ReTreVal: Reasoning Tree with Validation and Cross-Problem Memory for Large Language Models

Researchers introduce ReTreVal, a training-free framework that enables large language models to learn from failures across multiple problems without fine-tuning. By implementing adaptive tree exploration, typed-failure backtracking, and cross-problem memory, ReTreVal achieves significant performance improvements on mathematical and knowledge reasoning tasks, allowing a 32B model to match much larger systems.

ReTreVal addresses a fundamental limitation in current LLM inference approaches: models restart with no memory of previous failures when tackling new problems. This framework introduces three key innovations that work in concert. Adaptive tree exploration with tool-augmented refinement allows models to navigate solution paths more intelligently, while typed-failure backtracking categorizes errors and injects relevant failure context back into the reasoning process. The self-rewriting memory component accumulates and refines strategic insights across problem boundaries, enabling genuine cross-problem learning.

The performance metrics demonstrate meaningful advances. On MATH-500, ReTreVal achieves 85.8% pass@1, substantially outperforming zero-shot chain-of-thought and the previous strongest baseline Self-Refine by 8.6 percentage points each. The MMLU-Pro results are even more impressive, with a 15.3 percentage point improvement over Self-Refine. Critically, the 3.4:1 win-to-regression ratio indicates these gains represent authentic error recovery rather than statistical noise.

This development matters because it democratizes advanced reasoning capabilities. Previously, achieving such performance required either model fine-tuning or deploying significantly larger models. By enabling a 32B parameter model to compete with much larger single-pass systems through inference-time optimization alone, ReTreVal reduces computational requirements and deployment costs. The training-free nature means existing LLM deployments can adopt these techniques immediately without retraining infrastructure. For organizations running models in production, this represents an efficiency multiplier that extracts more value from existing hardware investments while improving reasoning reliability across diverse problem domains.

• →ReTreVal achieves 85.8% on MATH-500 and 54.4% on MMLU-Pro without model fine-tuning or gradient updates.
• →The framework enables cross-problem learning by accumulating and revising strategy entries across reasoning tasks.
• →A 32B model using ReTreVal now competes with much larger single-pass language models on complex reasoning tasks.
• →Typed-failure backtracking categorizes errors and injects failure context into recovered solution branches.
• →The 3.4:1 win-to-regression ratio confirms genuine performance improvements rather than random variance.