AI Paper Atlas

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

CaTS: Calibrated Test-Time Scaling for Efficient LLM Reasoning

Chengsong Huang, Langlin Huang, Jixuan Leng, Jiacheng Liu, Jiaxin Huang

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Routing, pruning and distillation for reasoning

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.

Proposes CaTS, a calibrated test-time scaling method for LLM reasoning that adaptively decides how many samples to draw per query instead of using a fixed N.

Problem

Best-of-N and Self-Consistency with majority voting use a fixed number of samples per query regardless of difficulty, wasting computation on simple questions and under-exploring complex ones.

Approach

Adaptive test-time scaling that calibrates how many responses to sample per query based on query complexity, allocating more computation to harder questions and less to easier ones. Specific mechanism not stated in the available abstract.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
weight votes by self-calibrated confidence (CaTS-SC)Self-Consistency on Llama-3.1-8B-Instruct · MathQA, budget 16accuracy (%): 73.7 -> 83.6Improved
early stop Best-of-N at a confidence threshold (CaTS-ES)Best-of-N on DeepSeek-R1-Distill-1.5B · ARC Challenge, budget 16accuracy (%): 54.1 -> 66.5Improved
confidence-weighted adaptive SC (CaTS-ASC)Adaptive Self-Consistency on Llama-3.1-8B-Instruct · MathQA, budget 16accuracy (%): 72.7 -> 81.9Improved
use CaTS-ES instead of Best-of-NQwen2.5-7B-Instruct · ARC Challenge, budget 16accuracy (%): 91.1 -> 90.8Hurt
use CaTS-ES with a small sample budgetBest-of-N on Qwen2.5-7B-Instruct · MathQAaccuracyHurt
distill self-consistency confidence into the model (Self-Calibration)Llama-3.1-8B-Instruct · GSM8KECE: 13.70 -> 3.79Improved
Self-Calibration trainingQwen2.5-7B-Instruct · MathQAaccuracy (%): 49.85 -> 64.18Improved
use soft self-consistency score instead of P(True) or plain SC for labels (ablation)Llama-3.1-8B-Instruct · GSM8KECE: 12.03 (P(True)), 4.48 (SC) -> 3.42 (SSC)Improved
remove soft self-consistency from Self-Calibration (ablation)Self-Calibration · Object CountingECE: 16.02 -> 9.69 (adding SSC)Improved
remove entropy-based dynamic temperature (ablation)Self-Calibration · Object Countingaccuracy (%): 62.88 -> 65.88 (adding EDT)Improved
use SmoothL1 instead of MSE (ablation)Self-Calibration · Object Countingaccuracy (%): 56.48 -> 65.88Improved
use self-calibrated confidence instead of a separate reward model for Best-of-16Llama-3.1-8B-Instruct · MathQAaccuracy (%): 82.1 -> 84.0No clear effect

Experimental setup

Baselines: Best-of-N sampling Self-Consistency with majority voting

Builds on: Best-of-N sampling Self-Consistency

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