Reproduce “DiSCTT: Consensus-Guided Self-Curriculum for Efficient Test-Time Adaptation in Reasoning” on OpenResearch

About this paper

Mohammad Mahdi Moradi, Sudhir Mudur

Researchers at Concordia University developed DiSCTT, a framework for efficient test-time adaptation in reasoning that dynamically allocates supervised fine-tuning and reinforcement learning based on an online, consensus-guided estimation of problem difficulty. This approach significantly improves accuracy, stability, and computational efficiency across various reasoning benchmarks compared to uniform adaptation strategies.

Key insights

• Effective test-time adaptation requires dynamically allocating different learning objectives based on the model's instance-level uncertainty, moving beyond uniform optimization.
• The degree of agreement (consensus) among independently sampled reasoning paths can serve as a robust, label-free proxy for estimating epistemic uncertainty and problem difficulty in LLMs.
• Combining a correctness-gated, population-relative novelty reward with semantic relevance filtering is essential for stabilizing label-free reinforcement learning and ensuring focused, meaningful exploration during test-time adaptation.

Summary by alphaXiv.