EDO integrates exploration objectives into RL post-training of LLMs, yielding greater solution diversity, 1.0-1.3% gains on in-distribution reasoning benchmarks, and 1.5% on out-of-distribution tasks when paired with test-time methods.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Exploration-Driven Optimization for Test-Time Large Language Model Reasoning
EDO integrates exploration objectives into RL post-training of LLMs, yielding greater solution diversity, 1.0-1.3% gains on in-distribution reasoning benchmarks, and 1.5% on out-of-distribution tasks when paired with test-time methods.