H. B. Jiang

Qiyong Zhong, Mao Zheng, Mingyang Song et al.

Tool-integrated reasoning (TIR) is difficult to scale to small language models due to instability in long-horizon tool interactions and limited model capacity. While reinforcement learning methods like group relative policy optimization provide only sparse outcome-level rewards. Recently, on-policy distillation (OPD) has gained popularity by supplying dense token-level supervision from a teacher on student-generated trajectories. However, our experiments indicate that applying OPD to TIR leads to a critical failure mode: erroneous tool calls tend to cascade across subsequent reasoning steps, progressively amplifying student-teacher divergence and rendering the teacher's token-level supervision increasingly unreliable. To address this, we propose SOD, a step-wise on-policy distillation framework for small language model agents, which adaptively reweights distillation strength at each step based on step-level divergence. Therefore, SOD can attenuate potentially misleading teacher signals in high-divergence regions while preserving dense guidance in well-aligned states. Experiments on challenging math, science, and code benchmarks show that SOD achieves up to 20.86% improvement over the second-best baseline. Notably, our 0.6B student achieves 26.13% on AIME 2025, demonstrating effective transfer of agentic reasoning to lightweight models. Our code is available at https://github.com/YoungZ365/SOD.

Junfeng Fang, Zhepei Hong, Mao Zheng et al.

On-policy distillation (OPD) is a powerful paradigm for model alignment, yet its reliance on teacher logits restricts its application to white-box scenarios. We contend that structured semantic rubrics can serve as a scalable alternative to teacher logits, enabling OPD using only teacher-generated responses. To prove it, we introduce ROPD, a simple yet foundational framework for rubric-based OPD. Specifically, ROPD induces prompt-specific rubrics from teacher-student contrasts, and then utilizes these rubrics to score the student rollouts for on-policy optimization. Empirically, ROPD outperforms the advanced logit-based OPD methods across most scenarios, and achieving up to a 10x gain in sample efficiency. These results position rubric-based OPD as a flexible, black-box-compatible alternative to the prevailing logit-based OPD, offering a simple yet strong baseline for scalable distillation across proprietary and open-source LLMs. Code is available at https://github.com/Peregrine123/ROPD_official.

Houcheng Jiang, Jiajun Fu, Junfeng Fang et al.

Multimodal large language models are increasingly expected to perform thinking with images, yet existing visual latent reasoning methods still rely on explicit textual chain-of-thought interleaved with visual latent tokens. This interleaved design limits efficiency and keeps reasoning fragmented across separate text and vision channels. We propose UniVLR, a unified visual latent reasoning framework that treats textual reasoning and auxiliary visual evidence as a shared visual workspace. Instead of preserving text CoT as an independent inference-time path, UniVLR renders reasoning traces together with auxiliary images and learns to compress this unified representation into compact visual latent tokens. At inference time, the model reasons only through visual latents and directly decodes the final answer, avoiding both external tool calls and verbose text reasoning. Experiments on real-world perception and visual reasoning tasks show that UniVLR outperforms prior visual latent reasoning methods while using substantially fewer generated reasoning tokens, suggesting a more unified and efficient paradigm for visual thinking in MLLMs.

Xingrui Qin, Wentao Zhao, Chuan Cao et al.

Dense depth estimation using millimeter-wave radar typically requires dense LiDAR supervision, generated via multi-frame projection and interpolation, for guiding the learning of accurate depth from sparse radar measurements and RGB images. However, this paradigm is both costly and data-intensive. To address this, we propose RaCalNet, a novel framework that eliminates the need for dense supervision by using sparse LiDAR to supervise the learning of refined radar measurements, resulting in a supervision density of merely around 1\% compared to dense-supervised methods. RaCalNet is composed of two key modules. The Radar Recalibration module performs radar point screening and pixel-wise displacement refinement, producing accurate and reliable depth priors from sparse radar inputs. These priors are then used by the Metric Depth Optimization module, which learns to infer scene-level scale priors and fuses them with monocular depth predictions to achieve metrically accurate outputs. This modular design enhances structural consistency and preserves fine-grained geometric details. Despite relying solely on sparse supervision, RaCalNet produces depth maps with clear object contours and fine-grained textures, demonstrating superior visual quality compared to state-of-the-art dense-supervised methods. Quantitatively, it achieves performance comparable to existing methods on the ZJU-4DRadarCam dataset and yields a 34.89\% RMSE reduction in real-world deployment scenarios. We plan to gradually release the code and models in the future at https://github.com/818slam/RaCalNet.git.