ReasonBENCH: Benchmarking the (In)Stability of LLM Reasoning

TL;DR

ReasonBENCH targets a blind spot in LLM evaluation by quantifying run-to-run instability in reasoning. It introduces a modular library and a ten-run, variance-aware benchmarking protocol across diverse tasks, methods, and models, complemented by a public leaderboard. Key findings show that many high-performing reasoning approaches are brittle and that prompts, parsing, and scaling significantly influence stability, with cost not reliably tracking reliability. The work advocates reproducibility and uncertainty quantification as essential, offering a foundation for more robust reasoning methods and variance-aware evaluation in the field.

Abstract

Large language models (LLMs) are increasingly deployed in settings where reasoning, such as multi-step problem solving and chain-of-thought, is essential. Yet, current evaluation practices overwhelmingly report single-run accuracy while ignoring the intrinsic uncertainty that naturally arises from stochastic decoding. This omission creates a blind spot because practitioners cannot reliably assess whether a method's reported performance is stable, reproducible, or cost-consistent. We introduce ReasonBENCH, the first benchmark designed to quantify the underlying instability in LLM reasoning. ReasonBENCH provides (i) a modular evaluation library that standardizes reasoning frameworks, models, and tasks, (ii) a multi-run protocol that reports statistically reliable metrics for both quality and cost, and (iii) a public leaderboard to encourage variance-aware reporting. Across tasks from different domains, we find that the vast majority of reasoning strategies and models exhibit high instability. Notably, even strategies with similar average performance can display confidence intervals up to four times wider, and the top-performing methods often incur higher and less stable costs. Such instability compromises reproducibility across runs and, consequently, the reliability of reported performance. To better understand these dynamics, we further analyze the impact of prompts, model families, and scale on the trade-off between solve rate and stability. Our results highlight reproducibility as a critical dimension for reliable LLM reasoning and provide a foundation for future reasoning methods and uncertainty quantification techniques. ReasonBENCH is publicly available at https://github.com/au-clan/ReasonBench .

Figures (4)

• Figure 1: Instability in LLM Reasoning. For the same query, different reasoning models (top) and reasoning strategies (middle) produce distinct chains of thought and frequently contradictory conclusions. Even when working from identical instructions, methods vary widely in their intermediate reasoning steps and the correctness of their final answers. The bottom panel summarizes this variability quantitatively, showing the relative deviation from average performance across reasoning models and strategies.
• Figure 2: ReasonBench architecture. Methods orchestrate the three core components: Agents, Environments, and Models. Agents translate states into prompts, query models, and parse responses into actions. Environments are drawn from a large task library and offer functions such as next-step transitions and scoring heuristics. Models provide a unified interface to external LLM APIs. States record the intermediate configurations of reasoning, enabling reproducibility and fair comparison across tasks and methods.
• Figure 3: Scaling Effects within a Model Family. Quality distributions for gpt-4.1-nano and gpt-4.1-mini across multiple reasoning strategies. The larger model achieves higher mean performance and exhibits markedly lower variance, suggesting greater stability in its reasoning behavior.
• Figure 4: Correlation between Quality and Cost. For FoA, quality scales positively with cost across all benchmarks. ReAct exhibits a consistent negative slope, indicating diminishing returns at higher costs. GoT does not follow a uniform pattern, with its cost–quality relationship varying substantially by task.