Sharpening Tax: Post-Training Cuts "Sometimes Solved" Tasks from 87.6% to 30% at 128 Attempts
iScienceLuvr · x · 2026-10-02
Project page for "Sharpening Tax in Post-Training" (Changdae Oh et al., work done at Meta). The sharpening hypothesis: post-training amplifies existing behaviors, improving first-try accuracy while narrowing solution coverage. Agents make a revealing test bed since they must execute structured actions and stay coherent across turns. Experiments across WebShop, BFCL, ACEBench on multiple model pairs (Gemma 4 31B, Qwen 2.5 32B, Qwen 3.5 35B-A3B, Ministral 3 14B) with 500 tasks and 128 rollouts per task show tasks that were "sometimes solved" fall from 87.6% to 30.0% post-training. The fix: explore on hard prompts, stay precise on easy ones, via posterior-tempered group sampling (PTGS).
More from Research
- Sparse MoE residual activations beat dense on interpretability, reports collab — amplifiedamp · 2026-10-25
- CoDA builds qualitative codebooks in under 2 hours with 5x fewer examples — BrihiJ · 2026-10-25
- KLD Is Misunderstood: Why It Can't Rank Quantized Models Reliably — Positive-Resource922 · 2026-10-25
- Interpretability meme ranks the field from linear probes to 'Sacred Geometry' oracle mysticism — voooooogel · 2026-10-25
- Researcher sorts problems by permanence: O(n log n) sorting lower bound holds forever — mboehme_ · 2026-10-25
- SuperMimic code still unreleased 3 days on, already derivative reimplementations: just ship code — carlosdponx · 2026-10-25