FAER: Auditable Utility-Aligned Trajectory Replay for Language Model Post-Training

2026-09-30

AI Summary

This study addresses the misalignment between format-based feedback and downstream learning utility in trajectory replay by proposing an auditable full-trajectory replay framework. Methodologically, a learner-aware selector is introduced to optimize post-training effectiveness, while an optimizer-aware virtual update constructs a magnitude-aware utility surface, enabling training-independent baseline comparisons. Reliability is further ensured through normalized gradient alignment, metadata cross-fitting calibration, and frozen auditing contracts. Experimental results on the GSM8K dataset demonstrate that the proposed approach significantly improves generation quality to 0.6624, outperforming both uniform sampling and format-feedback baselines. Overall, this work establishes an efficient and auditable data selection paradigm for model post-training.

Abstract

Replay selectors often rank cached trajectories by format feedback, confidence, freshness, or response length, although cache-level correctness and downstream learner utility are distinct objectives. We formalize this selection-to-learning gap and introduce FAER as an auditable full-trajectory replay framework. Its training-free fixed selector is a protocol baseline; FAER-UTILITY is the learner-aware selector fitted on disjoint calibration blocks. The normalized gradient alignment is reported as a baseline, while a disposable optimizer-aware virtual update supplies a magnitude-aware utility surface. The audit contract freezes observed fields and replay traces before evaluation labels are joined. On GSM8K with Qwen2.5-1.5B-Instruct, the matched learner study reports quality 0.6329 for the fixed selector, compared with 0.5482 for uniform and 0.6037 for format-feedback under 128 updates. Metadata-only cross-fitted calibration reaches $0.6476\!\pm\!0.0139$ over eight seeds (median 0.6481; paired 95% interval $[+0.079,+0.122]$) at 63,276 target-run tokens; its recorded full cost is 189,642 tokens and 3.48 GPU-hours including calibration. The completed FAER-UTILITY row reaches 0.6624 at 62,844 target-run tokens and 4.26 GPU-hours. Format-feedback selects records with correctness 0.6953, compared with 0.3594 for the fixed selector, despite the different downstream ranking. The completed comparison surfaces report the learner-aware ablation, same-seed gap, policy-optimization rows, and strict zero-shot transfer.

Problem

• trajectory replay
• language model post-training
• selection-to-learning gap
• learner utility
• auditable framework

Innovation

• trajectory replay
• utility-aligned selection
• auditable framework
• post-training
• virtual update