AI Paper Atlas

Research map / Large language models / Agents, code and math

OptimalThinkingBench: Evaluating Over and Underthinking in LLMs

Pranjal Aggarwal, Seungone Kim, Jack Lanchantin, Sean Welleck, Jason E Weston, Ilia Kulikov, Swarnadeep Saha

ICLR 2026 · Large language models › Agents, code and math · approach: Mathematical and structural reasoning benchmarks · 40 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add AdaptThink length-aware RL trainingR1-Distill-Qwen-7B · OptimalThinkingBenchF1: 24.5 -> 38.3Improved
add L1 length-control trainingR1-Distill-Qwen-7B · OptimalThinkingBenchF1: 24.5 -> 20.8Hurt
add SB-DS efficient reasoningR1-Distill-Qwen-7B · UnderthinkingBenchaccuracy: 23.6 -> 14.5Hurt
add VeriThinker verification trainingR1-Distill-Qwen-7B · OptimalThinkingBenchF1: 24.5 -> 27.4Mixed
merge thinking and non-thinking weightsQwen3-8B · OptimalThinkingBenchF1: 24.3 -> 38.2Improved
add L1 length-control trainingQwen3-8B · UnderthinkingBenchaccuracy: 47.7 -> 34.6Hurt
route between thinking and non-thinking with a trained routerQwen3 family · OptimalThinkingBenchF1: 24.3 -> 46.9Improved
use oracle router instead of trained routerQwen3 family · OptimalThinkingBenchF1: 46.9 -> 61.2Improved
add Don't Overthink prompt suffixQwen3 models · OverthinkingBenchthinking tokens: 1493 -> 1147Improved
add Let's think step-by-step prompt suffixQwen3 models · OptimalThinkingBenchF1: 26.3 -> 18.3Hurt
add irrelevant MCQ distractor optionsQwen3 models · OverthinkingBenchthinking tokens: 42 tokens per option, R^2=0.94Hurt
enable thinking mode vs non-thinking mode (ablation)Qwen3-14B · UnderthinkingBenchaccuracy: 14% -> 52.4%Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy