TailSFT Paper: Filtering Fitted Sequences in SFT Boosts Post-RL pass@1 by up to 4%
rohanpaul_ai · x · 2026-09-04
- arXiv paper "TailSFT: Filtered Fine-Tuning Improves Post-Training Performance" (Sadhika Malladi, Samy Jelassi, et al., Microsoft) argues existing SFT pipelines may not yield checkpoints best suited for RL, building on coverage/pass@K as predictors of post-RL performance.
- Method: filter out already-fit sequences during SFT, focusing learning on the under-modeled tail of the data distribution; design choices validated via controlled experiments and theory.
- Results: on OLMo-3 7B, pass@16 on math/coding improves up to 17% absolute with minimal overhead, and subsequent GRPO runs gain up to 4% absolute pass@1. Includes a lightweight diagnostic for when TailSFT helps, advocating stage-aware checkpoint selection.
Related event: Microsoft's TailSFT: Filtered Fine-Tuning Improves RL Performance(2 posts)→
More from Research
- Stanford team launches ScholarCatalyst benchmark for inspiration retrieval — lateinteraction · 2026-10-21
- Holon-S1 predicts gene activity and biomarkers from routine pathology slides — altryne · 2026-10-21
- Three reasons AI is not conscious and may never be, per Nature essay — anilkseth · 2026-10-21
- Aviation tracks every part's provenance — agents can reason over that graph — colinmcnamara · 2026-10-21
- SFT details: instruction/reasoning/code/math mix, synthetic + human traces, no epoching — TimDarcet · 2026-10-21
- Mid-training recipe disclosed: 32k→131k context extension, dropping logit distillation for tokens — TimDarcet · 2026-10-21