VeryTrace: Verifying Reasoning Traces through Compilable Formalism and Structured Verification

A new framework called VeryTrace aims to make multi-step reasoning by large language models more reliable by converting their natural-language thought processes into a structured, verifiable format, according to research submitted in 2026 [1]. The framework, detailed in a paper on arXiv, addresses a persistent weakness in Chain-of-Thought (CoT) prompting, where logical errors or hallucinations in early reasoning steps can cascade, leading to confident but incorrect final answers [1]. VeryTrace operates by formalizing a model's reasoning trace into a compilable representation using a custom Domain-Specific Language (DSL) [1]. This DSL makes step dependencies explicit, converts quantitative content into executable expressions, and structures semantic inferences using deduction schemas [1]. The system then applies a hybrid verification process that combines deterministic checks for computational correctness and constraint satisfaction with targeted audits by a language model for semantic judgments that cannot be mechanically verified [1]. This approach allows for step-level error localization and repair [1]. The researchers tested VeryTrace across three distinct domains: competition mathematics using the AIME 2025 benchmark, robotics planning with LLM-BabyBench, and kinship reasoning on the CLUTRR dataset [1]. The framework improved accuracy over zero-shot baselines on state-of-the-art language models without requiring domain-specific training or in-context examples, demonstrating that formalized trace verification can achieve both precision and generalization [1]. The concept of formalizing reasoning to eliminate ambiguity has deep roots. In mathematics, results are established by deductive proof rather than by experiment, a standard that began in Ancient Greek mathematics and was exemplified in Euclid's Elements [3]. The challenge VeryTrace tackles is analogous to the distinction between a "fuzzy concept," which has variable boundaries of application, and a "crisp concept," which is precise [4]. By forcing a language model's reasoning into a compilable DSL, the framework pushes the model's output from the fuzzy domain of natural language toward the crisp domain of formal logic, where deterministic checks can be applied [1][4].