Can models think visually in space as humans do? Introducing: Imaginative Perception Tokens by UW, OpenAI, Microsoft, and AI2.</p>\n<p>Imaginative Perception improves spatial reasoning in multimodal language models by teaching them to imagine useful visual perspectives as images. These imagined images help the model reason beyond the original view and answer spatial questions more accurately.</p>\n<p>📄 Paper: Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models<br>arXiv: <a href=\"https://arxiv.org/abs/2606.03988\" rel=\"nofollow\">https://arxiv.org/abs/2606.03988</a></p>\n<p>💻 Code<br>• Training: <a href=\"https://github.com/weikaih04/Imaginative-Perception-Token\" rel=\"nofollow\">https://github.com/weikaih04/Imaginative-Perception-Token</a><br>• Evaluation: <a href=\"https://github.com/weikaih04/Imaginative-Perception-Token-Eval\" rel=\"nofollow\">https://github.com/weikaih04/Imaginative-Perception-Token-Eval</a></p>\n<p>🤗 Data & Benchmarks<br>• Datasets (MVC + PET + PT): <a href=\"https://huggingface.co/collections/weikaih/imaginative-perception-token-data\">https://huggingface.co/collections/weikaih/imaginative-perception-token-data</a><br>• Spatial Mental Modeling Benchmark (human-verified): <a href=\"https://huggingface.co/collections/weikaih/spatial-mental-modeling-benchmark\">https://huggingface.co/collections/weikaih/spatial-mental-modeling-benchmark</a></p>\n<p><a href=\"https://cdn-uploads.huggingface.co/production/uploads/65cb0dcc4913057ac82a7a31/F0xfybI7tSgNOpnB_ntaw.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/65cb0dcc4913057ac82a7a31/F0xfybI7tSgNOpnB_ntaw.png\" alt=\"overview (1)\"></a></p>\n","updatedAt":"2026-06-08T18:34:58.308Z","author":{"_id":"65cb0dcc4913057ac82a7a31","avatarUrl":"/avatars/034f39dc2a5af0d42015c013aad44490.svg","fullname":"Weikai Huang","name":"weikaih","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":8,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7195900082588196},"editors":["weikaih"],"editorAvatarUrls":["/avatars/034f39dc2a5af0d42015c013aad44490.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2606.03988","authors":[{"_id":"6a2700e16dde1c5ef75bcd17","name":"Mahtab Bigverdi","hidden":false},{"_id":"6a2700e16dde1c5ef75bcd18","name":"Linjie Li","hidden":false},{"_id":"6a2700e16dde1c5ef75bcd19","name":"Weikai Huang","hidden":false},{"_id":"6a2700e16dde1c5ef75bcd1a","name":"Yiming Liu","hidden":false},{"_id":"6a2700e16dde1c5ef75bcd1b","name":"Jaemin Cho","hidden":false},{"_id":"6a2700e16dde1c5ef75bcd1c","name":"Jieyu Zhang","hidden":false},{"_id":"6a2700e16dde1c5ef75bcd1d","name":"Tuhin Kundu","hidden":false},{"_id":"6a2700e16dde1c5ef75bcd1e","name":"Chris Dangjoo Kim","hidden":false},{"_id":"6a2700e16dde1c5ef75bcd1f","name":"Zelun Luo","hidden":false},{"_id":"6a2700e16dde1c5ef75bcd20","name":"Linda Shapiro","hidden":false},{"_id":"6a2700e16dde1c5ef75bcd21","name":"Ranjay Krishna","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/65cb0dcc4913057ac82a7a31/rtNQ69U31Xzz-1ZjuL2MB.mp4","https://cdn-uploads.huggingface.co/production/uploads/65cb0dcc4913057ac82a7a31/CTFcV392R2sW8SqVACriQ.png","https://cdn-uploads.huggingface.co/production/uploads/65cb0dcc4913057ac82a7a31/yuuK10V1tiL7uC3XHXN8U.png","https://cdn-uploads.huggingface.co/production/uploads/65cb0dcc4913057ac82a7a31/P_QIx49kmRcr7Gqx46nUp.png","https://cdn-uploads.huggingface.co/production/uploads/65cb0dcc4913057ac82a7a31/7dCOkI6SkTtODg-exREOn.png"],"publishedAt":"2026-06-03T00:00:00.000Z","submittedOnDailyAt":"2026-06-08T00:00:00.000Z","title":"Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models","submittedOnDailyBy":{"_id":"65cb0dcc4913057ac82a7a31","avatarUrl":"/avatars/034f39dc2a5af0d42015c013aad44490.svg","isPro":true,"fullname":"Weikai Huang","user":"weikaih","type":"user","name":"weikaih"},"summary":"Vision language models (VLMs) excel at many tasks but still struggle with spatial reasoning when critical information is not directly observable. Many such problems require imaginative perception: inferring what would be seen from an unseen viewpoint, tracing paths through occluded spaces, or integrating partial observations into a coherent spatial representation. We introduce Imaginative Perception Tokens (IPT), intermediate perceptual representations that externalize what a VLM would perceive under alternative spatial configurations while remaining consistent with the observed input.\n To study this capability, we formulate three tasks, Perspective Taking (PET), Path Tracing (PT), and Multiview Counting (MVC), and construct datasets of approximately 20K examples with ground truth imaginations, answers, and evaluation benchmarks. Using the unified VLM BAGEL as the backbone, IPT supervision consistently improves spatial reasoning and often outperforms textual chain of thought training, even without generating images at inference time. On MVC, IPT improves accuracy by 3.4% and achieves competitive performance with strong closed-source models on PT. We further find that combining IPT and label-only supervision yields additional gains, whereas textual chain of thought can substantially degrade performance, suggesting a modality mismatch when spatial computation is forced through language. Overall, IPT provides a principled supervision signal for reasoning about unobserved spatial structure, improving generalization while producing interpretable intermediate representations.","upvotes":1,"discussionId":"6a2700e26dde1c5ef75bcd22","projectPage":"https://mahtabbigverdi.github.io/Imaginative-tokens.github.io/","githubRepo":"https://github.com/weikaih04/Imaginative-Perception-Token","githubRepoAddedBy":"user","ai_summary":"Imaginative Perception Tokens (IPT) enhance vision-language models' spatial reasoning by providing intermediate perceptual representations that externalize what the model would perceive from alternative viewpoints, outperforming traditional text-based reasoning methods.","ai_keywords":["Vision language models","spatial reasoning","imaginative perception","Imaginative Perception Tokens","perspective taking","path tracing","multiview counting","BAGEL","supervision","intermediate representations","generalization"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":2,"organization":{"_id":"6315a1bb86b3db2ac420100e","name":"UW","fullname":"University of Washington","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/61ac8f8a00d01045fca0ad2f/gr5B_WVvbMr4kTox5UkwZ.jpeg"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"65cb0dcc4913057ac82a7a31","avatarUrl":"/avatars/034f39dc2a5af0d42015c013aad44490.svg","isPro":true,"fullname":"Weikai Huang","user":"weikaih","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6315a1bb86b3db2ac420100e","name":"UW","fullname":"University of Washington","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/61ac8f8a00d01045fca0ad2f/gr5B_WVvbMr4kTox5UkwZ.jpeg"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2606/2606.03988.md"}">
Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models
Authors: ,
,
,
,
,
,
,
,
,
,
Abstract
Imaginative Perception Tokens (IPT) enhance vision-language models' spatial reasoning by providing intermediate perceptual representations that externalize what the model would perceive from alternative viewpoints, outperforming traditional text-based reasoning methods.
Vision language models (VLMs) excel at many tasks but still struggle with spatial reasoning when critical information is not directly observable. Many such problems require imaginative perception: inferring what would be seen from an unseen viewpoint, tracing paths through occluded spaces, or integrating partial observations into a coherent spatial representation. We introduce Imaginative Perception Tokens (IPT), intermediate perceptual representations that externalize what a VLM would perceive under alternative spatial configurations while remaining consistent with the observed input.
To study this capability, we formulate three tasks, Perspective Taking (PET), Path Tracing (PT), and Multiview Counting (MVC), and construct datasets of approximately 20K examples with ground truth imaginations, answers, and evaluation benchmarks. Using the unified VLM BAGEL as the backbone, IPT supervision consistently improves spatial reasoning and often outperforms textual chain of thought training, even without generating images at inference time. On MVC, IPT improves accuracy by 3.4% and achieves competitive performance with strong closed-source models on PT. We further find that combining IPT and label-only supervision yields additional gains, whereas textual chain of thought can substantially degrade performance, suggesting a modality mismatch when spatial computation is forced through language. Overall, IPT provides a principled supervision signal for reasoning about unobserved spatial structure, improving generalization while producing interpretable intermediate representations.
Community
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2606.03988 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.