Selected work

Representative Papers

Improving 3D Labeling in Self-Driving by Inferring Vehicle Information using Vision Language Models

May 20, 2026

This work addresses the challenges of error-prone and inefficient manual 3D vehicle annotation in complex autonomous driving scenarios, particularly under occlusion. To overcome these limitations, the study introduces vision-language models (VLMs) into the 3D annotation pipeline for the first time, leveraging zero-shot inference to predict vehicle make, model, and generation from cropped image regions and generate accurate initial 3D bounding box dimensions. By integrating iterative prompt engineering with Vehicle Make and Model Recognition (VMMR) techniques, the proposed method demonstrates strong generalization across both public and proprietary datasets. It significantly outperforms conventional LiDAR-assisted annotation approaches, markedly improving annotation accuracy while substantially reducing human effort, especially in mitigating annotation failures caused by occlusion.

0 citationsRead paper

Expanding the Capabilities of Reinforcement Learning via Text Feedback

Feb 02, 2026

This work proposes a multi-turn reinforcement learning framework with textual feedback (RLTF) to address the limitations of sparse scalar rewards and the high cost of full demonstration supervision in large language model post-training. RLTF leverages natural language feedback—provided by humans or automated evaluators—as an efficient supervisory signal that lies between scalar rewards and complete demonstrations. The authors introduce two novel approaches, self-distillation (RLTF-SD) and feedback modeling (RLTF-FM), which effectively internalize feedback during training, enabling the model to produce higher-quality single-turn outputs at inference time without relying on external feedback. Extensive experiments on reasoning puzzles, competitive mathematics, and creative writing tasks demonstrate that RLTF significantly outperforms strong baselines, highlighting the potential of textual feedback as a scalable signal for reinforcement learning in language models.

0 citationsRead paper

To Distill or Decide? Understanding the Algorithmic Trade-off in Partially Observable Reinforcement Learning

Oct 03, 2025

This work addresses the critical algorithmic choice of whether to employ privileged expert distillation in partially observable reinforcement learning (PORL). To characterize how stochasticity in latent state dynamics impedes policy learning efficiency, we propose a perturbed Block MDP theoretical framework. Our analysis reveals that latent-state stochasticity is the decisive factor governing distillation efficacy: under high stochasticity, standard RL outperforms distillation, and the optimal latent policy may be unsuitable as a distillation teacher. We further introduce approximate decodability analysis and belief contraction theory to derive formal sufficiency conditions for expert distillation applicability. Through rigorous theoretical modeling and empirical validation, we establish—for the first time—the precise effectiveness boundary of expert distillation in PORL. This yields an interpretable, verifiable criterion for algorithm selection in partially observable settings, significantly enhancing both robustness and sample efficiency of policy learning.

0 citationsRead paper

Efficient Construction of Implicit Surface Models From a Single Image for Motion Generation

Sep 24, 2025

Reconstructing implicit signed distance fields (SDFs) from a single RGB image remains a challenging problem, as state-of-the-art approaches typically require multi-view inputs and suffer from inefficient training. This paper introduces FINS, the first lightweight framework enabling fast, high-fidelity neural surface reconstruction from a single image. Methodologically, FINS integrates geometric priors from pretrained vision foundation models, employs multi-resolution hash grid encoding, and adopts lightweight geometry- and appearance-decoding heads; it further accelerates convergence via an approximate second-order optimizer. On multiple benchmark datasets, FINS significantly outperforms prior art: training is 3.2× faster, and reconstruction accuracy—measured by Chamfer distance—improves by 37% on average. We further demonstrate FINS’s practical utility in robotic real-time surface tracking and path planning, substantially reducing both data acquisition overhead and computational cost.

0 citationsRead paper

BEVCon: Advancing Bird's Eye View Perception with Contrastive Learning

Aug 06, 2025

To address insufficient representation learning in BEV perception for autonomous driving, this paper proposes a dual-path contrastive learning framework that enforces semantic consistency at both the instance level (BEV features) and the view level (image backbone outputs). Specifically, it introduces an instance-feature contrastive module and a perspective-to-BEV contrastive module, integrated with dense pixel- and instance-level contrastive strategies. The approach enhances the discriminability and geometric robustness of BEV representations without increasing inference overhead. It seamlessly integrates into existing BEV detection pipelines by adding only a lightweight contrastive regularization term to the standard detection loss. Evaluated on the nuScenes dataset, the method achieves significant improvements across 3D detection, segmentation, and trajectory prediction—yielding up to +2.4% mAP gain over state-of-the-art baselines with negligible computational cost. This work provides the first systematic empirical validation that contrastive learning substantially improves BEV representation quality.

0 citationsRead paper

Recent publications

Latest Papers

Improving 3D Labeling in Self-Driving by Inferring Vehicle Information using Vision Language Models

May 20, 2026

This work addresses the challenges of error-prone and inefficient manual 3D vehicle annotation in complex autonomous driving scenarios, particularly under occlusion. To overcome these limitations, the study introduces vision-language models (VLMs) into the 3D annotation pipeline for the first time, leveraging zero-shot inference to predict vehicle make, model, and generation from cropped image regions and generate accurate initial 3D bounding box dimensions. By integrating iterative prompt engineering with Vehicle Make and Model Recognition (VMMR) techniques, the proposed method demonstrates strong generalization across both public and proprietary datasets. It significantly outperforms conventional LiDAR-assisted annotation approaches, markedly improving annotation accuracy while substantially reducing human effort, especially in mitigating annotation failures caused by occlusion.

0 citationsRead paper

Expanding the Capabilities of Reinforcement Learning via Text Feedback

Feb 02, 2026

This work proposes a multi-turn reinforcement learning framework with textual feedback (RLTF) to address the limitations of sparse scalar rewards and the high cost of full demonstration supervision in large language model post-training. RLTF leverages natural language feedback—provided by humans or automated evaluators—as an efficient supervisory signal that lies between scalar rewards and complete demonstrations. The authors introduce two novel approaches, self-distillation (RLTF-SD) and feedback modeling (RLTF-FM), which effectively internalize feedback during training, enabling the model to produce higher-quality single-turn outputs at inference time without relying on external feedback. Extensive experiments on reasoning puzzles, competitive mathematics, and creative writing tasks demonstrate that RLTF significantly outperforms strong baselines, highlighting the potential of textual feedback as a scalable signal for reinforcement learning in language models.

0 citationsRead paper

To Distill or Decide? Understanding the Algorithmic Trade-off in Partially Observable Reinforcement Learning

Oct 03, 2025

This work addresses the critical algorithmic choice of whether to employ privileged expert distillation in partially observable reinforcement learning (PORL). To characterize how stochasticity in latent state dynamics impedes policy learning efficiency, we propose a perturbed Block MDP theoretical framework. Our analysis reveals that latent-state stochasticity is the decisive factor governing distillation efficacy: under high stochasticity, standard RL outperforms distillation, and the optimal latent policy may be unsuitable as a distillation teacher. We further introduce approximate decodability analysis and belief contraction theory to derive formal sufficiency conditions for expert distillation applicability. Through rigorous theoretical modeling and empirical validation, we establish—for the first time—the precise effectiveness boundary of expert distillation in PORL. This yields an interpretable, verifiable criterion for algorithm selection in partially observable settings, significantly enhancing both robustness and sample efficiency of policy learning.

0 citationsRead paper

Efficient Construction of Implicit Surface Models From a Single Image for Motion Generation

Sep 24, 2025

Reconstructing implicit signed distance fields (SDFs) from a single RGB image remains a challenging problem, as state-of-the-art approaches typically require multi-view inputs and suffer from inefficient training. This paper introduces FINS, the first lightweight framework enabling fast, high-fidelity neural surface reconstruction from a single image. Methodologically, FINS integrates geometric priors from pretrained vision foundation models, employs multi-resolution hash grid encoding, and adopts lightweight geometry- and appearance-decoding heads; it further accelerates convergence via an approximate second-order optimizer. On multiple benchmark datasets, FINS significantly outperforms prior art: training is 3.2× faster, and reconstruction accuracy—measured by Chamfer distance—improves by 37% on average. We further demonstrate FINS’s practical utility in robotic real-time surface tracking and path planning, substantially reducing both data acquisition overhead and computational cost.

0 citationsRead paper

BEVCon: Advancing Bird's Eye View Perception with Contrastive Learning

Aug 06, 2025

To address insufficient representation learning in BEV perception for autonomous driving, this paper proposes a dual-path contrastive learning framework that enforces semantic consistency at both the instance level (BEV features) and the view level (image backbone outputs). Specifically, it introduces an instance-feature contrastive module and a perspective-to-BEV contrastive module, integrated with dense pixel- and instance-level contrastive strategies. The approach enhances the discriminability and geometric robustness of BEV representations without increasing inference overhead. It seamlessly integrates into existing BEV detection pipelines by adding only a lightweight contrastive regularization term to the standard detection loss. Evaluated on the nuScenes dataset, the method achieves significant improvements across 3D detection, segmentation, and trajectory prediction—yielding up to +2.4% mAP gain over state-of-the-art baselines with negligible computational cost. This work provides the first systematic empirical validation that contrastive learning substantially improves BEV representation quality.

0 citationsRead paper