Provenance-Grounded Gating and Adaptive Recovery in Synthetic Post-Training Data Curation

2026-06-09

AI Summary

This work addresses the limitations of existing post-hoc training data filtering methods, which lack traceability to the origins of generated samples and often discard rejected instances without effective recovery mechanisms. To overcome these issues, the paper introduces a novel framework that jointly integrates a source-evidence-based credibility gating mechanism with an adaptive recovery pipeline. By leveraging adversarially injected corpora to provide ground-truth failure labels, the approach combines hallucination detection, reward-model-based gating, and targeted regeneration to enable precise failure diagnosis and high-fidelity sample restoration. Experimental results demonstrate that source tracing substantially enhances the fidelity of strong discriminator-based gating; the adaptive recovery strategy consistently outperforms naive resampling in terms of yield, recovery rate, and injection recall; and while fine-tuning quality is primarily driven by generator scale, both filtering and recovery strategies contribute significant additional gains.

Abstract

Synthetic post-training pipelines commonly filter generated samples with reward models or holistic LLM judges, yet two practices remain rarely examined together: whether the filtering signal is grounded in the source evidence that induced each generation, and whether rejected samples can be systematically recovered rather than permanently discarded. We present a controlled study of both questions across gate configurations, recovery strategies, and generator scales, using adversarially injected corpora to provide ground-truth failure labels. We find that exact source provenance improves faithfulness gating for stronger judges, that hallucination and reward gates reject largely disjoint sample populations making both necessary, and that an adaptive recovery pipeline combining failure diagnosis with targeted regeneration achieves higher yield, recovery rate, and injection recall than naive resampling. Downstream fine-tuning quality is driven primarily by generator scale, with filtration and recovery conditions contributing meaningfully but secondarily.

Problem

• provenance
• data curation
• synthetic post-training
• sample recovery
• hallucination

Innovation

• provenance-grounded gating
• adaptive recovery
• synthetic post-training
• failure diagnosis
• targeted regeneration