Skip to content

Diversifying RLVR Rollouts via First-Token Exploration

May 2026 · 2 citations · 61 references
Computer Science

TL;DR

This work identifies the first token of the response as a structurally distinct target for diversification, largely overlooked in prior work, and finds that the first-token distribution is sharply concentrated and only weakly related to downstream correctness, as lower-probability candidates can yield similarly accurate responses.

Abstract

Reinforcement learning with verifiable rewards (RLVR) trains reasoning models without labeled trajectories, using groups of verifier-scored rollouts to explore alternative reasoning paths. Limited rollout diversity is a central bottleneck, typically addressed through adjustments to temperature, prefixes, or rollout selection. We identify the first token of the response as a structurally distinct target for diversification, largely overlooked in prior work. We find that the first-token distribution is sharply concentrated and only weakly related to downstream correctness, as lower-probability candidates can yield similarly accurate responses. Diversifying the first token can therefore broaden the reasoning paths explored within each rollout group with little loss in response quality. Motivated by this observation, we introduce REFT (Rollout Exploration with First-Token Diversification), a lightweight modification to RLVR. REFT samples first tokens uniformly from the policy's top-$N$ candidates and allocates rollouts evenly across the sampled tokens, leaving the rest of the pipeline unchanged. We evaluate REFT on eight models spanning multiple architectures and sizes (0.5B-14B), with mathematical reasoning and code-generation tasks under GRPO and DAPO. Across these settings, REFT consistently improves Pass@1, Pass@8, and Pass@64. It also outperforms competing diversification methods at every evaluated budget, incurring the lowest rollout cost.

View source

Similar papers

Preprint Oct 2026

Activation-Space Exploration for Visual and Textual Reasoning with RLVR

Reinforcement learning with verifiable rewards (RLVR) improves visual and textual reasoning by sampling responses, scoring them with an automatic verifier, and updating the policy from the resulting rewards. The policy therefore receives direct learning signals only from trajectories that appear in its rollouts. Yet re...

Yu-Nan Zeng, Jin-Hao Jing, Xi Cheng et al. · 0 citations
#artificial intelligence Preprint Oct 2026

Decoupling Exploration from Optimization in RLVR

Modern language models undergo reinforcement learning with verifiable rewards (RLVR) on top of already-trained checkpoints. A key promise of RLVR is the discovery of new reasoning strategies. In principle, a model can sample novel ideas absent from its prior training data. In practice, however, augmenting RLVR with str...

Saif Punjwani, Micah Goldblum · 0 citations
Preprint Oct 2026

Adaptive Decoding Methods for Token-Efficient Reinforcement Learning

Rollout decoding affects both which trajectories are used for reinforcement learning with verifiable rewards (RLVR) and how many completion tokens are spent collecting them. We study whether candidate-token support should remain fixed or widen during training in the context of rollout sampling efficiency. Our main sche...

Zhao-Nan Li, Zelei Cheng, Amritansh Mishra et al. · 0 citations
Preprint Oct 2026

ReAPO: Retry-Aware Policy Optimization for Efficient Exploration in LLM Reasoning

RLVR improves language-model reasoning, but commonly relies on independent stochastic rollouts for exploration, which can repeatedly sample similar failed strategies. Prior work has seen success with approaches that guide exploration by introducing dependence into the sampling process by including incorrect solutions i...

Nghi Dao, Calder Morton-Ferguson, Noah Golowich et al. · 1 citation
#artificial intelligence Preprint Oct 2026

Hindsight-Privileged RLVR: Single-Rollout Token Credit Without Group Sampling

Reinforcement learning with verifiable rewards (RLVR) trains language models from checked task outcomes, but assigning a terminal reward to individual reasoning tokens or actions remains difficult. Group-relative methods such as Group Relative Policy Optimization (GRPO) sample several attempts per prompt, but lose thei...

Wei Chen, Yu-Bing Wu, Zhou-Shi Xu et al. · 0 citations

Related blog posts

MIT News · Artificial Intelligence Nov 17, 2026

Learning the language of machine learning

By studying the computational foundations of language learning, Associate Professor Jacob Andreas builds intelligent systems that can communicate effectively with humans.

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.