{"paper":{"title":"When Attention Closes: How LLMs Lose the Thread in Multi-Turn Interaction","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"LLMs lose track of instructions in multi-turn chats when attention to goal tokens fades, though residual streams may still encode the needed information.","cross_cats":["cs.CL"],"primary_cat":"cs.AI","authors_text":"Dilek Hakkani-T\\\"ur, Joseph Hsieh, Seunghyun Yoon, Trung Bui, Vardhan Dongre, Viet Dac Lai","submitted_at":"2026-05-13T02:58:18Z","abstract_excerpt":"Large language models can follow complex instructions in a single turn, yet over long multi-turn interactions they often lose the thread of instructions, persona, and rules. This degradation has been measured behaviorally but not mechanistically explained. We propose a channel-transition account: goal-defining tokens become less accessible through attention, while goal-related information may persist in residual representations. We introduce the Goal Accessibility Ratio (GAR), measuring attention from generated tokens to task-defining goal tokens, and combine it with sliding-window ablations a"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across architectures and model scales, the gap between attention loss and residual decodability predicts whether goal-conditioned behavior survives channel closure.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the linear probes on residual streams recover causally relevant goal information rather than spurious correlations, and that the sliding-window ablations isolate the attention channel without confounding other mechanisms.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Attention to goal tokens declines in multi-turn LLM interactions while residual representations often retain decodable goal information, and the gap between these predicts whether goal-conditioned behavior survives.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"LLMs lose track of instructions in multi-turn chats when attention to goal tokens fades, though residual streams may still encode the needed information.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"66c60cca120bc073a4b1d313963698529129e4d39ab4a99e19780123d969800d"},"source":{"id":"2605.12922","kind":"arxiv","version":1},"verdict":{"id":"dbd61569-cdfb-4f4d-81c5-ff1b18d7e154","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:16:15.218921Z","strongest_claim":"Across architectures and model scales, the gap between attention loss and residual decodability predicts whether goal-conditioned behavior survives channel closure.","one_line_summary":"Attention to goal tokens declines in multi-turn LLM interactions while residual representations often retain decodable goal information, and the gap between these predicts whether goal-conditioned behavior survives.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the linear probes on residual streams recover causally relevant goal information rather than spurious correlations, and that the sliding-window ablations isolate the attention channel without confounding other mechanisms.","pith_extraction_headline":"LLMs lose track of instructions in multi-turn chats when attention to goal tokens fades, though residual streams may still encode the needed information."},"references":{"count":39,"sample":[{"doi":"","year":null,"title":"LLMs Get Lost In Multi-Turn Conversation","work_id":"01fa2ab4-4b40-4e8a-9487-2de7db19cfe2","ref_index":1,"cited_arxiv_id":"2505.06120","is_internal_anchor":true},{"doi":"","year":null,"title":"Multi-if: Benchmarking llms on multi-turn and multilingual instructions following","work_id":"2e46fb28-13e3-4c54-88b5-a2d47f2a064b","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Transactions of the association for computational linguistics , volume=","work_id":"35c18954-8a0c-4935-b39c-5981194920ba","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Roformer: Enhanced transformer with rotary position embedding , author=. Neurocomputing , volume=. 2024 , publisher=","work_id":"0cfd4ae7-837c-4c4f-bf6d-afc10f5a8ed1","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=","work_id":"756542c0-c2f6-4557-9f1f-50c1be6099a8","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":39,"snapshot_sha256":"e9bd9f0947179197d22903ff1caab17d502cfdd23f9c82b831f5c672746b816f","internal_anchors":17},"formal_canon":{"evidence_count":2,"snapshot_sha256":"c864106dc6b25539c1110e1faf9125625aec269b7b8ef15cc8dbc800ea61e563"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}