sampling and functional relationship modeling in latent space, leveraging system sparsity priors to enhance generalization. Crucially, explicit incorporation of structural uncertainty guides representation learning to balance robustness and efficiency. Evaluations on simulated and real-world platforms—including manipulator and mobile robots—demonstrate significant improvements: 23.6% higher planning success rate under environmental disturbances and input noise, and 1.8× faster inference speed.

0 citationsRead paper

Fractional Collisions: A Framework for Risk Estimation of Counterfactual Conflicts using Autonomous Driving Behavior Simulations

Jun 09, 2025

This study addresses the challenge of collision risk assessment for autonomous driving systems (ADS) by introducing the “scored collision” paradigm—a decoupled, probabilistic quantification of both bodily injury and property damage risks. Methodologically, it constructs counterfactual simulation scenarios from real-world driving data, integrating multi-agent conflict detection, agent role identification, response-point localization, probabilistic human driver behavior modeling, counterfactual trajectory sampling, and physics-informed collision severity estimation. Its key contributions are the first fair, apples-to-apples risk comparison between ADS and human drivers, and native support for multi-source uncertainty fusion. Evaluated on the SHRP2/Nexar dataset, the method achieves 99% prediction accuracy. Empirical evaluation shows that a specific ADS version reduces natural collisions by 4× and lowers scored collision risk by 62%; over 250,000 miles of testing, 96% of ADS-initiated conflicts exhibit lower risk than the human baseline.

0 citationsRead paper

Evaluating Global Geo-alignment for Precision Learned Autonomous Vehicle Localization using Aerial Data

Mar 18, 2025

This work addresses the insufficient geospatial alignment accuracy between aerial imagery and vehicle-mounted sensor data in learning-based autonomous driving localization systems. We first systematically demonstrate that global geo-alignment quality critically impacts overall localization performance. To this end, we propose two factor-graph-optimization-based cross-view geometric alignment methods that tightly integrate multimodal feature matching with deep-learning-based localization models. Rigorous ablation studies quantify the contribution of each component. Evaluated on a 1600-km real-world driving dataset, our approach achieves a mean positioning error of 0.28 m and heading error of 0.47°, significantly outperforming prior methods and meeting L4-level autonomous driving requirements. The core innovation lies in the deep coupling of end-to-end learning with geometry-prior-driven global alignment, establishing a novel paradigm for trustworthy, multi-source, heterogeneous sensor data alignment.

0 citationsRead paper

Recent publications

Latest Papers

Proof-of-Perception: Certified Tool-Using Multimodal Reasoning with Compositional Conformal Guarantees

Feb 27, 2026

This work addresses the unreliability of multimodal reasoning, which often leads to error propagation and hallucination due to insufficient uncertainty calibration. The authors propose a novel executable reasoning graph framework that models perceptual and logical operations as nodes producing conformal prediction sets, thereby providing calibrated, stepwise uncertainty guarantees. A lightweight controller dynamically schedules tool invocations based on available computational budget. This approach establishes, for the first time, a compositional conformal guarantee mechanism that enables verifiable, evidence-backed reasoning, suppresses error accumulation, and allows controllable trade-offs between computation and accuracy. Experiments demonstrate consistent superiority over strong baselines across document, chart, and multi-image question-answering benchmarks in terms of performance, reliability, and computational efficiency.

0 citationsRead paper

SCOUT: A Lightweight Framework for Scenario Coverage Assessment in Autonomous Driving

Oct 28, 2025

Current autonomous driving scenario coverage evaluation relies either on labor-intensive manual annotation or computationally expensive large vision-language models (LVLMs), hindering scalable deployment. This paper proposes SCOUT, a lightweight framework that enables efficient coverage assessment for the first time using precomputed sensor-derived implicit representations. SCOUT employs knowledge distillation to transfer semantic labels—originally generated by LVLMs—into a compact surrogate network, which directly predicts coverage scores from perception features, thereby eliminating both manual annotation and real-time LVLM inference. The method integrates implicit representation extraction, an optimized neural architecture, and rigorous validation on large-scale real-world driving data. Experiments demonstrate that SCOUT maintains high accuracy while accelerating inference by two orders of magnitude and substantially reducing computational cost. As a result, SCOUT establishes a scalable, low-cost, automated paradigm for evaluating scenario coverage in large-scale autonomous driving systems.

0 citationsRead paper

Learning Causal Structure Distributions for Robust Planning

Aug 08, 2025

This work addresses the insufficient planning robustness and excessive computational overhead in robotic dynamic modeling caused by uncertainty in causal structure. We propose a lightweight dynamic learning method based on probabilistic causal structure modeling. Our approach treats the graph structure of structural causal models as a random variable and learns its posterior distribution—rather than assuming a fixed topology. We design an encoder–multi-decoder probabilistic framework that jointly optimizes structure distribution sampling and functional relationship modeling in latent space, leveraging system sparsity priors to enhance generalization. Crucially, explicit incorporation of structural uncertainty guides representation learning to balance robustness and efficiency. Evaluations on simulated and real-world platforms—including manipulator and mobile robots—demonstrate significant improvements: 23.6% higher planning success rate under environmental disturbances and input noise, and 1.8× faster inference speed.

0 citationsRead paper

Fractional Collisions: A Framework for Risk Estimation of Counterfactual Conflicts using Autonomous Driving Behavior Simulations

Jun 09, 2025

This study addresses the challenge of collision risk assessment for autonomous driving systems (ADS) by introducing the “scored collision” paradigm—a decoupled, probabilistic quantification of both bodily injury and property damage risks. Methodologically, it constructs counterfactual simulation scenarios from real-world driving data, integrating multi-agent conflict detection, agent role identification, response-point localization, probabilistic human driver behavior modeling, counterfactual trajectory sampling, and physics-informed collision severity estimation. Its key contributions are the first fair, apples-to-apples risk comparison between ADS and human drivers, and native support for multi-source uncertainty fusion. Evaluated on the SHRP2/Nexar dataset, the method achieves 99% prediction accuracy. Empirical evaluation shows that a specific ADS version reduces natural collisions by 4× and lowers scored collision risk by 62%; over 250,000 miles of testing, 96% of ADS-initiated conflicts exhibit lower risk than the human baseline.

0 citationsRead paper

Evaluating Global Geo-alignment for Precision Learned Autonomous Vehicle Localization using Aerial Data

Mar 18, 2025

This work addresses the insufficient geospatial alignment accuracy between aerial imagery and vehicle-mounted sensor data in learning-based autonomous driving localization systems. We first systematically demonstrate that global geo-alignment quality critically impacts overall localization performance. To this end, we propose two factor-graph-optimization-based cross-view geometric alignment methods that tightly integrate multimodal feature matching with deep-learning-based localization models. Rigorous ablation studies quantify the contribution of each component. Evaluated on a 1600-km real-world driving dataset, our approach achieves a mean positioning error of 0.28 m and heading error of 0.47°, significantly outperforming prior methods and meeting L4-level autonomous driving requirements. The core innovation lies in the deep coupling of end-to-end learning with geometry-prior-driven global alignment, establishing a novel paradigm for trustworthy, multi-source, heterogeneous sensor data alignment.

0 citationsRead paper