Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2609.38777

Distill the Visual Evidence, Not Just the Answer: Cross-World On-Policy Distillation for Vision-Language Models

Yuanhao Sun·Huawei Ji·Jiaxin Ding·Luoyi Fu·Xinbing Wang
View original source ↗Read source document ↗Export metadata ↓

Abstract

A central goal of vision-language model (VLM) distillation is to transfer both the teacher's language capabilities and its visual understanding. However, existing methods primarily supervise the student's output, leaving visual understanding implicit. Our analysis reveals that a student can match the teacher's answer without relying on the same visual evidence, raising the question: how can we ensure the student responds to the visual information that actually determines the answer? To this end, we propose \textbf{Cross-World On-Policy Distillation (CW-OPD)}, which explicitly supervises the student's response to changes in visual evidence. For each example, CW-OPD constructs two visual worlds that share the question and scene context but differ in answer-critical evidence, yielding different answers. We perform on-policy distillation in both worlds and distill the teacher's cross-world belief transition, encouraging the student to match not only \emph{what} the teacher predicts but also \emph{why} its prediction changes with the evidence. A gradient analysis shows that this term is invariant to errors shared by both worlds and supplies a corrective signal invisible to endpoint matching alone. In this way, CW-OPD makes reliance on the relevant visual evidence an explicit distillation target rather than an implicit consequence of output matching. To diagnose whether a model truly grounds its answers in visual evidence, we introduce CWBench, which measures cross-world consistency via Cross-World Pair Accuracy (CWPA). Experiments on Qwen3.5-4B show that CW-OPD outperforms the strongest baseline by \textbf{1.2} points on average, and the 4B student exceeds DeepSeek-V4.1 (552B) by \textbf{22.4} CWPA points on CWBench. Code is released in https://github.com/baokou-fw2/CWAD.

Explore related subjects

cs.CV

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Yuanhao Sun, Huawei Ji, Jiaxin Ding, Luoyi Fu, Xinbing Wang. 2026-10-19. Distill the Visual Evidence, Not Just the Answer: Cross-World On-Policy Distillation for Vision-Language Models. https://arxiv.org/abs/2609.38777

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Taskonomy: Disentangling Task Transfer Learning

Do visual tasks have a relationship, or are they unrelated? For instance, could having surface normals simplify estimating the depth of an image? Intuition answers these questions positively, implying existence of a structure among visual tasks. Knowing this structure has notable values; it is the concept underlying transfer learning and provides a principled way for identifying redundancies across tasks, e.g., to seamlessly reuse supervision among related tasks or solve many tasks in one system without piling up the complexity. We proposes a fully computational approach for modeling the structure of space of visual tasks. This is done via finding (first and higher-order) transfer learning dependencies across a dictionary of twenty six 2D, 2.5D, 3D, and semantic tasks in a latent space. The product is a computational taxonomic map for task transfer learning. We study the consequences of this structure, e.g. nontrivial emerged relationships, and exploit them to reduce the demand for labeled data. For example, we show that the total number of labeled datapoints needed for solving a set of 10 tasks can be reduced by roughly 2/3 (compared to training independently) while keeping the performance nearly the same. We provide a set of tools for computing and probing this taxonomical structure including a solver that users can employ to devise efficient supervision policies for their use cases.

cs.CV↗
arXiv2026

WildPPG: A Real-World PPG Dataset of Long Continuous Recordings

Reflective photoplethysmography (PPG) has become the default sensing technique in wearable devices to monitor cardiac activity via a person's heart rate (HR). However, PPG-based HR estimates can be substantially impacted by factors such as the wearer's activities, sensor placement and resulting motion artifacts, as well as environmental characteristics such as temperature and ambient light. These and other factors can significantly impact and decrease HR prediction reliability. In this paper, we show that state-of-the-art HR estimation methods struggle when processing \emph{representative} data from everyday activities in outdoor environments, likely because they rely on existing datasets that captured controlled conditions. We introduce a novel multimodal dataset and benchmark results for continuous PPG recordings during outdoor activities from 16 participants over 13.5 hours, captured from four wearable sensors, each worn at a different location on the body, totaling 216\,hours. Our recordings include accelerometer, temperature, and altitude data, as well as a synchronized Lead I-based electrocardiogram for ground-truth HR references. Participants completed a round trip from Zurich to Jungfraujoch, a tall mountain in Switzerland over the course of one day. The trip included outdoor and indoor activities such as walking, hiking, stair climbing, eating, drinking, and resting at various temperatures and altitudes (up to 3,571\,m above sea level) as well as using cars, trains, cable cars, and lifts for transport -- all of which impacted participants' physiological dynamics. We also present a novel method that estimates HR values more robustly in such real-world scenarios than existing baselines.

cs.CV↗
arXiv2026

Scene-R1: Video-Grounded Large Language Models for 3D Scene Reasoning with Lightweight Annotations

Large language models are increasingly being applied to 3D scene understanding. However, existing 3D-aware LLMs largely operate as black boxes: they produce bounding boxes or textual responses without exposing the reasoning process, and they typically depend on pre-trained 3D detectors for object proposals. We present Scene-R1, a video-grounded framework for 3D scene reasoning that avoids both pre-trained 3D object detectors and dense point-cloud backbones. The framework combines reinforcement-learning-based reasoning with a two-stage grounding pipeline. In the first stage, Scene-R1 performs temporal grounding by identifying the video segments most relevant to an open-ended query. In the second stage, it grounds the target in image space by predicting a 2D bounding box. This prediction is then refined into a precise 3D localization using high-quality candidates from a zero-shot segmentation module, enabling fine-grained geometric reasoning without detector-based proposals. Scene-R1 can also be adapted to 3D visual question answering, generating free-form answers directly from RGB-D video. Our training pipeline uses lightweight 2D supervision projected from sparse scans, avoiding the computational overhead required to train 3D instance segmentation models. Across multiple datasets, Scene-R1 outperforms existing open-vocabulary baselines while also providing interpretable intermediate reasoning steps. These results suggest that reinforcement-learning-based reasoning over RGB-D video can support accurate 3D scene understanding with task-level supervision. The training, inference, and 3D refinement code is available at https://github.com/819346675/Scene-R1.

cs.CV↗