Research map / Large language models / Reasoning and chain of thought

A Communication-Theoretic Framework for LLM Agents: Cost-Aware Adaptive Reliability

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Efficient LLM serving · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Frames LLM sampling as a discrete stochastic channel and maps six communication-theoretic reliability operators (diversity combining, HARQ, turbo-style generator-critic, fountain sampling, FEC-style parity, ACM routing) onto agent techniques, plus a cost-aware semantic-kNN router with a single Lagrangian knob.

Problem

Agent reliability techniques (retry, majority voting, self-consistency, refinement, routing) are developed separately without a common analytical framework, and no fixed model-technique-budget choice dominates across tasks.

Approach

Treats an LLM sampled at temperature T as a discrete stochastic channel p(y|x) and recasts existing inference-time reliability methods as special cases of six classical communication operators; adds a semantic-kNN router whose lambda trades quality against normalized cost without retraining.

Key results

• On a 300-item hard split of MMLU/GSM8K/HumanEval, the semantic-kNN router at lambda=3 beats cross-validated fixed-best by 6.8% quality at matched normalized cost (Wilcoxon p<1e-13).
• At matched quality (~0.818), the router at lambda=38 runs at 55.7% lower normalized cost than fixed-best.
• On 14B, HARQ-IR reaches +14.0% coding gain and turbo +14.2%; Diversity-EGC beats MRC on noisier judge CSI configurations (14B and Ollama-cloud).

Surprising findings

• Equal-gain combining outperforms quality-weighted (MRC) combining on the local and Ollama-cloud configurations where judge scores are noisier.
• Self-Refine slightly beats HARQ-IR on Ollama-cloud configurations, and Self-Consistency is statistically tied with the paper's Diversity-MRC-Discrete-N at some operating points.

Limitations

• Judge-score noise from binary-checklist scorer and dual-role judge confound (judge-free decoders still win about 39% of oracle assignments).
• Curated task selection bias; 69 curated tasks and 36 benchmark tasks referenced in supplement only.
• Theoretical results rely on linear-Gaussian idealization and first-order approximations; LLM combiner is not a linear combiner, so MRC optimality is not established end-to-end.

Experimental setup

Datasets: MMLU GSM8K HumanEval curated 69-task suite

Benchmarks: MMLU GSM8K HumanEval curated 69-task suite

Metrics: quality coding gain normalized cost coding-gain efficiency Pareto frontier

Baselines: Self-Consistency Best-of-N Weighted Best-of-N CISC Mixture-of-Agents Self-Refine Chain-of-Verification RouteLLM FrugalGPT

Builds on: Self-Consistency Mixture-of-Agents Self-Refine FrugalGPT RouteLLM

Closest papers by meaning

• KVComm: Enabling Efficient LLM Communication through Selective KV Sharing (ICLR 2026 · similarity 0.88)
• PACE: Two-Timescale Self-Evolution for Small Language Model Agents (NeurIPS 2026 · similarity 0.87)
• Cost-Effective Communication: An Auction-based Method for Language Agent Interaction (AAAI 2026 · similarity 0.87)
• MESS+: Dynamically Learned Inference-Time LLM Routing in Model Zoos with Service Level Guarantees (NeurIPS 2025 · similarity 0.87)
• Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems (ICLR 2025 · similarity 0.86)
• MOC: Multi-Order Communication in LLM-based Multi-Agent Systems (ICML 2026 · similarity 0.86)