{"paper":{"title":"MAP: A Map-then-Act Paradigm for Long-Horizon Interactive Agent Reasoning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"The Map-then-Act Paradigm lets LLM agents build environment maps before execution to escape trial-and-error cycles.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Jinwei Xiao, Lei Zhang, Mingye Zhu, Qi Gu, Xunliang Cai, Yueqing Sun, Yuxin Liu, Zhuowen Han, Ziang Ye","submitted_at":"2026-05-13T05:46:29Z","abstract_excerpt":"Current interactive LLM agents rely on goal-conditioned stepwise planning, where environmental understanding is acquired reactively during execution rather than established beforehand. This temporal inversion leads to Delayed Environmental Perception: agents must infer environmental constraints through trial-and-error, resulting in an Epistemic Bottleneck that traps them in inefficient failure cycles. Inspired by human affordance perception and cognitive map theory, we propose the Map-then-Act Paradigm (MAP), a plug-and-play framework that shifts environment understanding before execution. MAP"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On ARC-AGI-3, MAP enables frontier models to surpass near-zero baseline performance in 22 of 25 game environments. We further introduce MAP-2K, a dataset of map-then-act trajectories, and show that training on it outperforms expert execution traces.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That global exploration can efficiently acquire accurate environment-general priors and that the resulting structured cognitive map will remain valid and useful during subsequent task execution without introducing new errors or excessive overhead.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"MAP improves LLM agent reasoning by constructing a structured cognitive map of the environment before task execution, yielding performance gains on benchmarks like ARC-AGI-3 and superior training data via the new MAP-2K dataset.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"The Map-then-Act Paradigm lets LLM agents build environment maps before execution to escape trial-and-error cycles.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"c0a064e91ec1dd89b35d3a987f4549b8818931849f01bd388a14f31a2e076ebe"},"source":{"id":"2605.13037","kind":"arxiv","version":1},"verdict":{"id":"42b10749-d2a2-45a5-9eca-5de9996b0616","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:43:04.508085Z","strongest_claim":"On ARC-AGI-3, MAP enables frontier models to surpass near-zero baseline performance in 22 of 25 game environments. We further introduce MAP-2K, a dataset of map-then-act trajectories, and show that training on it outperforms expert execution traces.","one_line_summary":"MAP improves LLM agent reasoning by constructing a structured cognitive map of the environment before task execution, yielding performance gains on benchmarks like ARC-AGI-3 and superior training data via the new MAP-2K dataset.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That global exploration can efficiently acquire accurate environment-general priors and that the resulting structured cognitive map will remain valid and useful during subsequent task execution without introducing new errors or excessive overhead.","pith_extraction_headline":"The Map-then-Act Paradigm lets LLM agents build environment maps before execution to escape trial-and-error cycles."},"references":{"count":50,"sample":[{"doi":"","year":2023,"title":"GPT-4 Technical Report","work_id":"b928e041-6991-4c08-8c81-0359e4097c7b","ref_index":1,"cited_arxiv_id":"2303.08774","is_internal_anchor":true},{"doi":"","year":2023,"title":"Self-rag: Learn- ing to retrieve, generate, and critique through self-reflection","work_id":"2070772a-8584-4ab0-99be-17b284fd4927","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Remember Me, Refine Me: A Dynamic Procedural Memory Framework for Experience-Driven Agent Evolution","work_id":"e1808e0d-921b-46dd-b221-96110080aa71","ref_index":3,"cited_arxiv_id":"2512.10696","is_internal_anchor":true},{"doi":"","year":2025,"title":"MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention","work_id":"c59fbe20-f41e-4140-a81c-40a12e7e8364","ref_index":4,"cited_arxiv_id":"2506.13585","is_internal_anchor":true},{"doi":"","year":2026,"title":"Learning to self-verify makes language models better reasoners.CoRR, abs/2602.07594","work_id":"ff39b989-0a7a-491a-88b0-1a47366715ff","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":50,"snapshot_sha256":"4a9b7dfdccd1e989f5d84398ff435daf14df7eb99430c71f227a518296cc6d80","internal_anchors":18},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}