Think Short, Defer Smart, Act, and Repeat: Calibrated Reasoning and Uncertainty-Aware Deferral for Edge LLM Agents
stat.ML, cs.AI, cs.IT, cs.LG, math.IT
Submitted: 2026-07-29
Updated: 2026-09-25
Code: https://github.com/amirfar76/think-short-defer-smart
License: http://creativecommons.org/licenses/by/4.0/
The gist: LLM agents following the ReAct paradigm are promising enablers of complex multi-step tasks, including multi-hop question answering, code generation, and control of physical AI systems.
Terminology
Abstract
LLM agents following the ReAct paradigm are promising enablers of complex multi-step tasks, including multi-hop question answering, code generation, and control of physical AI systems. Yet, when deployed at the edge, they must tightly manage their reasoning budget while remaining reliable and deferring to a cloud-side model only when local uncertainty is too high to act safely. We propose Think Short, Defer Smart (TSDS), a framework that synergistically integrates a lightweight convergence probe, which halts on-device reasoning once the intended action has stabilized, with a perplexity-based deferral rule that escalates uncertain actions to a cloud-side model. Both mechanisms are jointly calibrated on end-to-end episode trajectories via a multi-objective Learn-Then-Test (LTT) procedure, providing simultaneous finite-sample guarantees on expected episode reward and cloud-call rate. We evaluate TSDS on four ReAct benchmarks spanning arithmetic reasoning (GSM8K), multi-hop question answering (HotpotQA), code generation (MBPP), and multi-step embodied planning (household robot), and compare against thought-calibration-only and calibrated-deferral-only standalone baselines. TSDS reduces per-episode thinking compute by 43%-73% over deferral-only baselines across HotpotQA, MBPP, and the household robot task, while maintaining certified reward and cloud-call rate guarantees.
Sources
- Do As I Can, Not As I Say: Grounding Language in Robotic Affordances
- Learn then Test: Calibrating Predictive Algorithms to Achieve Risk Control
- Program Synthesis with Large Language Models
- FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance
- Training Verifiers to Solve Math Word Problems
- Therefore I am. I Think
- Statistically Valid Post-Training Hyperparameter Selection: From Tuning to Guarantees
- Entropy Collapse in Mobile Sensors: The Hidden Risks of Sensor-Based Security
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
- Building Production-Ready Probes For Gemini
- Uncertainty Estimation in Autoregressive Structured Prediction
- RouteLLM: Learning to Route LLMs with Preference Data
- ReDAct: Uncertainty-Aware Deferral for LLM Agents
- Voyager: An Open-Ended Embodied Agent with Large Language Models
- No Free Labels: Limitations of LLM-as-a-Judge Without Human Grounding
- Thought calibration: Efficient and confident test-time scaling
- On-Device Language Models: A Comprehensive Review
- Large Language Model Cascades with Mixture of Thoughts Representations for Cost-efficient Reasoning
- Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models
Related papers
- Behavior of prediction performance metrics with rare events
- Optimal Estimation of Generic Dynamics by Path-Dependent Neural Jump ODEs
- A Posterior-Dynamics Framework for Imaging Inverse Problems with Pretrained Diffusion Priors
- Scalable Posterior Uncertainty for Flexible Density-Based Clustering
- A Transport Geometry of Benign Overfitting
- Efficient quadratic entropy with distance sketches