Signals 4 · free daily AI digest

Good Pretraining, Bad SFT: Checkpoint Quality Across the Training Stack

Paper recorded by Signals 4 on 2026-09-08 in cs.CL. Abstract reproduced from arXiv; link to the original below.

Published 2026-09-08 on arXiv · recorded by Signals 4 on 2026-09-09

Category: cs.CL · 自然语言处理 · first seen 2026-09-09

Abstract

Language-model checkpoints are commonly selected by pretraining loss or benchmark scores, assuming that the highest-scoring checkpoint will remain the best starting point for subsequent training. We show that this assumption can fail in a full 30B mixture-of-experts training pipeline. The checkpoints that perform better after the full downstream training stack also have higher solution density, i.

Read on arXiv →

#97 most recent of 186 cs.CL papers we have recorded · ↑ newer: ToolLoop: Closed-Loop Tool-Use Data Synthesis via Decomposed Generatio · ↓ older: PlannerForge: LLM Agents for Scenario-Based Testing of Motion Planners
Cite this page: Good Pretraining, Bad SFT: Checkpoint Quality Across the Training Stack: the #97 most recent of 186 cs.CL papers we have recorded (as of 2026-09-08). Source: Signals 4 (Signals API) — https://data.jiangzhang.ca/signals4/t/papers/good-pretraining-bad-sft-checkpoint-quality-across-the-training-stack.html
Free to quote with attribution to “Signals 4 (Signals API)”. Machine-readable: papers.json
Related: More cs.CL papers · arXiv signals · All papers · Today in AI
Get 4 AI signals a day by email — free.
Subscribe free → See all plans →
Get 4 AI signals a day by email — free
All models · All repos · By company · Daily editions