{"paper":{"title":"Switching Successor Measures for Hierarchical Zero-shot Reinforcement Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Switching successor measures arise naturally from classical ones and let a single forward-backward representation produce both high-level subgoals and low-level actions in zero-shot hierarchical RL.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Alexandre Proutiere, Stefan Stojanovic","submitted_at":"2026-05-13T08:58:33Z","abstract_excerpt":"Hierarchical reinforcement learning can improve generalization by decomposing long-horizon decision-making into simpler subproblems. However, existing approaches often rely on restrictive design choices, such as fixed temporal abstractions or goal-conditioned objectives, which largely confine them to goal-reaching tasks and limit their applicability to general reward functions. In this paper, we introduce switching successor measures, an extension of successor measures that enables hierarchical control in zero-shot reinforcement learning without additional supervision, fixed horizons, or manua"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Switching successor measures arise naturally from classical successor measures while preserving their underlying structure, allowing FB π-Switch to extract both a high-level subgoal-selection policy and a low-level control policy directly from forward-backward representations for hierarchical zero-shot RL without additional supervision, fixed horizons, or manually designed subgoals.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That switching successor measures can be derived from classical ones in a way that preserves structure sufficiently to support emergent hierarchical behavior from a single FB representation across both goal-conditioned and general reward tasks.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Switching successor measures extend classical successor measures to enable hierarchical zero-shot RL via the FB π-Switch algorithm that extracts subgoal-selection and control policies from forward-backward representations.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Switching successor measures arise naturally from classical ones and let a single forward-backward representation produce both high-level subgoals and low-level actions in zero-shot hierarchical RL.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"0bda80f9832fa6543a6f09680df36d3a17fe87b5ec8a6b1857d6edc24bcc995b"},"source":{"id":"2605.13207","kind":"arxiv","version":1},"verdict":{"id":"f4c4fbb0-0e27-49fc-a8e2-0277f5c3967f","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:23:39.992215Z","strongest_claim":"Switching successor measures arise naturally from classical successor measures while preserving their underlying structure, allowing FB π-Switch to extract both a high-level subgoal-selection policy and a low-level control policy directly from forward-backward representations for hierarchical zero-shot RL without additional supervision, fixed horizons, or manually designed subgoals.","one_line_summary":"Switching successor measures extend classical successor measures to enable hierarchical zero-shot RL via the FB π-Switch algorithm that extracts subgoal-selection and control policies from forward-backward representations.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That switching successor measures can be derived from classical ones in a way that preserves structure sufficiently to support emergent hierarchical behavior from a single FB representation across both goal-conditioned and general reward tasks.","pith_extraction_headline":"Switching successor measures arise naturally from classical ones and let a single forward-backward representation produce both high-level subgoals and low-level actions in zero-shot hierarchical RL."},"references":{"count":64,"sample":[{"doi":"","year":2021,"title":"Deep reinforcement learning at the edge of the statistical precipice","work_id":"c3d48de0-debb-4594-8d88-4bbd76ba95ee","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"A unified framework for unsupervised reinforcement learning al- gorithms","work_id":"599e9445-4bfe-4ead-9512-2ab76ede8723","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Proto successor measure: Representing the behavior space of an RL agent.arXiv preprint arXiv:2411.19418, 2024","work_id":"e2101019-a668-4daf-be2d-55e515f58012","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Option-aware temporally abstracted value for offline goal-conditioned reinforcement learning","work_id":"178e6002-05c1-4b61-838e-d8e5477a8d7c","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"OPAL: Offline primitive discovery for accelerating offline reinforcement learning","work_id":"286b3006-aac9-45e5-bec8-8122a192c114","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":64,"snapshot_sha256":"d827bc72f2d7a74a477eb500d5465e590674be9d1ffa73c3a7267c0f8e18c273","internal_anchors":3},"formal_canon":{"evidence_count":2,"snapshot_sha256":"461adf71a93f3a186bf64f09d8638da67f16902012ca9553c18b810495f73e45"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}