AIAny
Icon for item

Knowing When Not to Reuse: Conditional Experience Transfer in Autonomous LLM Post-Training

Decides when past post-training updates should be reused for autonomous LLM adaptation by introducing Boundary-Calibrated Intervention Transfer (BCIT). BCIT binds effects to source context, checks applicability and hard conflicts, and runs bounded trials to obtain current-state evidence—reducing harmful updates and improving equal-budget final-model quality.

Introduction

Accumulating successful post-training updates and treating them as unconditional permissions to reuse is surprisingly risky: later weight changes can make prior evidence irrelevant or harmful. This paper reframes that risk as an explicit authorization problem—conditional experience transfer—and proposes a practical guardrail for autonomous LLM post-training.

Key Findings
  • Core idea: BCIT (Boundary-Calibrated Intervention Transfer) ties an observed update effect to the source context, evaluates applicability conditions, vetoes named hard conflicts, and—when needed—runs a short, bounded training trial to get fresh evidence before reusing experience.
  • Empirical result: On a 4B model adapted across finance reasoning, text-to-SQL, and function calling, candidate updates show heterogeneous target and retention effects; BCIT authorizes fewer harmful updates and yields higher final-model quality under equal compute/budget compared to naïve reuse baselines.
  • Practical implication: Safe reuse requires lightweight state checks and limited trials, not blind replay of past successes. This reduces wasted compute and the risk of promoting updates that degrade downstream training trajectories.
Method overview

BCIT operates before weight-changing training: it records source-context bindings for observed effects, enforces applicability conditions, applies explicit vetoes for hard conflicts, and executes bounded trial training when current-state evidence is missing. Fully trained candidates still pass a shared adoption rule; only observed events extend memory.

Who it's for + tradeoffs

Great fit if you run autonomous post-training loops that propose and evaluate model updates (e.g., continual adaptation, domain transfer, tool-enabled agents) and need to limit regressions from naive experience reuse. Look elsewhere if you cannot afford any bounded trial overhead or if updates are so low-stakes that simple heuristics suffice—the BCIT protocol adds orchestration cost and requires instrumentation to bind effects to contexts.

Where it fits

BCIT sits between retrieval/selection of candidate updates and full promotion: it separates the authorization problem from candidate generation/evaluation and complements existing post-training and unsupervised update techniques by adding applicability checks and verification trials.

Information

  • Websitearxiv.org
  • OrganizationsAbudukelimu Wuerkaixi, Guohua Liu, Yuewei Zhang
  • AuthorsTingyun Li, Wenfeng Feng, Weiqing Li, Abudukelimu Wuerkaixi, Guohua Liu, Yuewei Zhang
  • Published date2026/08/27

More Items

Reconstructs executable terminal workspaces from recorded agent trajectories and synthesizes verifiable single- and multi-round coding tasks for agent training; it replays file operations, uses an LLM completion agent to fill missing files/dependencies, and verifies tasks with autogenerated test suites.

Converts a natural-language function specification into a reusable local neural function by using teacher models to synthesize examples and finetuning a small adapter for a compact interpreter. Achieves higher semantic accuracy (83.6% on FuzzyBench-Hard) at the cost of roughly one minute compile time; produces versionable PAW artifacts for local deployment.

Studies on-policy distillation (OPD) at the data-minimal limit by training on a single query, measuring state coverage and alignment dynamics, and showing OPD is often data-overfed but algorithm-starved.