e1: Learning Adaptive Control of Reasoning Effort

TL;DR

This paper introduces Adaptive Effort Control, a reinforcement learning method that allows AI models to dynamically allocate reasoning effort based on user preferences, improving cost-accuracy tradeoffs and efficiency across various model sizes.

Contribution

It presents a novel self-adaptive reinforcement learning approach enabling models to adjust reasoning effort proportionally to task difficulty without prior tuning.

Findings

2-3x reduction in reasoning length while maintaining performance

Better cost-accuracy tradeoff curves compared to standard methods

Effective across models from 1.5B to 32B parameters

Abstract

Increasing the thinking budget of AI models can significantly improve accuracy, but not all questions warrant the same amount of reasoning. Users may prefer to allocate different amounts of reasoning effort depending on how they value output quality versus latency and cost. To leverage this tradeoff effectively, users need fine-grained control over the amount of thinking used for a particular query, but few approaches enable such control. Existing methods require users to specify the absolute number of desired tokens, but this requires knowing the difficulty of the problem beforehand to appropriately set the token budget for a query. To address these issues, we propose Adaptive Effort Control, a self-adaptive reinforcement learning method that trains models to use a user-specified fraction of tokens relative to the current average chain-of-thought length for each query. This approach eliminates dataset- and phase-specific tuning while producing better cost-accuracy tradeoff curves compared to standard methods. Users can dynamically adjust the cost-accuracy trade-off through a continuous effort parameter specified at inference time. We observe that the model automatically learns to allocate resources proportionally to the task difficulty and, across model scales ranging from 1.5B to 32B parameters, our approach enables a 2-3x reduction in chain-of-thought length while maintaining or improving performance relative to the base model used for RL training.