Saved in:
Bibliographic Details
Main Authors: Dineen, Jacob, RRV, Aswin, Xu, Zhikun, Zhou, Ben
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2604.03472
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910173953523712
author Dineen, Jacob
RRV, Aswin
Xu, Zhikun
Zhou, Ben
author_facet Dineen, Jacob
RRV, Aswin
Xu, Zhikun
Zhou, Ben
contents Co-evolutionary self-play, where one language model generates problems and another solves them, promises autonomous curriculum learning without human supervision. In practice, the proposer quickly converges to a narrow distribution of problems that satisfy the reward function. This diversity collapse renders the curriculum uninformative for the solver, stalling the co-evolutionary loop. We introduce vocabulary dropout, a random mask applied to the proposer's output logits during both policy training and curriculum generation, as a lightweight mechanism to sustain diversity. The mask is hard and non-stationary, preventing the proposer from locking into fixed token sequences. Training Qwen3-4B and Qwen3-8B on mathematical reasoning via R-Zero, we find that vocabulary dropout sustains proposer diversity across lexical, semantic, and functional metrics throughout training, and yields solver improvements averaging +4.4 points at 8B, with the largest gains on competition-level benchmarks. Our findings suggest that explicit action-space constraints, analogous to the structural role that game rules play in classical self-play, can help sustain productive co-evolution in language. Vocabulary dropout is one simple instantiation of this principle.
format Preprint
id arxiv_https___arxiv_org_abs_2604_03472
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Vocabulary Dropout for Curriculum Diversity in LLM Co-Evolution
Dineen, Jacob
RRV, Aswin
Xu, Zhikun
Zhou, Ben
Computation and Language
Artificial Intelligence
Co-evolutionary self-play, where one language model generates problems and another solves them, promises autonomous curriculum learning without human supervision. In practice, the proposer quickly converges to a narrow distribution of problems that satisfy the reward function. This diversity collapse renders the curriculum uninformative for the solver, stalling the co-evolutionary loop. We introduce vocabulary dropout, a random mask applied to the proposer's output logits during both policy training and curriculum generation, as a lightweight mechanism to sustain diversity. The mask is hard and non-stationary, preventing the proposer from locking into fixed token sequences. Training Qwen3-4B and Qwen3-8B on mathematical reasoning via R-Zero, we find that vocabulary dropout sustains proposer diversity across lexical, semantic, and functional metrics throughout training, and yields solver improvements averaging +4.4 points at 8B, with the largest gains on competition-level benchmarks. Our findings suggest that explicit action-space constraints, analogous to the structural role that game rules play in classical self-play, can help sustain productive co-evolution in language. Vocabulary dropout is one simple instantiation of this principle.
title Vocabulary Dropout for Curriculum Diversity in LLM Co-Evolution
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2604.03472