Test-time Recursive Thinking: Self-Improvement without External Feedback

2026-02-03

AI Summary

Large language models struggle to self-improve without external feedback or additional training, primarily due to difficulties in generating high-quality candidate solutions and selecting correct answers in an unsupervised setting. This work proposes the Test-time Recursive Thinking (TRT) framework, which enables iterative self-optimization during inference through strategy-guided reasoning, knowledge accumulation, and self-generated validation signals. TRT achieves effective self-improvement for the first time without relying on reinforcement learning or human annotations, establishing an end-to-end test-time optimization pipeline. Experimental results demonstrate that open-source models attain 100% accuracy on AIME-25/24, while closed-source models show performance gains of 10.4–14.8 percentage points on the most challenging problems in LiveCodeBench.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search
• Natural Language Processing: (Large) Language Models

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Modern Large Language Models (LLMs) have shown rapid improvements in reasoning capabilities, driven largely by reinforcement learning (RL) with verifiable rewards. Here, we ask whether these LLMs can self-improve without the need for additional training. We identify two core challenges for such systems: (i) efficiently generating diverse, high-quality candidate solutions, and (ii) reliably selecting correct answers in the absence of ground-truth supervision. To address these challenges, we propose Test-time Recursive Thinking (TRT), an iterative self-improvement framework that conditions generation on rollout-specific strategies, accumulated knowledge, and self-generated verification signals. Using TRT, open-source models reach 100% accuracy on AIME-25/24, and on LiveCodeBench's most difficult problems, closed-source models improve by 10.4-14.8 percentage points without external feedback.

Problem

• self-improvement
• test-time
• large language models
• no external feedback
• reasoning

Innovation

• Test-time Recursive Thinking
• self-improvement
• large language models
• iterative reasoning
• self-generated verification