Hugging Face Daily Papers · July 1, 2026 · 6 min read

Does VLA Even Know the Basics? Measuring Commonsense and World Knowledge Retention in Vision-Language-Action Models

#model-release #multimodal #agents #hardware

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Like Read original ↗

<video src=\"https://cdn-uploads.huggingface.co/production/uploads/64705ef84be5cf1f3348e283/Kh4-yceGXH5VOe0HnTMt_.mp4\" controls=\"\" class=\"max-w-full!\"></video>\n\n\nArXiv: <a href=\"https://arxiv.org/abs/2606.19297\" rel=\"nofollow\">https://arxiv.org/abs/2606.19297</a> Project Page: <a href=\"https://tttonyalpha.github.io/act2answer\" rel=\"nofollow\">https://tttonyalpha.github.io/act2answer</a> Code: <a href=\"https://github.com/CognitiveAISystems/Act2Answer\" rel=\"nofollow\">https://github.com/CognitiveAISystems/Act2Answer</a>\n","updatedAt":"2026-07-01T15:45:23.856Z","author":{"_id":"64705ef84be5cf1f3348e283","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64705ef84be5cf1f3348e283/mE8jsFgEk7AN_8KiH8NNW.jpeg","fullname":"Nikita","name":"tttonyalpha","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":3,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.5963894724845886},"editors":["tttonyalpha"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/64705ef84be5cf1f3348e283/mE8jsFgEk7AN_8KiH8NNW.jpeg"],"reactions":[{"reaction":"🔥","users":["ANDRYHA"],"count":1}],"isReport":false}},{"id":"6a45c3587e3d33cf0af16074","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":372,"isUserFollowing":false},"createdAt":"2026-07-02T01:48:08.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"This is an automated message from the [Librarian Bot](https://huggingface.co/librarian-bots). I found the following papers similar to this paper. \n\nThe following papers were recommended by the Semantic Scholar API \n\n* [RoboSemanticBench: Diagnosing Semantic Grounding in Action Prediction for VLA Models](https://huggingface.co/papers/2606.02277) (2026)\n* [VLA-Trace: Diagnosing Vision-Language-Action Models through Representation and Behavior Tracing](https://huggingface.co/papers/2605.30117) (2026)\n* [Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments](https://huggingface.co/papers/2605.30280) (2026)\n* [Rethinking VLM Representation for VLA Initialization](https://huggingface.co/papers/2605.25802) (2026)\n* [When Does Language Matter? Multilingual Instructions Reveal Step-wise Language Sensitivity in Vision-Language-Action Models](https://huggingface.co/papers/2606.11906) (2026)\n* [GuidedVLA: Specifying Task-Relevant Factors via Plug-and-Play Action Attention Specialization](https://huggingface.co/papers/2605.12369) (2026)\n* [LA4VLA: Learning to Act without Seeing via Language-Action Pretraining](https://huggingface.co/papers/2606.27295) (2026)\n\n\n Please give a thumbs up to this comment if you found it helpful!\n\n If you want recommendations for any Paper on Hugging Face checkout [this](https://huggingface.co/spaces/librarian-bots/recommend_similar_papers) Space\n\n You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: `@librarian-bot recommend`","html":"This is an automated message from the <a href=\"https://huggingface.co/librarian-bots\">Librarian Bot</a>. I found the following papers similar to this paper. \nThe following papers were recommended by the Semantic Scholar API \n<ul>\n<li><a href=\"https://huggingface.co/papers/2606.02277\">RoboSemanticBench: Diagnosing Semantic Grounding in Action Prediction for VLA Models</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2605.30117\">VLA-Trace: Diagnosing Vision-Language-Action Models through Representation and Behavior Tracing</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2605.30280\">Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2605.25802\">Rethinking VLM Representation for VLA Initialization</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2606.11906\">When Does Language Matter? Multilingual Instructions Reveal Step-wise Language Sensitivity in Vision-Language-Action Models</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2605.12369\">GuidedVLA: Specifying Task-Relevant Factors via Plug-and-Play Action Attention Specialization</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2606.27295\">LA4VLA: Learning to Act without Seeing via Language-Action Pretraining</a> (2026)</li>\n</ul>\n Please give a thumbs up to this comment if you found it helpful!\n If you want recommendations for any Paper on Hugging Face checkout <a href=\"https://huggingface.co/spaces/librarian-bots/recommend_similar_papers\">this</a> Space\n You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: <code>@librarian-bot recommend</code>\n","updatedAt":"2026-07-02T01:48:08.759Z","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":372,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.6915649771690369},"editors":["librarian-bot"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2606.19297","authors":[{"_id":"6a3a826efdcd3514343bb88c","name":"Nikita Kachaev","hidden":false},{"_id":"6a3a826efdcd3514343bb88d","name":"Andrey Moskalenko","hidden":false},{"_id":"6a3a826efdcd3514343bb88e","name":"Matvey Skripkin","hidden":false},{"_id":"6a3a826efdcd3514343bb88f","name":"Nikita Kurlaev","hidden":false},{"_id":"6a3a826efdcd3514343bb890","name":"Daria Pugacheva","hidden":false},{"_id":"6a3a826efdcd3514343bb891","name":"Albina Burlova","hidden":false},{"_id":"6a3a826efdcd3514343bb892","name":"Mikhail Kolosov","hidden":false},{"_id":"6a3a826efdcd3514343bb893","name":"Denis Shepelev","hidden":false},{"_id":"6a3a826efdcd3514343bb894","name":"Andrey Kuznetsov","hidden":false},{"_id":"6a3a826efdcd3514343bb895","name":"Elena Tutubalina","hidden":false},{"_id":"6a3a826efdcd3514343bb896","name":"Aleksandr I. Panov","hidden":false},{"_id":"6a3a826efdcd3514343bb897","name":"Alexey K. Kovalev","hidden":false},{"_id":"6a3a826efdcd3514343bb898","name":"Vlad Shakhuro","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/64705ef84be5cf1f3348e283/uWJrMxWyi2E5TV-t7VXpn.mp4"],"publishedAt":"2026-06-17T00:00:00.000Z","submittedOnDailyAt":"2026-07-01T00:00:00.000Z","title":"Does VLA Even Know the Basics? Measuring Commonsense and World Knowledge Retention in Vision-Language-Action Models","submittedOnDailyBy":{"_id":"64705ef84be5cf1f3348e283","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64705ef84be5cf1f3348e283/mE8jsFgEk7AN_8KiH8NNW.jpeg","isPro":false,"fullname":"Nikita","user":"tttonyalpha","type":"user","name":"tttonyalpha"},"summary":"Embodied Vision-Language-Action (VLA) models are typically obtained by fine-tuning powerful pretrained VLMs on robotics data, yet it is unclear how much commonsense and factual knowledge they retain after adaptation. Failures on knowledge-sensitive tasks are ambiguous, conflating missing knowledge with poor generalization of low-level control. We introduce Act2Answer, a lightweight protocol that adapts VLM knowledge benchmarks to VLA evaluation by requiring agents to answer through action. Each question becomes a short tabletop episode where the agent performs a single object-placement action to select among candidate answers, yielding an action-grounded success rate with reduced control confounds. We curate a test suite of such environments across diverse commonsense and world-knowledge categories and introduce layerwise intent probing to localize answer-relevant information across the VLM backbone and action head. In a large-scale study of 7 VLA models and 9 VLM baselines, we systematically rank models across categories, finding that VLAs show solid performance on simple concepts while exhibiting larger gaps on richer semantic categories relative to their source VLMs, that VQA co-training is associated with better knowledge retention, and that answer-relevant signals peak in middle VLA layers but attenuate in upper layers. Act2Answer is available at https://tttonyalpha.github.io/act2answer/.","upvotes":41,"discussionId":"6a3a826efdcd3514343bb899","projectPage":"https://tttonyalpha.github.io/act2answer/","githubRepo":"https://github.com/CognitiveAISystems/Act2Answer","githubRepoAddedBy":"user","ai_summary":"Act2Answer protocol evaluates embodied vision-language-action models by having agents answer questions through physical actions, revealing knowledge retention and generalization patterns across different semantic categories.","ai_keywords":["Vision-Language-Action models","pretrained VLMs","robotics data","knowledge-sensitive tasks","action-grounded success rate","commonsense knowledge","world-knowledge categories","layerwise intent probing","VQA co-training","semantic categories"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":2},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"623a3b20fa4890c51b04cba7","avatarUrl":"/avatars/ebc71284fab04d506471aa5b1792157b.svg","isPro":false,"fullname":"Andrey Moskalenko","user":"ANDRYHA","type":"user"},{"_id":"64705ef84be5cf1f3348e283","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64705ef84be5cf1f3348e283/mE8jsFgEk7AN_8KiH8NNW.jpeg","isPro":false,"fullname":"Nikita","user":"tttonyalpha","type":"user"},{"_id":"6626c5d0a329de26e7eb16fa","avatarUrl":"/avatars/124f389f768fb666efd8b5a9b54c3b3c.svg","isPro":false,"fullname":"Matvey Skripkin","user":"barracuda049","type":"user"},{"_id":"6169a581d05945bfd8718dfa","avatarUrl":"/avatars/1892ab06a7ddb557232777de3cbec470.svg","isPro":false,"fullname":"Ivan Oseledets","user":"oseledets","type":"user"},{"_id":"6453df2b908e259483c12bdb","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6453df2b908e259483c12bdb/j0QZDDWffeZNiTMEgQTtS.jpeg","isPro":false,"fullname":"Oleg Y. Rogov","user":"qubitter","type":"user"},{"_id":"6668687caee0993c95b0eb81","avatarUrl":"/avatars/301fe1f395e0a129b1c9785868fa9858.svg","isPro":false,"fullname":"Egor Cherepanov","user":"avanturist","type":"user"},{"_id":"665b10fb270e47e678f2ddf1","avatarUrl":"/avatars/1bc7a9211acf767f7bfca998c24315a0.svg","isPro":false,"fullname":"max","user":"maksimko123","type":"user"},{"_id":"6746007fec397570749d9f8e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6746007fec397570749d9f8e/EV3W2emcC1Yb2JQAnmWMs.jpeg","isPro":false,"fullname":"Mikhail Kolosov","user":"idkolosofff","type":"user"},{"_id":"698a047b84704fee74653012","avatarUrl":"/avatars/b04efeb86c08ea64221c8da64a04858d.svg","isPro":false,"fullname":"I M","user":"iamakarov","type":"user"},{"_id":"6392017db52ed222df025ef4","avatarUrl":"/avatars/a12bc218bcffd370b510a3538627e6a7.svg","isPro":false,"fullname":"Denis Shepelev","user":"denzist","type":"user"},{"_id":"67bcb1012906865678a11f91","avatarUrl":"/avatars/80fb0cc24f0d16c4740f9115b680df0f.svg","isPro":false,"fullname":"Vladimir Korviakov","user":"korviakov","type":"user"},{"_id":"663692c75f67f8da32723bf8","avatarUrl":"/avatars/258264afe2ea5048a4a7a8e9945d2f5b.svg","isPro":false,"fullname":"Alexander Yakovenko","user":"a-yakovenko","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2606/2606.19297.md","query":{}}">

Papers

arxiv:2606.19297

Does VLA Even Know the Basics? Measuring Commonsense and World Knowledge Retention in Vision-Language-Action Models

Published on Jun 17

· Submitted by

Nikita on Jul 1

Upvote

Authors:

Abstract

Act2Answer protocol evaluates embodied vision-language-action models by having agents answer questions through physical actions, revealing knowledge retention and generalization patterns across different semantic categories.

Generated by Qwen/Qwen2.5-Coder-32B-Instruct

Embodied Vision-Language-Action (VLA) models are typically obtained by fine-tuning powerful pretrained VLMs on robotics data, yet it is unclear how much commonsense and factual knowledge they retain after adaptation. Failures on knowledge-sensitive tasks are ambiguous, conflating missing knowledge with poor generalization of low-level control. We introduce Act2Answer, a lightweight protocol that adapts VLM knowledge benchmarks to VLA evaluation by requiring agents to answer through action. Each question becomes a short tabletop episode where the agent performs a single object-placement action to select among candidate answers, yielding an action-grounded success rate with reduced control confounds. We curate a test suite of such environments across diverse commonsense and world-knowledge categories and introduce layerwise intent probing to localize answer-relevant information across the VLM backbone and action head. In a large-scale study of 7 VLA models and 9 VLM baselines, we systematically rank models across categories, finding that VLAs show solid performance on simple concepts while exhibiting larger gaps on richer semantic categories relative to their source VLMs, that VQA co-training is associated with better knowledge retention, and that answer-relevant signals peak in middle VLA layers but attenuate in upper layers. Act2Answer is available at https://tttonyalpha.github.io/act2answer/.