Mitigating Error Propagation in LLM Mathematical Reasoning via Disambiguation and Back-Verification

While Large Language Models (LLMs) have demonstrated emerging capabilities in complex mathematical reasoning, simply scaling inference-time search (e.g., Tree-of-Thought or Monte Carlo Tree Search) often yields diminishing returns. We identify two fundamental bottlenecks driving this inefficiency: ambiguity propagation, where early misinterpretations of implicit problem structures cascade through the reasoning process, and correlated errors, where shared biases across search paths render standard aggregation methods ineffective. To overcome these limitations, we propose BEATS (Back-Verify and Adaptive Disambiguation Tree Search), a unified, zero-shot inference framework. BEATS intervenes at the root by adaptively disambiguating problem statements before reasoning begins. It then employs a Constraint-based Pruning Tree Search (CPTS) to optimize search efficiency within bounded polynomial complexity, directly addressing the compute-performance trade-off. Finally, rather than relying on costly human-annotated Process Reward Models (PRMs), BEATS introduces an adaptive back-verification mechanism that functions as a zero-shot PRM by leveraging the model's inherent capacity to reverse-validate logical flows. Comprehensive evaluations across five rigorous mathematical benchmarks and three model families demonstrate that BEATS significantly mitigates error propagation, improving reasoning accuracy while strictly controlling inference-time compute.