TL;DR

A controlled study of where reasoning diversity disappears during RLVR, distinguishing failure to initiate a solution family from failure to execute it after the first reasoning choice.

Why it matters

Higher pass@1 can hide a narrower policy. If training concentrates probability on fewer initial strategies, the model may look stronger on one sample while producing less value from additional test-time sampling.

Key findings

  1. 01

    Across PPO and GRPO setups on the Countdown task, the authors report solution coverage falling by as much as 67% even as single-sample accuracy improves.

  2. 02

    Likelihood shifts before the first arithmetic operation are reported to be 11–16× larger than shifts during downstream reasoning.

  3. 03

    An entrance-targeted parameter intervention increases solution coverage by 37% with no reported loss in pass@1, while surface prompting does not recover the diversity.

Scaling dimensions

Models, methods & benchmarks

Models
Qwen2.5-3B, Qwen2.5-3B-Instruct
Algorithms
RLVR, PPO, GRPO
Benchmarks
Countdown

Topics

Read the original sourcearXiv