{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:7C53S2RCYVBEV4RJMXKOND544V","short_pith_number":"pith:7C53S2RC","schema_version":"1.0","canonical_sha256":"f8bbb96a22c5424af22965d4e68fbce568cf9d1bfcea6f2520552a063ef2a088","source":{"kind":"arxiv","id":"2607.21273","version":1},"attestation_state":"computed","paper":{"title":"The Dark Room in the Reward Channel: Dense Prediction Rewards Collapse GRPO-Trained LLM Agents -- and What Actually Works","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Yu Wang","submitted_at":"2026-07-23T12:50:18Z","abstract_excerpt":"Dense per-step supervision is an appealing remedy for sparse-reward, long-horizon LLM agents: reward the agent for predicting its next observation, and memory should follow. We show that under group-normalized RL (GRPO), this recipe does not merely fail -- it destroys the policy. Across Qwen3-1.7B/4B/8B on ALFWorld, a potential-based prediction reward drives every run into a degenerate absorbing state (prediction accuracy -> 1.0, task success -> 0,episode length pinned at the horizon): the \"dark room\" pathology, built automatically by the optimizer. A single-factor ablation localizes the cause"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2607.21273","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-07-23T12:50:18Z","cross_cats_sorted":[],"title_canon_sha256":"adcf42aa0868c43d10cda759c147949cf02473513533091f7ab3059533fae88e","abstract_canon_sha256":"768f9ab4abfd6b010062b86b2627885d16de488cf9ed69fe6f56eba1dc5361fa"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-24T01:24:23.133052Z","signature_b64":"7HUCiJ4AotF9gI47j6ATEAdG8XlGxuwRLGh+TWaNgS/j2IfsR8eWW7ZK+Oif4r1v6hZt/EI5FPS6br2G45AoAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f8bbb96a22c5424af22965d4e68fbce568cf9d1bfcea6f2520552a063ef2a088","last_reissued_at":"2026-07-24T01:24:23.132098Z","signature_status":"signed_v1","first_computed_at":"2026-07-24T01:24:23.132098Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"The Dark Room in the Reward Channel: Dense Prediction Rewards Collapse GRPO-Trained LLM Agents -- and What Actually Works","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Yu Wang","submitted_at":"2026-07-23T12:50:18Z","abstract_excerpt":"Dense per-step supervision is an appealing remedy for sparse-reward, long-horizon LLM agents: reward the agent for predicting its next observation, and memory should follow. We show that under group-normalized RL (GRPO), this recipe does not merely fail -- it destroys the policy. Across Qwen3-1.7B/4B/8B on ALFWorld, a potential-based prediction reward drives every run into a degenerate absorbing state (prediction accuracy -> 1.0, task success -> 0,episode length pinned at the horizon): the \"dark room\" pathology, built automatically by the optimizer. A single-factor ablation localizes the cause"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2607.21273","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2607.21273/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2607.21273","created_at":"2026-07-24T01:24:23.132556+00:00"},{"alias_kind":"arxiv_version","alias_value":"2607.21273v1","created_at":"2026-07-24T01:24:23.132556+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2607.21273","created_at":"2026-07-24T01:24:23.132556+00:00"},{"alias_kind":"pith_short_12","alias_value":"7C53S2RCYVBE","created_at":"2026-07-24T01:24:23.132556+00:00"},{"alias_kind":"pith_short_16","alias_value":"7C53S2RCYVBEV4RJ","created_at":"2026-07-24T01:24:23.132556+00:00"},{"alias_kind":"pith_short_8","alias_value":"7C53S2RC","created_at":"2026-07-24T01:24:23.132556+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/7C53S2RCYVBEV4RJMXKOND544V","json":"https://pith.science/pith/7C53S2RCYVBEV4RJMXKOND544V.json","graph_json":"https://pith.science/api/pith-number/7C53S2RCYVBEV4RJMXKOND544V/graph.json","events_json":"https://pith.science/api/pith-number/7C53S2RCYVBEV4RJMXKOND544V/events.json","paper":"https://pith.science/paper/7C53S2RC"},"agent_actions":{"view_html":"https://pith.science/pith/7C53S2RCYVBEV4RJMXKOND544V","download_json":"https://pith.science/pith/7C53S2RCYVBEV4RJMXKOND544V.json","view_paper":"https://pith.science/paper/7C53S2RC","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2607.21273&json=true","fetch_graph":"https://pith.science/api/pith-number/7C53S2RCYVBEV4RJMXKOND544V/graph.json","fetch_events":"https://pith.science/api/pith-number/7C53S2RCYVBEV4RJMXKOND544V/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/7C53S2RCYVBEV4RJMXKOND544V/action/timestamp_anchor","attest_storage":"https://pith.science/pith/7C53S2RCYVBEV4RJMXKOND544V/action/storage_attestation","attest_author":"https://pith.science/pith/7C53S2RCYVBEV4RJMXKOND544V/action/author_attestation","sign_citation":"https://pith.science/pith/7C53S2RCYVBEV4RJMXKOND544V/action/citation_signature","submit_replication":"https://pith.science/pith/7C53S2RCYVBEV4RJMXKOND544V/action/replication_record"}},"created_at":"2026-07-24T01:24:23.132556+00:00","updated_at":"2026-07-24T01:24:23.132556+00:00"}