AI Paper Atlas

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

Task-and-Model-Aware Fractal-Consistency for Efficient LLM Reasoning

Ziqiu Luo, Jianmin Liu, Yukai Miao, Li Chen, Dan Li

ICML 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 a task-and-model-aware, fractal-style consistency method for LLM reasoning that aims to reduce the compute cost of self-consistency by evaluating consistency at finer granularity and using adaptive rather than static stopping/aggregation.

Problem

Self-consistency methods aggregate many stochastic LLM samples to improve answer correctness, but they are computationally expensive because they judge output consistency monolithically and use static sampling/aggregation settings.

Approach

Evaluates consistency fractally, so partially correct answers across samples can be combined efficiently, and adapts the consistency procedure to the task and model rather than using static settings.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add Fractal-Consistency and ASCC (TMAFC) over Adaptive-ConsistencyGPT-4 sampling · structured tasks (NER)sampling cost: 36.81 -> 16.44Improved
add Fractal-Consistency and ASCC (TMAFC) over Adaptive-ConsistencyGPT-4 sampling · EE (structured)accuracy: 41.8% -> 43.6%Improved
add component-wise Fractal-Consistency (remove ASCC) (ablation)TMAFC · NER (structured)sampling cost: 36.81 (AC) -> 19.32 (w/o ASCC)Improved
add ASCC on top of Fractal-Consistency (ablation)TMAFC · NER (structured)sampling cost: 19.32 -> 16.44Improved
add ASCC on top of Fractal-Consistency (ablation)TMAFC · EE (structured)accuracy: 43.7% -> 43.6%No clear effect
add Fractal-Consistency on top of ASCC (w/o FC variant) (ablation)TMAFC · Math-Reasoning (Hard)accuracy: 68.1% -> 68.5%Improved
add Fractal-Consistency on top of ASCC (w/o FC variant) (ablation)TMAFC · Mathematical Reasoning (unstructured)sampling cost: 28.52 -> 21.24Improved
adaptive prior (alpha0, beta0) instead of fixed (1,1)Adaptive-Consistency · hybrid tasks, GPT-4sampling cost: 8.61 -> 4.14Improved
adaptive prior on weak model on hard questionsAdaptive-Consistency · hybrid tasks, Llama3-8B-Instructsampling cost: 24.66 -> 22.4 (-9.1%)Improved
inject synthetic anchor errors (0% to 50%) (ablation)TMAFC · unstructured tasksaccuracy and sampling costHurt
replace LLM-based difficulty classifier with XGBoost or DistilBERT (ablation)ASCC · cross-domain 300 questionsend-to-end performanceHurt
vary classifier accuracy 65% to 90% via noise injection (ablation)ASCC · cross-domain 300 questionsaccuracy and costNo clear effect

Limitations

Experimental setup

Baselines: self-consistency

Builds on: 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