Meta Paper: Post-Training Boosts pass@1 but Shrinks LLM Agents' pass@K Solution Coverage
iScienceLuvr · x · 2026-10-02
Meta researchers (Superintelligence Labs) introduce "Sharpening Tax," a diagnostic metric measuring how post-training reduces test-time scalability of LLM agents. Surprising finding: base models with a light inference harness often surpass post-trained counterparts in pass@K coverage despite lower pass@1, since post-training pushes per-task success toward 0% or 100%. They also propose PTGS (posterior-tempered group sampling), a plug-and-play Bayesian sampler that adapts temperature per prompt difficulty.
More from Research
- Paper resolving a 15-year-old AdaBoost question, aided by LLMs, accepted to NeurIPS 2026 — chetanp · 2026-11-19
- Anthropic analyzes 207k research sessions: data analysis dominates, formal sciences use Claude most — plopesresearch · 2026-11-19
- Ramp releases SWE-Bench v2: 500 production-built tasks, 6.2 Astra tops at 76.4% — AniBaddepudi · 2026-11-19
- Stanford DWD paper: linear head predicts pixels directly, ditching the diffusion decoder — kwangmoo_yi · 2026-11-19
- Stanford's Le Cong to talk AI co-scientists for RNA design at 14th mRNA Health Conference — MengdiWang10 · 2026-11-19
- New paper shows DiT latents can be linearly probed to predict pixels, skipping the VAE decoder — kwangmoo_yi · 2026-11-19