Suppress reasoning-model overthinking with batch prompting — ~76% fewer reasoning tokens

OpenReview·high signal

Batching multiple questions into a single prompt acts as an implicit regularizer against overthinking in reasoning models, cutting reasoning tokens by ~76% on DeepSeek-R1 and o1-class models while maintaining or improving accuracy — with no model changes. Complementary 2026 tactics include generating two low-effort solutions and selecting the one with the lower 'overthinking score' (~43% cheaper). For cost-sensitive agentic loops, the reasoning trace is usually the dominant token cost, so this is a high-leverage knob.