AI Paper Atlas

Research map / Large language models / Reasoning and chain of thought

When Simple Problems Wear Complex Costumes: Improving Efficiency in LRM's Adaptive Reasoning

Junnan Ren, Yan Zhang, Qian Chen, Yunhang Shen, Ke Li, Shengchuan Zhang, Liujuan Cao, Rongrong Ji

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Addresses overthinking in large reasoning models by improving adaptive reasoning, so simple problems get direct answers and complex ones get explicit reasoning (details limited to the abstract).

Problem

Large reasoning models overthink simple tasks, applying complex multi-step reasoning that wastes computation and adds latency; existing adaptive reasoning models that switch between explicit reasoning and direct answers are the focus of this work.

Approach

Improves adaptive reasoning in LRMs so the model decides when to reason explicitly versus answer directly; the specific mechanism is not stated in the available abstract.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
apply AdaRS two-stage training (SFT with redundancy augmentation + GRPO)DeepSeek-R1-Distill-Qwen-1.5B Think · GSM8K, Math500, AIME 2024 averageaccuracy: 62.4 -> 65.6Improved
apply AdaRS two-stage trainingDeepSeek-R1-Distill-Qwen-1.5B Think · GSM8K, Math500, AIME 2024 averagetokens per instance: 7413 -> 3250Improved
add Cold Start SFT with concise and redundant augmented queries (ablation)DeepSeek-R1-Distill-Qwen-1.5B · GSM8Ktokens per instance: 1634 -> 1047Improved
add Cold Start SFT (ablation)DeepSeek-R1-Distill-Qwen-1.5B · ASDIVaccuracy: 89.3 -> 91.8Improved
use GRPO with naive reward that ignores reasoning cost (ablation)Cold Start model · GSM8Ktokens per instance: 1047 -> 1809Hurt
use mode-aware reward (Enhancing Adaptive Policy) instead of naive reward (ablation)GRPO stage · GSM8Ktokens per instance: 1809 -> 503Improved
use mode-aware reward instead of naive reward (ablation)GRPO stage · ASDIVtokens per instance: 1766 -> 308Improved
add mode-aware reward on top of Cold Start (ablation)Cold Start model · GSM8Kaccuracy: 82.7 -> 83.5Improved
train with redundancy-augmented data and custom reward vs prior adaptive modelsadaptive reasoning models · ASDIV-redundancyThink%: 58.5% (AdaptThink) -> 33.1% (AdaRS)Improved
add redundant phrasing to simple queriesAdaRS · ASDIVThink%: 4.3% -> 33.1%Hurt
add redundant phrasing to simple queriesAdaptThink · ASDIVThink%: 21.9% -> 58.5%Hurt
use AdaRS adaptive switching instead of forced Nothink modeDeepSeek-R1-Distill-Qwen-1.5B · GSM8K, Math500, AIME 2024 averageaccuracy: 52.2 -> 65.6Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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