{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:WEUR63GCLWX24NJ42VP6JS2VQ6","short_pith_number":"pith:WEUR63GC","canonical_record":{"source":{"id":"2605.07021","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-07T23:05:50Z","cross_cats_sorted":[],"title_canon_sha256":"1ceb985b886599ff635b05a228404db77f4ff4615b65a9bf992570f272ff7482","abstract_canon_sha256":"30e05c3adb0fc0e6680b12fbc18a3ed184e0baddf91c6fb1a4930a808f781f9e"},"schema_version":"1.0"},"canonical_sha256":"b1291f6cc25dafae353cd55fe4cb5587a5c88c327317f6bd0f55a6acb47d95a6","source":{"kind":"arxiv","id":"2605.07021","version":2},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.07021","created_at":"2026-05-21T01:04:27Z"},{"alias_kind":"arxiv_version","alias_value":"2605.07021v2","created_at":"2026-05-21T01:04:27Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.07021","created_at":"2026-05-21T01:04:27Z"},{"alias_kind":"pith_short_12","alias_value":"WEUR63GCLWX2","created_at":"2026-05-21T01:04:27Z"},{"alias_kind":"pith_short_16","alias_value":"WEUR63GCLWX24NJ4","created_at":"2026-05-21T01:04:27Z"},{"alias_kind":"pith_short_8","alias_value":"WEUR63GC","created_at":"2026-05-21T01:04:27Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:WEUR63GCLWX24NJ42VP6JS2VQ6","target":"record","payload":{"canonical_record":{"source":{"id":"2605.07021","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-07T23:05:50Z","cross_cats_sorted":[],"title_canon_sha256":"1ceb985b886599ff635b05a228404db77f4ff4615b65a9bf992570f272ff7482","abstract_canon_sha256":"30e05c3adb0fc0e6680b12fbc18a3ed184e0baddf91c6fb1a4930a808f781f9e"},"schema_version":"1.0"},"canonical_sha256":"b1291f6cc25dafae353cd55fe4cb5587a5c88c327317f6bd0f55a6acb47d95a6","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-21T01:04:27.042278Z","signature_b64":"Zvqb9VnUJ4H0dEKC8+8zvHQz7q032LAqwBAxcrKju4DydQ1hTKWSHvBXPL1TfGgFczY2A5Vam5ZfxtDSHGmiCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"b1291f6cc25dafae353cd55fe4cb5587a5c88c327317f6bd0f55a6acb47d95a6","last_reissued_at":"2026-05-21T01:04:27.041460Z","signature_status":"signed_v1","first_computed_at":"2026-05-21T01:04:27.041460Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2605.07021","source_version":2,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-21T01:04:27Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"KV+t5wPOcarM4VofkHo5mrl5iYPNDmzT6ajAZiQVMoc3LbU1kC08E2asgbjD14DDuuOeW519G01Uz9s1FoZ3DQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-30T02:26:25.889330Z"},"content_sha256":"8823596905e1753b9c7631b60d1f4dcaed9156496f54350e7bbf96ff9f3f4ae4","schema_version":"1.0","event_id":"sha256:8823596905e1753b9c7631b60d1f4dcaed9156496f54350e7bbf96ff9f3f4ae4"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:WEUR63GCLWX24NJ42VP6JS2VQ6","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Behavior Cue Reasoning: Monitorable Reasoning Improves Efficiency and Safety through Oversight","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Training LLMs to emit Behavior Cues before behaviors makes reasoning monitorable, allowing recovery of safe actions from 80% of unsafe traces and doubling success rates.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Christopher Z. Cui, Prithviraj Ammanabrolu, Taylor W. Killian","submitted_at":"2026-05-07T23:05:50Z","abstract_excerpt":"Reasoning in Large Language Models (LLMs) poses a challenge for oversight as many misaligned behaviors do not surface until reasoning concludes. To address this, we introduce Behavior Cue Reasoning for making LLM reasoning more controllable and monitorable. Behavior Cues are special token sequences that a model is trained to emit immediately before specific implicit and explicit behaviors, acting as dual purpose signal and control levers. When fine-tuning a weaker external monitor with Reinforcement Learning for reasoning oversight, a compressed view of only information surfaced by Behavior Cu"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"When leveraged by an almost optimal rule-based monitor in an environment where excessive constraint violations results in failure, Behavior Cue Reasoning allows for the recovery of safe actions from 80% of reasoning traces that would otherwise end with the proposal of an unsafe action, more than doubling the success rate from 46% to 96%.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That LLMs can be reliably fine-tuned to emit Behavior Cues immediately before target behaviors without any degradation to core reasoning performance, and that a compressed cue-only view provides sufficient information for external monitors to make effective pruning or recovery decisions.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Behavior Cue Reasoning trains LLMs to emit special tokens before behaviors, enabling monitors to prune up to 50% of wasted tokens and recover safe actions from 80% of unsafe traces, more than doubling success rates with no performance loss.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Training LLMs to emit Behavior Cues before behaviors makes reasoning monitorable, allowing recovery of safe actions from 80% of unsafe traces and doubling success rates.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"039f3f361c266384c81aed828ecebd3601327ecd19af302b640a5e98de214559"},"source":{"id":"2605.07021","kind":"arxiv","version":2},"verdict":{"id":"92ddc632-05d9-4400-a272-b5b68e6a7f66","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-11T00:50:04.133299Z","strongest_claim":"When leveraged by an almost optimal rule-based monitor in an environment where excessive constraint violations results in failure, Behavior Cue Reasoning allows for the recovery of safe actions from 80% of reasoning traces that would otherwise end with the proposal of an unsafe action, more than doubling the success rate from 46% to 96%.","one_line_summary":"Behavior Cue Reasoning trains LLMs to emit special tokens before behaviors, enabling monitors to prune up to 50% of wasted tokens and recover safe actions from 80% of unsafe traces, more than doubling success rates with no performance loss.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That LLMs can be reliably fine-tuned to emit Behavior Cues immediately before target behaviors without any degradation to core reasoning performance, and that a compressed cue-only view provides sufficient information for external monitors to make effective pruning or recovery decisions.","pith_extraction_headline":"Training LLMs to emit Behavior Cues before behaviors makes reasoning monitorable, allowing recovery of safe actions from 80% of unsafe traces and doubling success rates."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.07021/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"claim_evidence","ran_at":"2026-05-20T11:42:03.433512Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-20T06:37:42.565227Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T17:31:18.941554Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T12:11:24.935444Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"b2edc4fed531163f0fd9f23ff91b2d5c0a5068b7850171ffff7e3c244e15e43c"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"5ea8923e6b2477a127db1fed46be8f0aca66026adda92398553239f5f934bc18"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"92ddc632-05d9-4400-a272-b5b68e6a7f66"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-21T01:04:27Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"hrT39Zi4AHd1B15Y0//fWysoTIbd6qO01fRs/bIw001dxa1CMIx10H14XweEsL5sDoB9bTcBkUNT0oPossI5AA==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-30T02:26:25.890032Z"},"content_sha256":"72c88c3b278afed108155138184a0038b3f1c0ffdb5ee0350b754c878248f6bc","schema_version":"1.0","event_id":"sha256:72c88c3b278afed108155138184a0038b3f1c0ffdb5ee0350b754c878248f6bc"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/WEUR63GCLWX24NJ42VP6JS2VQ6/bundle.json","state_url":"https://pith.science/pith/WEUR63GCLWX24NJ42VP6JS2VQ6/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/WEUR63GCLWX24NJ42VP6JS2VQ6/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-05-30T02:26:25Z","links":{"resolver":"https://pith.science/pith/WEUR63GCLWX24NJ42VP6JS2VQ6","bundle":"https://pith.science/pith/WEUR63GCLWX24NJ42VP6JS2VQ6/bundle.json","state":"https://pith.science/pith/WEUR63GCLWX24NJ42VP6JS2VQ6/state.json","well_known_bundle":"https://pith.science/.well-known/pith/WEUR63GCLWX24NJ42VP6JS2VQ6/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:WEUR63GCLWX24NJ42VP6JS2VQ6","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"30e05c3adb0fc0e6680b12fbc18a3ed184e0baddf91c6fb1a4930a808f781f9e","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-07T23:05:50Z","title_canon_sha256":"1ceb985b886599ff635b05a228404db77f4ff4615b65a9bf992570f272ff7482"},"schema_version":"1.0","source":{"id":"2605.07021","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.07021","created_at":"2026-05-21T01:04:27Z"},{"alias_kind":"arxiv_version","alias_value":"2605.07021v2","created_at":"2026-05-21T01:04:27Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.07021","created_at":"2026-05-21T01:04:27Z"},{"alias_kind":"pith_short_12","alias_value":"WEUR63GCLWX2","created_at":"2026-05-21T01:04:27Z"},{"alias_kind":"pith_short_16","alias_value":"WEUR63GCLWX24NJ4","created_at":"2026-05-21T01:04:27Z"},{"alias_kind":"pith_short_8","alias_value":"WEUR63GC","created_at":"2026-05-21T01:04:27Z"}],"graph_snapshots":[{"event_id":"sha256:72c88c3b278afed108155138184a0038b3f1c0ffdb5ee0350b754c878248f6bc","target":"graph","created_at":"2026-05-21T01:04:27Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"When leveraged by an almost optimal rule-based monitor in an environment where excessive constraint violations results in failure, Behavior Cue Reasoning allows for the recovery of safe actions from 80% of reasoning traces that would otherwise end with the proposal of an unsafe action, more than doubling the success rate from 46% to 96%."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That LLMs can be reliably fine-tuned to emit Behavior Cues immediately before target behaviors without any degradation to core reasoning performance, and that a compressed cue-only view provides sufficient information for external monitors to make effective pruning or recovery decisions."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Behavior Cue Reasoning trains LLMs to emit special tokens before behaviors, enabling monitors to prune up to 50% of wasted tokens and recover safe actions from 80% of unsafe traces, more than doubling success rates with no performance loss."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Training LLMs to emit Behavior Cues before behaviors makes reasoning monitorable, allowing recovery of safe actions from 80% of unsafe traces and doubling success rates."}],"snapshot_sha256":"039f3f361c266384c81aed828ecebd3601327ecd19af302b640a5e98de214559"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"5ea8923e6b2477a127db1fed46be8f0aca66026adda92398553239f5f934bc18"},"integrity":{"available":true,"clean":true,"detectors_run":[{"findings_count":0,"name":"claim_evidence","ran_at":"2026-05-20T11:42:03.433512Z","status":"completed","version":"1.0.0"},{"findings_count":0,"name":"ai_meta_artifact","ran_at":"2026-05-20T06:37:42.565227Z","status":"completed","version":"1.0.0"},{"findings_count":0,"name":"doi_title_agreement","ran_at":"2026-05-19T17:31:18.941554Z","status":"completed","version":"1.0.0"},{"findings_count":0,"name":"doi_compliance","ran_at":"2026-05-19T12:11:24.935444Z","status":"completed","version":"1.0.0"}],"endpoint":"/pith/2605.07021/integrity.json","findings":[],"snapshot_sha256":"b2edc4fed531163f0fd9f23ff91b2d5c0a5068b7850171ffff7e3c244e15e43c","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"abstract_excerpt":"Reasoning in Large Language Models (LLMs) poses a challenge for oversight as many misaligned behaviors do not surface until reasoning concludes. To address this, we introduce Behavior Cue Reasoning for making LLM reasoning more controllable and monitorable. Behavior Cues are special token sequences that a model is trained to emit immediately before specific implicit and explicit behaviors, acting as dual purpose signal and control levers. When fine-tuning a weaker external monitor with Reinforcement Learning for reasoning oversight, a compressed view of only information surfaced by Behavior Cu","authors_text":"Christopher Z. Cui, Prithviraj Ammanabrolu, Taylor W. Killian","cross_cats":[],"headline":"Training LLMs to emit Behavior Cues before behaviors makes reasoning monitorable, allowing recovery of safe actions from 80% of unsafe traces and doubling success rates.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-07T23:05:50Z","title":"Behavior Cue Reasoning: Monitorable Reasoning Improves Efficiency and Safety through Oversight"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.07021","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-11T00:50:04.133299Z","id":"92ddc632-05d9-4400-a272-b5b68e6a7f66","model_set":{"reader":"grok-4.3"},"one_line_summary":"Behavior Cue Reasoning trains LLMs to emit special tokens before behaviors, enabling monitors to prune up to 50% of wasted tokens and recover safe actions from 80% of unsafe traces, more than doubling success rates with no performance loss.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Training LLMs to emit Behavior Cues before behaviors makes reasoning monitorable, allowing recovery of safe actions from 80% of unsafe traces and doubling success rates.","strongest_claim":"When leveraged by an almost optimal rule-based monitor in an environment where excessive constraint violations results in failure, Behavior Cue Reasoning allows for the recovery of safe actions from 80% of reasoning traces that would otherwise end with the proposal of an unsafe action, more than doubling the success rate from 46% to 96%.","weakest_assumption":"That LLMs can be reliably fine-tuned to emit Behavior Cues immediately before target behaviors without any degradation to core reasoning performance, and that a compressed cue-only view provides sufficient information for external monitors to make effective pruning or recovery decisions."}},"verdict_id":"92ddc632-05d9-4400-a272-b5b68e6a7f66"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:8823596905e1753b9c7631b60d1f4dcaed9156496f54350e7bbf96ff9f3f4ae4","target":"record","created_at":"2026-05-21T01:04:27Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"30e05c3adb0fc0e6680b12fbc18a3ed184e0baddf91c6fb1a4930a808f781f9e","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-07T23:05:50Z","title_canon_sha256":"1ceb985b886599ff635b05a228404db77f4ff4615b65a9bf992570f272ff7482"},"schema_version":"1.0","source":{"id":"2605.07021","kind":"arxiv","version":2}},"canonical_sha256":"b1291f6cc25dafae353cd55fe4cb5587a5c88c327317f6bd0f55a6acb47d95a6","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"b1291f6cc25dafae353cd55fe4cb5587a5c88c327317f6bd0f55a6acb47d95a6","first_computed_at":"2026-05-21T01:04:27.041460Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-21T01:04:27.041460Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"Zvqb9VnUJ4H0dEKC8+8zvHQz7q032LAqwBAxcrKju4DydQ1hTKWSHvBXPL1TfGgFczY2A5Vam5ZfxtDSHGmiCQ==","signature_status":"signed_v1","signed_at":"2026-05-21T01:04:27.042278Z","signed_message":"canonical_sha256_bytes"},"source_id":"2605.07021","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:8823596905e1753b9c7631b60d1f4dcaed9156496f54350e7bbf96ff9f3f4ae4","sha256:72c88c3b278afed108155138184a0038b3f1c0ffdb5ee0350b754c878248f6bc"],"state_sha256":"95d7481cae4e5213db68d6d56db3625afcaf9f2f2bdb44cc371c0379c9585995"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"z4jh1cWpvXOstUvhLN5VpcTzd72PCg9VbtHllJyvN6PT63B0F341JrsbZgn3Dl0lKXftF8P5pGV9J8H6+X8/DA==","signed_message":"bundle_sha256_bytes","signed_at":"2026-05-30T02:26:25.893301Z","bundle_sha256":"b04be543bc6985b62986cfe19973f9137921a430d771a6f37cf7987e9a633a5e"}}