The authors compare base and post-trained LLMs on agentic tasks (multi-turn tool use) and find that base models, despite lower single-shot accuracy (pass@1), often achieve higher solution coverage (pass@K) given repeated sampling. They propose a diagnostic metric called Sharpening Tax to quantify the loss in test-time scalability after post-training. This tax is prevalent across 14 model pairs from 4 families on 3 benchmarks (42 cases). To mitigate the trade-off, they introduce Posterior-Tempered Group Sampling (PTGS), a Bayesian sampler that adapts sampling temperature per prompt based on estimated difficulty. Applied during RL training, PTGS reduces the sharpening tax, solving more tasks under repeated sampling while also improving pass@1.
RL post-training improves agent consistency but narrows solution coverage
Across agentic benchmarks, base models explored more successful strategies under repeated sampling, while post-trained models performed better on individual attempts.
Top University
Changdae Oh · Qi Zeng · Qi Qi · Andrey Zhmoginov · Deren Lei · Yun He · +4 more
Meta Superintelligence Labs · University of Wisconsin–Madison · NYU · Stanford University
Research Digest··3 min read
In one line
Post-training RL sharpens LLMs, improving pass@1 but reducing pass@K coverage on agentic tasks.
Before you read this
Three ideas this paper builds onreinforcement learningA machine learning method where an agent learns to make sequences of decisions by receiving rewards or penalties for its actions.296 papers →large language modelA neural network trained on massive text corpora to generate coherent text and perform diverse language tasks.329 papers →agentic evaluationMeasuring AI agents' performance in complex tasks, highlighting failures in planning, monitoring, and real304 papers →
Read it as
The authors compare 14 base and post-trained model pairs on three interactive agent benchmarks, using a lightweight harness to let base language models call tools and act across multiple turns.
The study evaluates model pairs from four families across three agentic benchmarks, producing 42 base versus post-training comparisons.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.
What it released
Code