According to arxiv.org, researchers have introduced ReTreVal (Reasoning Tree with Validation), a training-free framework that enables large language models to learn from failures across multiple problems without requiring fine-tuning.
The research addresses what the authors describe as a fundamental limitation in existing inference-time reasoning frameworks: “Every existing inference-time reasoning framework discards all failure context at problem boundaries, leaving a model solving problem 500 no wiser than it was on problem 1,” according to the paper abstract.
ReTreVal implements three core mechanisms: adaptive tree exploration with tool-augmented node refinement, typed-failure backtracking that categorizes errors and injects context into recovered branches, and a self-rewriting memory system that accumulates and revises strategy entries across problems. According to the source, this approach enables “inference-time cross-problem learning on any fixed, unmodified LLM without fine-tuning.”
The framework achieved 85.8% pass@1 accuracy on the MATH-500 benchmark, representing an 8.6 percentage point improvement over both Zero-Shot Chain-of-Thought and the Self-Refine baseline. On MMLU-Pro, it reached 54.4% accuracy, a 15.3 percentage point improvement over Self-Refine. The paper reports a 3.4:1 win-to-regression ratio, which the authors say confirms “genuine error recovery rather than noise.”
According to arxiv.org, these capabilities allow a 32B parameter model to compete with larger single-pass systems using methods that previously required gradient updates.