Think Short, Defer Smart, Act, and Repeat: Calibrated Reasoning and Uncertainty-Aware Deferral for Edge LLM Agents

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Amirmohammad Farzaneh, Osvaldo Simeone
• Venue: arXiv: stat.ML
• Published: 2026-07-29
• Collected: 2026-07-30
• URL: http://arxiv.org/abs/2607.26865v1
• PDF: https://arxiv.org/pdf/2607.26865v1
• DOI: UNKNOWN

Plain-Language Brief

The paper proposes a framework called Think Short, Defer Smart (TSDS) for LLM agents that follow the ReAct paradigm. This framework helps manage reasoning budget and ensures reliability when deployed at the edge. It combines a lightweight convergence probe to halt on-device reasoning once an action stabilizes, with a perplexity-based deferral rule that escalates uncertain actions to a cloud-side model. The framework is calibrated using a multi-objective Learn-Then-Test (LTT) procedure and provides guarantees for expected episode reward and cloud-call rate. TSDS was tested on four ReAct benchmarks and showed significant reduction in per-episode thinking compute while maintaining certified reward and cloud-call rate guarantees.

What the Document Says

• The paper introduces a framework called Think Short, Defer Smart (TSDS) for LLM agents following the ReAct paradigm.
• The TSDS framework aims to manage reasoning budget and ensure reliability when deployed at the edge.
• It combines a lightweight convergence probe and a perplexity-based deferral rule.
• These mechanisms are jointly calibrated using a multi-objective Learn-Then-Test (LTT) procedure.
• The framework provides guarantees for expected episode reward and cloud-call rate.
• TSDS was tested on four ReAct benchmarks: GSM8K, HotpotQA, MBPP, and household robot.

Key Concepts Mentioned

• LLM agents
• ReAct paradigm
• Reasoning budget
• Think Short, Defer Smart (TSDS)
• Lightweight convergence probe
• Perplexity-based deferral rule
• Multi-objective Learn-Then-Test (LTT) procedure
• Expected episode reward
• Cloud-call rate

Claims Made by the Document

• TSDS reduces per-episode thinking compute by 43%-73% over deferral-only baselines across HotpotQA, MBPP, and the household robot task.

Evidence Used

• The text states that TSDS reduces per-episode thinking compute significantly when compared to deferral-only baselines on HotpotQA, MBPP, and the household robot task.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

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.