Selected work

Representative Papers

Latent-MOPD: Latent Multi-Teacher On-Policy Distillation

Oct 01, 2026

This study addresses the limitation of existing multi-teacher online distillation methods, which exploit only output distributions while neglecting internal representations. To overcome this, we propose the first representation-level multi-teacher online distillation framework. Specifically, this work introduces a novel dual-channel mechanism that integrates hidden states with token predictions to coordinate knowledge from multiple experts without additional training. Furthermore, it incorporates late-layer target selection, shared projection bridging, and domain-grouped updating techniques to enable dynamic adjustment of supervision weights. Extensive evaluations across nine benchmarks, including mathematical reasoning and code generation, demonstrate that the student model consistently outperforms baseline approaches and surpasses most single-domain expert teachers in overall performance.

0 citationsRead paper

MINT: Min-Selection Preference Distillation for Balanced Multi-Objective Alignment

Aug 14, 2026

This study addresses the optimization imbalance caused by reward aggregation in multi-objective preference training by proposing MINT. Grounded in the limit theory of generalized means, this method transforms traditional weighted summation into weakest-target ranking. By integrating minimal selection preference distillation with Direct Preference Optimization (DPO), MINT generalizes from weighted sums to worst-case selection with merely a single-line code modification. Experimental results demonstrate that MINT significantly improves scores for underperforming objectives while reducing performance disparity. Notably, it surpasses human expert performance in emotional support tasks, offering an efficient solution for balanced multi-objective alignment.

0 citationsRead paper

ProActor: Timing-Aware Reinforcement Learning for Proactive Task Scheduling Agents

May 24, 2026

This work addresses the lack of a generalizable end-to-end framework for effectively evaluating and optimizing the proactive behavior of task-oriented agents—specifically, their ability to trigger software actions at appropriate moments. The authors propose ProActor, a unified framework that innovatively integrates a RULER-based scoring reward mechanism with a phase-aware composite reward, enabling, for the first time, joint optimization of timing quality and action alignment in proactive behavior. They also introduce a scalable pipeline that automatically generates full opportunity time-window annotations. Leveraging GRPO reinforcement learning, LoRA fine-tuning, and a 4-bit quantized Qwen2.5-14B model—combined with an adaptive inference cluster and a single-node multi-GPU DDP training architecture (ART-F)—the approach achieves significant improvements in proactive timing accuracy on two new datasets while maintaining action consistency and accelerating training efficiency by 4–8×.

0 citationsRead paper

SELFDOUBT: Uncertainty Quantification for Reasoning LLMs via the Hedge-to-Verify Ratio

Apr 07, 2026

This work addresses the challenge of uncertainty estimation for large language models in black-box API settings, where existing methods are either computationally expensive or rely on unreliable single-pass signals. The authors propose SELFDOUBT, a novel framework that extracts the “Hesitation-to-Verification Ratio” (HVR) from a single natural language reasoning trajectory as a behavioral signal of uncertainty. This approach enables high-precision confidence gating at zero additional computational cost, without requiring multiple generations or internal model access. Evaluated across seven models and three benchmarks, SELFDOUBT demonstrates strong performance: trajectories without hesitation achieve 96% accuracy, while the full SELFDOUBT score outperforms semantic entropy at one-tenth the inference cost. In cascaded deployment, it achieves 90% accuracy with 71% coverage, showcasing its practical efficacy for real-world black-box APIs.

0 citationsRead paper

GaussFusion: Improving 3D Reconstruction in the Wild with A Geometry-Informed Video Generator

Mar 26, 2026

This work addresses floating artifacts, flickering, and blurriness in 3D Gaussian splatting reconstructions of wild scenes, which arise from camera pose errors, insufficient coverage, and noisy geometric initialization. To resolve these issues, the authors propose a geometry-guided video-to-video generation approach that refines rendered outputs with temporal consistency. Their method introduces, for the first time, a geometry-aware video generation framework that constructs a Gaussian primitive video buffer using depth, normals, opacity, and covariance. Combined with a synthetic data training strategy capable of simulating diverse degradation patterns, this approach significantly enhances generalization. The method achieves state-of-the-art performance on novel view synthesis benchmarks, with an efficient variant running at 21 FPS, enabling interactive applications.

0 citationsRead paper

Recent publications

Latest Papers

Latent-MOPD: Latent Multi-Teacher On-Policy Distillation

Oct 01, 2026

This study addresses the limitation of existing multi-teacher online distillation methods, which exploit only output distributions while neglecting internal representations. To overcome this, we propose the first representation-level multi-teacher online distillation framework. Specifically, this work introduces a novel dual-channel mechanism that integrates hidden states with token predictions to coordinate knowledge from multiple experts without additional training. Furthermore, it incorporates late-layer target selection, shared projection bridging, and domain-grouped updating techniques to enable dynamic adjustment of supervision weights. Extensive evaluations across nine benchmarks, including mathematical reasoning and code generation, demonstrate that the student model consistently outperforms baseline approaches and surpasses most single-domain expert teachers in overall performance.

0 citationsRead paper

MINT: Min-Selection Preference Distillation for Balanced Multi-Objective Alignment

Aug 14, 2026

This study addresses the optimization imbalance caused by reward aggregation in multi-objective preference training by proposing MINT. Grounded in the limit theory of generalized means, this method transforms traditional weighted summation into weakest-target ranking. By integrating minimal selection preference distillation with Direct Preference Optimization (DPO), MINT generalizes from weighted sums to worst-case selection with merely a single-line code modification. Experimental results demonstrate that MINT significantly improves scores for underperforming objectives while reducing performance disparity. Notably, it surpasses human expert performance in emotional support tasks, offering an efficient solution for balanced multi-objective alignment.

0 citationsRead paper

ProActor: Timing-Aware Reinforcement Learning for Proactive Task Scheduling Agents

May 24, 2026

This work addresses the lack of a generalizable end-to-end framework for effectively evaluating and optimizing the proactive behavior of task-oriented agents—specifically, their ability to trigger software actions at appropriate moments. The authors propose ProActor, a unified framework that innovatively integrates a RULER-based scoring reward mechanism with a phase-aware composite reward, enabling, for the first time, joint optimization of timing quality and action alignment in proactive behavior. They also introduce a scalable pipeline that automatically generates full opportunity time-window annotations. Leveraging GRPO reinforcement learning, LoRA fine-tuning, and a 4-bit quantized Qwen2.5-14B model—combined with an adaptive inference cluster and a single-node multi-GPU DDP training architecture (ART-F)—the approach achieves significant improvements in proactive timing accuracy on two new datasets while maintaining action consistency and accelerating training efficiency by 4–8×.

0 citationsRead paper

SELFDOUBT: Uncertainty Quantification for Reasoning LLMs via the Hedge-to-Verify Ratio

Apr 07, 2026

This work addresses the challenge of uncertainty estimation for large language models in black-box API settings, where existing methods are either computationally expensive or rely on unreliable single-pass signals. The authors propose SELFDOUBT, a novel framework that extracts the “Hesitation-to-Verification Ratio” (HVR) from a single natural language reasoning trajectory as a behavioral signal of uncertainty. This approach enables high-precision confidence gating at zero additional computational cost, without requiring multiple generations or internal model access. Evaluated across seven models and three benchmarks, SELFDOUBT demonstrates strong performance: trajectories without hesitation achieve 96% accuracy, while the full SELFDOUBT score outperforms semantic entropy at one-tenth the inference cost. In cascaded deployment, it achieves 90% accuracy with 71% coverage, showcasing its practical efficacy for real-world black-box APIs.

0 citationsRead paper

GaussFusion: Improving 3D Reconstruction in the Wild with A Geometry-Informed Video Generator

Mar 26, 2026

This work addresses floating artifacts, flickering, and blurriness in 3D Gaussian splatting reconstructions of wild scenes, which arise from camera pose errors, insufficient coverage, and noisy geometric initialization. To resolve these issues, the authors propose a geometry-guided video-to-video generation approach that refines rendered outputs with temporal consistency. Their method introduces, for the first time, a geometry-aware video generation framework that constructs a Gaussian primitive video buffer using depth, normals, opacity, and covariance. Combined with a synthetic data training strategy capable of simulating diverse degradation patterns, this approach significantly enhances generalization. The method achieves state-of-the-art performance on novel view synthesis benchmarks, with an efficient variant running at 21 FPS, enabling interactive applications.

0 citationsRead paper