{"paper":{"title":"Macro-Action Based Multi-Agent Instruction Following through Value Cancellation","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Correcting the Bellman backup target at each instruction boundary decouples value estimates across contexts, allowing a single policy to follow interrupting instructions while preserving base-task performance in multi-agent settings.","cross_cats":["cs.MA"],"primary_cat":"cs.AI","authors_text":"Enrico Marchesini Xiang Zhi Tan, Ethan Rathbun, Wo Wei Lin","submitted_at":"2026-05-12T19:01:16Z","abstract_excerpt":"Multi-agent reinforcement learning (MARL) in real-world use cases may need to adapt to external natural language instructions that interrupt ongoing behavior and conflict with long-horizon objectives. However, conditioning rewards on instructions introduces a fundamental failure mode as Bellman updates couple value estimates across instruction contexts, leading to inconsistent values when instructions interrupt macro-actions. We propose Macro-Action Value Correction for Instruction Compliance (MAVIC), which corrects Bellman backups at instruction boundaries by correcting the incoming instructi"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"MAVIC achieves high instruction compliance while preserving base task performance in increasingly complex cooperative multi-agent environments.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That correcting the bootstrapping target at instruction boundaries is sufficient to fully decouple value estimates across contexts without introducing new inconsistencies under stochastic switching or macro-action interruptions.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"MAVIC corrects Bellman backups at instruction boundaries by adjusting the incoming objective and restoring continuation value, enabling consistent estimation under stochastic instruction switching in a unified policy.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Correcting the Bellman backup target at each instruction boundary decouples value estimates across contexts, allowing a single policy to follow interrupting instructions while preserving base-task performance in multi-agent settings.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"db942b2271ccb422000eb5a68eb94c0fce7bf00258e2963a03f50d84c618cd04"},"source":{"id":"2605.12655","kind":"arxiv","version":1},"verdict":{"id":"7d79834b-368f-4ab7-815a-1cdca8494a7b","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:32:37.115796Z","strongest_claim":"MAVIC achieves high instruction compliance while preserving base task performance in increasingly complex cooperative multi-agent environments.","one_line_summary":"MAVIC corrects Bellman backups at instruction boundaries by adjusting the incoming objective and restoring continuation value, enabling consistent estimation under stochastic instruction switching in a unified policy.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That correcting the bootstrapping target at instruction boundaries is sufficient to fully decouple value estimates across contexts without introducing new inconsistencies under stochastic switching or macro-action interruptions.","pith_extraction_headline":"Correcting the Bellman backup target at each instruction boundary decouples value estimates across contexts, allowing a single policy to follow interrupting instructions while preserving base-task performance in multi-agent settings."},"references":{"count":4,"sample":[{"doi":"","year":2021,"title":"bring me the tomato,","work_id":"c529cb9f-0686-4e65-9c98-3a674afd2fcb","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1999,"title":"When c=c ′, the reward is unchanged","work_id":"b7590f98-9c18-416a-bcc3-e45947033d44","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Don’t use left cutting board","work_id":"adad8480-b299-45e0-8d41-eafaaff87697","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Go to small boxes","work_id":"ced46404-1005-42e0-8a0d-ba22a2e91f4c","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":4,"snapshot_sha256":"dc6afdeced95dd3b205a114113da17d69f20013332fe56e6db4f3b2b1d7a6f4d","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"}