Search papers, labs, and topics across Lattice.
This study investigates the effectiveness of provenance-grounded gating and adaptive recovery strategies in synthetic post-training data curation, focusing on the interplay between filtering signals and the potential for recovering rejected samples. By employing adversarially injected corpora to establish ground-truth failure labels, the authors demonstrate that grounding filtering signals in source evidence enhances the accuracy of gating mechanisms, while the combination of hallucination and reward gates reveals distinct sample populations that necessitate both approaches. The proposed adaptive recovery pipeline significantly outperforms naive resampling in terms of yield, recovery rate, and injection recall, although downstream fine-tuning quality is primarily influenced by the scale of the generator used.
Grounding filtering signals in source evidence can drastically improve data curation outcomes, revealing that both hallucination and reward gates are essential for effective sample selection.
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.
