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

• URL: http://arxiv.org/abs/2605.27186v1
• Date: Tue, 26 May 2026 15:38:46 GMT
• Title: MAIGO: Mitigating Lost-in-Conversation with History-Cleaned On-Policy Self-Distillation
• Abstract summary: Large language models often solve tasks from a fully specified prompt but degrade when the same requirements unfold over multiple turns.<n>We propose MAIGO, an on-policy self-distillation method that reduces this contamination using history-cleaned references from the model's own policy.<n>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.
• Score: 51.32912774464992
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: 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.

Related papers

• MemToC: Benchmarking Memory-Tool Conflict Resolution in Large Language Models [49.47300047353726]
  We introduce MemToC, a benchmark for post-tool-return arbitration with executable tools.<n>MemToC comprises 6,504 evaluation episodes constructed from 542 quality-controlled factual questions.<n>Across five open-weight 7-9B models, tool returns strongly dominate elicited closed-book answers.
  arXiv  Detail & Related papers  (2026-08-26T18:22:03Z)
• Same Agent, Different Answers: A Repeat-Aware Audit of Corpus-Induced Answer Churn in Retrieval-Augmented QA [0.4351760066527804]
  A retrieval-augmented QA system can return different answers after an index expansion even when its requested model identifier, prompt, retrieval policy, evidence depth, rendering, and exposed generation controls are held fixed.<n>We call the hidden phenomenon accuracy-blind answer churn and introduce the emphSnapshot Compatibility Audit, which estimates excess answer churn by subtracting same-snapshot repeat disagreement from cross-snapshot disagreement.
  arXiv  Detail & Related papers  (2026-08-24T06:40:09Z)
• Denoising Iterative Self-Correction: Structured Verification Loops for Reliable LLM Reasoning [5.634990956257954]
  Large language models produce fluent but often incorrect multi-step reasoning, and naive correction methods risk degrading already-correct answers.<n>We introduce Denoising Iterative Self-Correction (DISC), a test-time procedure that treats verification question outputs as noisy measurements of where a solution may be corrupted.<n>A binary judgment gate controls correction precision by blocking rewrites that would damage already-correct answers while the verifier and corrector together repair errors.
  arXiv  Detail & Related papers  (2026-06-19T20:23:14Z)
• Understanding and Mitigating Premature Confidence for Better LLM Reasoning [76.16007941549857]
  Long chains of thought (CoT) from current language models frequently contain logical gaps and unjustified leaps.<n>We find such a signal in how the model's confidence evolves during reasoning.<n>We exploit this in progressive confidence shaping, a reinforcement learning objective that trains models to update their confidence as they reason rather than commit early.
  arXiv  Detail & Related papers  (2026-05-23T04:42:45Z)
• CyberCorrect: A Cybernetic Framework for Closed-Loop Self-Correction in Large Language Models [9.281774217584289]
  CyberCorrect is a framework that formalizes large language model self-correction.<n>A type-directed Correction Controller generates repair instructions based on diagnosed error categories.<n>A Convergence Judge determines iteration termination using stability criteria adapted from control theory.
  arXiv  Detail & Related papers  (2026-05-17T07:47:34Z)
• A Heterogeneous Temporal Memory Governance Framework for Long-Term LLM Persona Consistency [2.447989276487424]
  Large language models often suffer from fact loss, timeline confusion, persona drift, and reduced stability during long-range interaction.<n>We introduce ARPM, an external temporal memory governance framework for long-term dialogue.<n>We show that ARPM treats continuity as a traceable, auditable, and transferable governance problem.
  arXiv  Detail & Related papers  (2026-05-14T13:12:49Z)
• When Reasoning Traces Become Performative: Step-Level Evidence that Chain-of-Thought Is an Imperfect Oversight Channel [5.106950500256654]
  Chain-of-thought (CoT) traces are increasingly used to improve language model capability and to audit model behavior.<n>We test this assumption with a step-level Detect-Classify-Compare framework built around an answer-commitment proxy.<n>Across nine models and seven reasoning benchmarks, latent commitment and explicit answer arrival align on only 61.9% of steps on average.
  arXiv  Detail & Related papers  (2026-05-12T08:24:47Z)
• Reinforced Agent: Inference-Time Feedback for Tool-Calling Agents [6.158612515104146]
  We introduce Helpfulness-Harmfulness metrics to measure the tradeoff between agent feedback and reviewer feedback.<n>We evaluate our approach on BFCL and Tau2-Bench (multi-turn stateful scenarios), achieving +5.5% on irrelevance detection and +7.1% on multi-turn tasks.<n>Our metrics reveal that reviewer model choice is critical: the reasoning model o3-mini achieves a 3:1 benefit-to-risk ratio versus 2.1:1 for GPT-4o.
  arXiv  Detail & Related papers  (2026-04-29T22:09:47Z)
• Reasoning or Rationalization? The Role of Justifications in Masked Diffusion Models for Fact Verification [0.0]
  Masked Diffusion Language Models generate tokens sequentially and benefit from reasoning-before-answering strategies.<n>We show that MDLMs converge on a verdict early in the diffusion process, treating it as a global anchor that is resolved before the justification is complete.<n>Experiments reveal that the model rationalizes incorrect forced verdicts in 56% of cases, and that verdicts are strongly causally dependent on justification quality.
  arXiv  Detail & Related papers  (2026-03-01T17:18:05Z)
• GATES: Self-Distillation under Privileged Context with Consensus Gating [89.62339954332248]
  We study self-distillation in settings where supervision is unreliable.<n>We focus on document-grounded question answering with asymmetric context.<n>We derive supervision online from tutor consensus by sampling multiple document-grounded reasoning traces.
  arXiv  Detail & Related papers  (2026-02-24T05:56:20Z)
• HLE-Verified: A Systematic Verification and Structured Revision of Humanity's Last Exam [63.84155758655084]
  Humanity's Last Exam (HLE) has become a widely used benchmark for evaluating frontier large language models.<n>We introduce HLE-Verified, a verified and revised version of HLE with a transparent verification protocol and fine-grained error taxonomy.<n>We evaluate seven state-of-the-art language models on HLE and HLE-Verified, observing an average absolute accuracy gain of 7--10 percentage points.
  arXiv  Detail & Related papers  (2026-02-15T02:50:15Z)
• CARE What Fails: Contrastive Anchored-REflection for Verifiable Multimodal [84.71254539482369]
  Group-relative reinforcement learning with verifiable rewards (RLVR) often wastes the most informative data it already has the failures.<n>We present CARE, a failure-centric post-training framework for multimodal reasoning that turns errors into supervision.<n> CARE improves accuracy and training smoothness while explicitly increasing the share of learning signal that comes from failures.
  arXiv  Detail & Related papers  (2025-12-22T16:34:21Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.