Conformal Thinking: Risk Control for Reasoning on a Compute Budget
Reasoning Large Language Models (LLMs) benefit from test-time scaling, with accuracy improving as token budget increases, motivating adaptive reasoning. However, setting token budgets and thresholds involves a risk-accuracy trade-off. This paper re-frames budget setting as risk control, limiting error rate while minimizing compute. It introduces an upper threshold to stop when confident and a lower threshold to stop unsolvable instances. Using distribution-free risk control, the framework optimally specifies stopping mechanisms given a target risk and validation set. Experiments across diverse tasks and models demonstrate computational efficiency gains while adhering to risk targets.
content type paperpublished July 2026
Conformal Thinking: Risk Control for Reasoning on a Compute Budget
AuthorsXi Wang†*, Anushri Suresh†*, Alvin Zhang†*, Rishi More†*, William Jurayj†, Benjamin Van Durme†, Mehrdad Farajtabar, Daniel Khashabi†, Eric Nalisnick†
View publication
Reasoning Large Language Models (LLMs) enable test-time scaling, with dataset-level accuracy improving as the token budget increases, motivating adaptive reasoning—spending tokens when they improve reliability and stopping early when additional computation is unlikely to help. However, setting the token budget, as well as the threshold for adaptive reasoning, is a practical challenge that entails a fundamental risk-accuracy trade-off. We re-frame the budget setting problem as risk control, limiting the error rate while minimizing compute. Our framework introduces an upper threshold that stops reasoning when the model is confident (risking incorrect output) and a novel parametric lower threshold that preemptively stops unsolvable instances (risking premature stoppage). Given a target risk and a validation set, we use distribution-free risk control to optimally specify these stopping mechanisms. Empirical results across diverse reasoning tasks and models demonstrate the effectiveness of our risk control approach, demonstrating computational efficiency gains from the lower threshold and ensemble stopping mechanisms, all while adhering to the user-specified risk target. Code is available at https://github.com/xidulu/reasoning_risk_control/.
† Johns Hopkins University
- Equal contribution
LaDiR: Latent Diffusion Enhances LLMs for Text Reasoning
April 28, 2026research area Speech and Natural Language Processingconference ICLR
Large Language Models (LLMs) demonstrate their reasoning ability through chain-of-thought (CoT) generation. However, LLM’s autoregressive decoding may limit the ability to revisit and refine earlier tokens in a holistic manner, which can also lead to inefficient exploration for diverse solutions. In this paper, we propose LaDiR (Latent Diffusion Reasoner), a novel reasoning framework that unifies the expressiveness of continuous latent…
Read more
Worst Cases Policy Gradients
November 9, 2019research area Methods and Algorithmsconference CoRL
Recent advances in deep reinforcement learning have demonstrated the capability of learning complex control policies from many types of environments. When learning policies for safety-critical applications, it is essential to be sensitive to risks and avoid catastrophic events. Towards this goal, we propose an actor-critic framework that models the uncertainty of the future and simultaneously learns a policy based on that uncertainty model…
Read more