{"paper":{"title":"Thinking with Patterns: Breaking the Perceptual Bottleneck in Visual Planning via Pattern Induction","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Vision-language models overcome perceptual limits in visual planning by inducing reusable patterns that build accurate internal world models step by step.","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.CV","authors_text":"Boyuan Xiao, Yao-Xiang Ding, Yichang Jian, Yifei Peng, Zhenyuan Huang","submitted_at":"2026-05-16T07:12:19Z","abstract_excerpt":"Planning from raw visual input remains a significant challenge for current Vision-Language Models (VLMs), when the complexity of input is beyond their one-step perception capability. Motivated by recent advances in Thinking with Images (TWI), a reasonable solution is to decompose the perception process into simpler steps by iteratively acquiring and incorporating local visual evidence. However, even though current VLMs are well-trained in general TWI ability, their perceptual bottleneck in the planning domain remains. To tackle this challenge, we formulate TWI as a tool to gradually build and "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"The resulting training-free planning strategy enables VLMs to solve tasks that are far beyond their initial capabilities, at the cost that too many TWI operations would significantly increase the computational overhead; Pattern Inference and Pattern Induction achieve a desirable balance between accuracy and efficiency.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That visual patterns can be treated as composite and reusable experts which are autonomously discovered and optimized from experience in a way that directly improves inference efficiency without requiring task-specific retraining or external supervision.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Pattern Induction discovers reusable visual patterns as experts via online inductive learning, and Pattern Inference uses them to let VLMs perform efficient multi-step visual planning beyond their native capabilities.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Vision-language models overcome perceptual limits in visual planning by inducing reusable patterns that build accurate internal world models step by step.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"c294cb134a2b2f1b7c9daf37b67e13956d89b52219959b0ba0c91f408cbae945"},"source":{"id":"2605.16848","kind":"arxiv","version":1},"verdict":{"id":"fd5e472c-fb94-465a-86b2-af22e36aba69","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T20:40:10.681899Z","strongest_claim":"The resulting training-free planning strategy enables VLMs to solve tasks that are far beyond their initial capabilities, at the cost that too many TWI operations would significantly increase the computational overhead; Pattern Inference and Pattern Induction achieve a desirable balance between accuracy and efficiency.","one_line_summary":"Pattern Induction discovers reusable visual patterns as experts via online inductive learning, and Pattern Inference uses them to let VLMs perform efficient multi-step visual planning beyond their native capabilities.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That visual patterns can be treated as composite and reusable experts which are autonomously discovered and optimized from experience in a way that directly improves inference efficiency without requiring task-specific retraining or external supervision.","pith_extraction_headline":"Vision-language models overcome perceptual limits in visual planning by inducing reusable patterns that build accurate internal world models step by step."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.16848/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T21:01:19.242682Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T20:51:12.384217Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T18:41:56.316321Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T18:33:26.389385Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"d621e9ba4b9f3cb28e219a88d2ad7a62409507638f2a985aa6798536079fa4b3"},"references":{"count":25,"sample":[{"doi":"","year":null,"title":"On top of this, we plan the shortest path","work_id":"7f6321af-c4fa-4e76-b14f-74dd49c88c6c","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"If there aren’t any unrevealed grids on the path, the algorithm stops and returns the path","work_id":"0ec10c1a-07df-4786-b1f5-be17982b985f","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"If there are, we check the unrevealed grids one by one from start to goal","work_id":"fc6f6b03-0f8e-414b-bf3d-40beb4cd5be9","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"If an impassable grid is ever encountered during this process, we immediately go back to step 1 to get another plan","work_id":"459f3547-8a23-4c23-86dc-ebcad48fa377","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"If all checked grids are passable, the algorithm stops and returns the path. Thepolicy-generation procedureworks at step (4) above. Each time, it outputs the first unrevealed grid from start to goal. ","work_id":"a2a15d64-511f-4c51-bf1d-8573017dae15","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":25,"snapshot_sha256":"ce303876fad08763714cf31a1592dcc6be00e44fecc6091ee1c5c289dd28863b","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"9186336b8dbf870eaf5554ad0dea06c701ed5784cc4e089032d5cf3101e448d3"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}