ReThinker: Scientific Reasoning by Rethinking with Guided Reflection and Confidence Control

TL;DR

ReThinker is a confidence-aware, adaptive reasoning framework for large language models that improves expert-level scientific reasoning by dynamically orchestrating tools, reflection, and multi-agent coordination.

Contribution

It introduces a novel Solver-Critic-Selector architecture with confidence-based computation allocation and a scalable training pipeline using reverse data synthesis and trajectory recycling.

Findings

Outperforms state-of-the-art models on HLE, GAIA, and XBench benchmarks.

Achieves state-of-the-art results on expert-level reasoning tasks.

Demonstrates robustness and efficiency in scientific reasoning scenarios.

Abstract

Expert-level scientific reasoning remains challenging for large language models, particularly on benchmarks such as Humanity's Last Exam (HLE), where rigid tool pipelines, brittle multi-agent coordination, and inefficient test-time scaling often limit performance. We introduce ReThinker, a confidence-aware agentic framework that orchestrates retrieval, tool use, and multi-agent reasoning through a stage-wise Solver-Critic-Selector architecture. Rather than following a fixed pipeline, ReThinker dynamically allocates computation based on model confidence, enabling adaptive tool invocation, guided multi-dimensional reflection, and robust confidence-weighted selection. To support scalable training without human annotation, we further propose a reverse data synthesis pipeline and an adaptive trajectory recycling strategy that transform successful reasoning traces into high-quality supervision. Experiments on HLE, GAIA, and XBench demonstrate that ReThinker consistently outperforms state-of-the-art foundation models with tools and existing deep research systems, achieving state-of-the-art results on expert-level reasoning tasks.