DiSCTT: Consensus-Guided Self-Curriculum for Efficient Test-Time Adaptation in Reasoning

TL;DR

DiSCTT introduces a dynamic, consensus-guided self-curriculum for test-time adaptation in reasoning models, improving accuracy and efficiency by tailoring optimization strategies based on instance difficulty and uncertainty.

Contribution

It presents a novel framework that adaptively applies different optimization methods based on instance-level uncertainty, enhancing test-time reasoning performance.

Findings

Outperforms existing test-time adaptation methods across reasoning benchmarks.

Achieves higher accuracy with lower computational costs.

Reduces variance in model performance.

Abstract

Test-time adaptation offers a promising avenue for improving reasoning performance in large language models without additional supervision, but existing approaches often apply a uniform optimization objective across all inputs, leading to inefficient or unstable adaptation on heterogeneous reasoning problems. We propose DiSCTT, a difficulty-aware, consensus-guided self-curriculum framework that dynamically allocates test-time optimization strategies based on instance-level epistemic uncertainty estimated from agreement among sampled reasoning trajectories. Inputs with high consensus are consolidated via supervised fine-tuning using majority-agreed solutions as pseudo-labels, while low-consensus inputs are optimized via reinforcement learning with a consensus-regularized objective that encourages diversity under relevance constraints. Across a broad suite of mathematical and general reasoning benchmarks, DiSCTT consistently outperforms strong test-time adaptation baselines, achieving higher accuracy with reduced variance and substantially lower computation and wall-clock training times. These results demonstrate that explicitly accounting for instance difficulty and uncertainty enables more stable, efficient, and effective test-time adaptation for reasoning models.