{"paper":{"title":"Revisiting DAgger in the Era of LLM-Agents","license":"http://creativecommons.org/licenses/by/4.0/","headline":"DAgger with turn-level interpolation mitigates covariate shift in multi-turn LLM agents while retaining dense teacher supervision.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Bo Dai, Changhao Li, Chao Zhang, Chenxiao Gao, Jiawei Huang, Niao He, Rushi Qiang","submitted_at":"2026-05-13T02:40:28Z","abstract_excerpt":"Long-horizon LM agents learn from multi-turn interaction, where a single early mistake can alter the subsequent state distribution and derail the whole trajectory. Existing recipes fall short in complementary ways: supervised fine-tuning provides dense teacher supervision but suffers from covariate shift because it is trained on off-policy teacher trajectories; while reinforcement learning with verifiable rewards avoids this off-policy mismatch by learning from on-policy rollouts but with only sparse outcome feedback. We address this dilemma by revisiting Dataset Aggregation (DAgger) for multi"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"DAgger-style training with turn-level interpolation of student and teacher policies mitigates covariate shift while retaining dense teacher supervision, producing +3.9 and +3.6 point gains on SWE-bench Verified for 4B and 8B models respectively.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That a reliable teacher policy remains available and affordable to query at every training step and that the environment state can be reset or continued after mixed student-teacher actions without introducing new distribution shifts.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"DAgger-style training with turn-level policy interpolation raises 4B and 8B LLM agents to 27.3% and 29.8% on SWE-bench Verified, beating several larger published systems.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"DAgger with turn-level interpolation mitigates covariate shift in multi-turn LLM agents while retaining dense teacher supervision.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"7a96e2bd3b01fb21b2f9c13017685477ff1ee08735f5dbd980b0fafb57e16a5b"},"source":{"id":"2605.12913","kind":"arxiv","version":1},"verdict":{"id":"a3874483-226c-44a6-9fdc-986324510ae4","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:50:59.965269Z","strongest_claim":"DAgger-style training with turn-level interpolation of student and teacher policies mitigates covariate shift while retaining dense teacher supervision, producing +3.9 and +3.6 point gains on SWE-bench Verified for 4B and 8B models respectively.","one_line_summary":"DAgger-style training with turn-level policy interpolation raises 4B and 8B LLM agents to 27.3% and 29.8% on SWE-bench Verified, beating several larger published systems.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That a reliable teacher policy remains available and affordable to query at every training step and that the environment state can be reset or continued after mixed student-teacher actions without introducing new distribution shifts.","pith_extraction_headline":"DAgger with turn-level interpolation mitigates covariate shift in multi-turn LLM agents while retaining dense teacher supervision."},"references":{"count":61,"sample":[{"doi":"","year":2023,"title":"GPT-4 Technical Report","work_id":"b928e041-6991-4c08-8c81-0359e4097c7b","ref_index":1,"cited_arxiv_id":"2303.08774","is_internal_anchor":true},{"doi":"","year":2024,"title":"On-policy distillation of language models: Learning from self-generated mistakes","work_id":"3733ff2d-cd95-4776-9fa9-1b2328326749","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Dream: Deep research evaluation with agentic metrics.arXiv preprint arXiv:2602.18940, 2026","work_id":"8010f1a4-ea3c-4f11-8c31-857e66722934","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Swe-rebench: An automated pipeline for task collection and decontaminated evaluation of software engineering agents","work_id":"c819f5fe-32b6-41d1-aeb5-2d80c6fa8474","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Qwen3-Coder-Next Technical Report","work_id":"ad966e68-641d-4b33-a9da-57cf741f35a6","ref_index":5,"cited_arxiv_id":"2603.00729","is_internal_anchor":true}],"resolved_work":61,"snapshot_sha256":"0c51235a8f3e9b1cfe5acdfae95e90eaa98571c42bcbc7d6544e0ec802102218","internal_anchors":22},"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"}