# Qwen3-VL

Knowledge distillation for driving vision-language models previously matched teacher distributions uniformly across answer tokens, diluting planning-relevant semantics and passing on confident teacher errors. RSKD first localizes task-critical phrases in the ground-truth answer, such as traffic participants, states and spatial relations, then uses a frozen sentence encoder to score how semantically consistent the teacher's prediction is with each phrase, so consistent positions take more teacher-distribution supervision and inconsistent ones take more ground-truth supervision. On DrivingVQA, LingoQA and MapLMv2, RSKD reaches 16.60, 0.672 and 81.82, beating the compared SFT, GKD and ToDi student baselines on the two QA benchmarks and nearly matching SFT's 81.85 on MapLMv2.

Knowledge distillation for driving vision-language models previously matched teacher distributions uniformly across answer tokens, diluting planning-relevant semantics and passing on confident teacher errors. RSKD first localizes task-critical phrases in the ground-truth answer, such as traffic participants, states and spatial relations, then uses a frozen sentence encoder to score how semantically consistent the teacher's prediction is with each phrase, so consistent positions take more teacher-distribution supervision and inconsistent ones take more ground-truth supervision. On DrivingVQA, LingoQA and MapLMv2, RSKD reaches 16.60, 0.672 and 81.82, beating the compared SFT, GKD and ToDi student baselines on the two QA benchmarks and nearly matching SFT's 81.85 on MapLMv2.

The internal choices inside a visual-head discovery score reshape which heads it returns in vision-language models. On Qwen3-VL-8B-Instruct, fully crossing query type, discovery corpus and the token positions whose attention is read, and testing all 12 head-sets with steerability and necessity interventions, only necessity follows the attention source: query-span head-sets are more necessary than last-token ones, and the two referring plus query-span sets lie 17–19 standard deviations above 250 layer-matched random sets. At the default strength, no discovered head-set steers better than its layer-matched random sets.

The internal choices inside a visual-head discovery score reshape which heads it returns in vision-language models. On Qwen3-VL-8B-Instruct, fully crossing query type, discovery corpus and the token positions whose attention is read, and testing all 12 head-sets with steerability and necessity interventions, only necessity follows the attention source: query-span head-sets are more necessary than last-token ones, and the two referring plus query-span sets lie 17–19 standard deviations above 250 layer-matched random sets. At the default strength, no discovered head-set steers better than its layer-matched random sets.

Vision-language models struggle with spatial reasoning because RGB inputs do not directly supply geometric evidence, and existing fixes either add 3D at inference, paying architecture and latency costs, or train on outcome rewards that supervise only the final answer. GPD renders depth, semantic segmentation and bird's-eye-view layout as compact text, routes the question-relevant subset to the teacher alongside the reference answer, and applies a privileged KL only to incorrect trajectories within GRPO, leaving the deployed model RGB-only. On the 4B backbone it reaches 57.1 on VSI-Bench and 37.6 average across MindCube, SPAR-Bench, MMSI-Bench and ViewSpatial, above GRPO and answer-privileged OPSD.

Vision-language models struggle with spatial reasoning because RGB inputs do not directly supply geometric evidence, and existing fixes either add 3D at inference, paying architecture and latency costs, or train on outcome rewards that supervise only the final answer. GPD renders depth, semantic segmentation and bird's-eye-view layout as compact text, routes the question-relevant subset to the teacher alongside the reference answer, and applies a privileged KL only to incorrect trajectories within GRPO, leaving the deployed model RGB-only. On the 4B backbone it reaches 57.1 on VSI-Bench and 37.6 average across MindCube, SPAR-Bench, MMSI-Bench and ViewSpatial, above GRPO and answer-privileged OPSD.

The authors introduce the AutoRegressive Retriever (ARR), a multimodal retrieval model that alternates between retrieving an item and updating the query embedding, appending each selected item's multimodal content to the query history before re-encoding and using the final embedding to rank the collection; training has two stages, supervised fine-tuning with stepwise contrastive supervision to teach feedback interpretation, then reinforcement learning that treats feedback items as actions and optimizes their selection using the final reciprocal rank of a relevant item, with a query-side adapter enabling this against a fixed item index.

The authors introduce the AutoRegressive Retriever (ARR), a multimodal retrieval model that alternates between retrieving an item and updating the query embedding, appending each selected item's multimodal content to the query history before re-encoding and using the final embedding to rank the collection; training has two stages, supervised fine-tuning with stepwise contrastive supervision to teach feedback interpretation, then reinforcement learning that treats feedback items as actions and optimizes their selection using the final reciprocal rank of a relevant item, with a query-side adapter enabling this against a fixed item index.

For the Explainable Deepfake Detection Challenge on XPlainVerse, the authors present a modular detection-and-explanation system in which a detector fuses several DINOv3 backbones with Mesorch forensic features, a Grounding-DINO pipeline converts local artifact descriptions from explanations into weak patch-level pseudo-masks supervising an Artifact Evidence Map, and a local patch-level contrastive objective separates artifact from authenticity evidence, while class-conditional Qwen3-VL models generate complex explanations and a GRPO-optimized model simplifies them, reaching 0.9349 detection accuracy, a 0.5571 explanation score, and a 0.7456 final challenge score on the full test split.

For the Explainable Deepfake Detection Challenge on XPlainVerse, the authors present a modular detection-and-explanation system in which a detector fuses several DINOv3 backbones with Mesorch forensic features, a Grounding-DINO pipeline converts local artifact descriptions from explanations into weak patch-level pseudo-masks supervising an Artifact Evidence Map, and a local patch-level contrastive objective separates artifact from authenticity evidence, while class-conditional Qwen3-VL models generate complex explanations and a GRPO-optimized model simplifies them, reaching 0.9349 detection accuracy, a 0.5571 explanation score, and a 0.7456 final challenge score on the full test split.