mlsanityData to 14 Oct 2026

MAIGO: Mitigating Lost-in-Conversation with History-Cleaned On-Policy Self-Distillation

arXiv 2605.27186 ↗First posted 26 May 2026cs.CLcited by 6

On the map: Knowledge distillation → On-policy and self-distillation for LLMs

Abstract (arXiv)

Large language models often solve tasks from a fully specified prompt but degrade when the same requirements unfold over multiple turns, known as the lost-in-conversation (LiC) gap. We trace part of this degradation to self-contamination: intermediate assistant replies enter later context and carry early deviations forward. Motivated by this mechanism, we propose MAIGO, an on-policy self-distillation method that reduces this contamination using history-cleaned references from the model's own policy. For middle turns, MAIGO removes prior assistant replies while preserving the user-visible sharded prefix; for answer turns, it distills from paired full-view references conditioned on the completed user-side dialogue. A reliability weight downweights middle-turn samples that disagree with the clean reference. MAIGO requires no verifier rewards, state labels, or inference-time scaffolding. Under the LiC paired-view protocol with deterministic verifiers, MAIGO improves Qwen2.5-7B-Instruct SHARDED accuracy from 52.8 to 66.1 and the SHARDED/FULL ratio from 66.5% to 84.1%, while keeping FULL accuracy within 2.3 points. These results show that self-contamination is a trainable component of the LiC gap.

Key facts. Cited by 6 papers in the map (data through 14 Oct 2026). Placed in On-policy and self-distillation for LLMs, one of 12 directions in Knowledge distillation. Builds on 31 of its 62 references that are in the map. First posted on arXiv 26 May 2026.

Builds on

31 of 62 references.

Closest ancestors · 5

Same direction, newest first

  1. 2026The Many Faces of On-Policy Distillation: Pitfalls, Mechanisms, and Fixes11 May
  2. 2026TCOD: Exploring Temporal Curriculum in On-Policy Distillation for Multi-turn Autonomous Agents27 Apr
  3. 2026OPSDL: On-Policy Self-Distillation for Long-Context Language Models19 Apr
  4. 2026SCOPE: Signal-Calibrated On-Policy Distillation Enhancement with Dual-Path Adaptive Weighting12 Apr
  5. 2026On-Policy Context Distillation for Language Models12 Feb

Other references · 26

  1. 2022Training language models to follow instructions with human feedbackcited by 15,971
  2. 2024The Llama 3 Herd of Modelscited by 14,715
  3. 2021Training Verifiers to Solve Math Word Problemscited by 10,109
  4. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 9,379
  5. 2021Evaluating Large Language Models Trained on Codecited by 8,320
  6. 2023Lost in the Middle: How Language Models Use Long Contextscited by 3,455
  7. 2018Spider: A Large-Scale Human-Labeled Dataset for Complex and Cross-Domain Semantic Parsing and Text-to-SQL Taskcited by 1,139
  8. 2023On-Policy Distillation of Language Models: Learning from Self-Generated Mistakescited by 939
  9. 2021Measuring Coding Challenge Competence With APPScited by 902
  10. 2015Scheduled Sampling for Sequence Prediction with Recurrent Neural Networkscited by 803
  11. 2025LLMs Get Lost In Multi-Turn Conversationcited by 429
  12. 2026Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillationcited by 77
14 more
  1. 2025AdaMCoT: Rethinking Cross-Lingual Factual Reasoning through Adaptive Multilingual Chain-of-Thoughtcited by 14
  2. 2025CCL-XCoT: An Efficient Cross-Lingual Knowledge Transfer Method for Mitigating Hallucination Generationcited by 9
  3. 2026Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversationcited by 9
  4. 2026Breaking Contextual Inertia: Reinforcement Learning with Single-Turn Anchors for Stable Multi-Turn Interactioncited by 9
  5. 2026Mitigating Conversational Inertia in Multi-Turn Agentscited by 7
  6. 2025Mitigating Lost in Multi-turn Conversation via Curriculum RL with Verifiable Accuracy and Abstention Rewardscited by 6
  7. 2026SEQUOR: A Multi-Turn Benchmark for Realistic Constraint Followingcited by 6
  8. 2026Useless but Safe? Benchmarking Utility Recovery with User Intent Clarification in Multi-Turn Conversationscited by 3
  9. 2026Models Recall What They Violate: Constraint Adherence in Multi-Turn LLM Ideationcited by 3
  10. 2026When Attention Closes: How LLMs Lose the Thread in Multi-Turn Interactioncited by 3
  11. 2026Grounded Continuation: A Linear-Time Runtime Verifier for LLM Conversationscited by 3
  12. 2026ICPO: Illocution-Calibrated Policy Optimization for Multi-Turn Conversationcited by 2
  13. 2026Pause or Fabricate? Training Language Models for Grounded Reasoningcited by 2
  14. 2026Evaluating Temporal Consistency in Multi-Turn Language Modelscited by 1

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 6.

6 papers.

  1. 2026Trajectory-Relative Hindsight Distillation for Agentic Reinforcement Learningcited by 1
  2. 2026What if LLMs Ate Their Words: Causal History Effects in Multi-Turn Interactioncited by 1
  3. 2026SMOPD: Selective Token-Entropy Masking for Dirty-History Multi-Turn On-Policy Self-Distillationcited by 0
  4. 2026You Changed Your Mind, The Model Didn't: Demystifying Intent in Multi-Turn Dialoguecited by 0
  5. 2026Privileged Context as Drift in On-Policy Self-Distillationcited by 0
  6. 2026State-OPD: State-Grounded On-Policy Distillation for Multi-Turn Agentscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Found in Conversation: LLMs Teach Themselves to Close the Multi-Turn Gap
  2. 2026Learning What Not to Forget: Long-Horizon Agent Memory from a Few Kilobytes of Learning
  3. 2026Same Evidence, Different Answers: Canonical-Context On-Policy Distillation for Multi-Turn Language Models
  4. 2026SeDT: Sentence-Transformer Decision-Transformer Conditioning for Multi-Turn Conversation Reliability
  5. 2026Can Agents Forget When Asked? A Benchmark for Memory Unlearning in LLM Agents
  6. 2026LSREP: A Longitudinal State-Replay Protocol for Evaluating Conversational Memory, with ICE v2 as an Audited Local-First Architecture
  7. 2026MemOps: Benchmarking Lifecycle Memory Operations in Long-Horizon Conversations
  8. 2026Verifiable Self-Evolution for Open-Ended Dialogue Skills via Future-Feedback Prediction