{"paper":{"title":"Learning POMDP World Models from Observations with Language-Model Priors","license":"http://creativecommons.org/licenses/by/4.0/","headline":"An LLM proposes and refines POMDP models from observation-action trajectories alone to match methods with hidden-state access.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Alfonso Amayuelas, Bernhard Sch\\\"olkopf, David Hyland, Frederik Panse, Lancelot Da Costa, Mathis Fajeau, Mridul Sharma, Philipp Hennig, Tim Z. Xiao, Valentin Six","submitted_at":"2026-05-13T16:18:15Z","abstract_excerpt":"Whether navigating a building, operating a robot, or playing a game, an agent that acts effectively in an environment must first learn an internal model of how that environment works. Partially-observable Markov decision processes (POMDPs) provide a flexible modeling class for such internal world models, but learning them from observation-action trajectories alone is challenging and typically requires extensive environment interaction. We ask whether language-model priors can reduce costly interaction by leveraging prior knowledge, and introduce \\emph{Pinductor} (POMDP-inductor): an LLM propos"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Despite using strictly less information, Pinductor matches the performance and sample efficiency of LLM-based POMDP learning methods that assume privileged access to the hidden state, while significantly surpassing the sample efficiency of tabular POMDP baselines.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That an LLM can reliably propose and iteratively refine POMDP transition and observation models whose belief-based likelihood on limited trajectories corresponds to the true underlying dynamics.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Pinductor leverages language-model priors to learn POMDP world models from limited trajectories, matching privileged-access methods in performance and exceeding tabular baselines in sample efficiency.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"An LLM proposes and refines POMDP models from observation-action trajectories alone to match methods with hidden-state access.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f7d2bf228ed43e4d54dc6b648eabacb78819e9d1a405ffa4cbcc52c5761e1263"},"source":{"id":"2605.13740","kind":"arxiv","version":1},"verdict":{"id":"c747305a-4889-429f-bd26-553d0d25ac2b","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:51:44.251932Z","strongest_claim":"Despite using strictly less information, Pinductor matches the performance and sample efficiency of LLM-based POMDP learning methods that assume privileged access to the hidden state, while significantly surpassing the sample efficiency of tabular POMDP baselines.","one_line_summary":"Pinductor leverages language-model priors to learn POMDP world models from limited trajectories, matching privileged-access methods in performance and exceeding tabular baselines in sample efficiency.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That an LLM can reliably propose and iteratively refine POMDP transition and observation models whose belief-based likelihood on limited trajectories corresponds to the true underlying dynamics.","pith_extraction_headline":"An LLM proposes and refines POMDP models from observation-action trajectories alone to match methods with hidden-state access."},"references":{"count":60,"sample":[{"doi":"","year":2018,"title":"Sutton and Andrew G","work_id":"5b66f797-bcf6-443f-a9c8-7e12732ddd11","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"World Models","work_id":"07227eee-8445-4c98-bce4-c6a6fd5ed907","ref_index":2,"cited_arxiv_id":"1803.10122","is_internal_anchor":true},{"doi":"","year":2025,"title":"Training Agents Inside of Scalable World Models","work_id":"f0464a07-aaee-486f-a0b1-a4bce0bbc3e4","ref_index":3,"cited_arxiv_id":"2509.24527","is_internal_anchor":true},{"doi":"10.1016/0022-247x(65)90154-x","year":1965,"title":"Aggregate: count rows whereCOUNTRY= Algeria. [target: Country]","work_id":"c1830d65-2327-4588-a847-abaa0f7884b2","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1016/s0004-3702(98)00023-x","year":1998,"title":"Planning and acting in partially observable stochastic domains","work_id":"b3ea0098-466e-4f58-abfe-1641bbce18de","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":60,"snapshot_sha256":"ba4aafca1db6ebe3431c0447bb9f8836a30c5547a0fdd0ec374f793813989a1c","internal_anchors":4},"formal_canon":{"evidence_count":2,"snapshot_sha256":"b20fe43997b716a34595e08c4412383ecde1f2be9d36ccb4b8dcd8d8c709d629"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}