LLMs Can Predict Failure Risk, But Struggle to Predict Which Collaboration Protocol Pays Off: Cost-Aware Protocol Routing Across Reasoning Tasks

2026-08-14

AI Summary

This study addresses the cost-effectiveness challenge in selecting collaboration protocols for multi-agent reasoning by proposing a cost-aware routing framework. Through comparative experiments involving four protocols integrated with failure risk probes and confidence gating mechanisms, results demonstrate that while models can effectively predict failure risks with an AUROC of 0.88, their ability to identify specific protocol value remains limited, achieving an AUPRC of only 0.17. These findings reveal critical limitations in LLMs' capacity to leverage confidence scores for initial escalation decisions, indicating difficulties in accurately assessing optimal collaboration strategies. Consequently, this work provides essential empirical evidence to inform efficient resource allocation within multi-agent systems, highlighting the gap between risk prediction and strategic protocol selection in current large language model capabilities.

Abstract

Multi-agent large language model (LLM) systems can improve reasoning by spending more computation, but deployment requires deciding when extra collaboration is worth its cost. We isolate this decision by running every problem under four protocols while holding the solver fixed within each setting: direct solving (Baseline), iterative self-correction (Single), planner-executor-reviewer collaboration (PER), and multi-agent deliberation (Broadcast). The primary benchmark comprises 4,181 competition-level math problems; paired robustness checks cover four benchmarks spanning competition math, biology, and broader science with two solver families. Across fixed policies, trained routers, and frozen LLM routers, conservative policies under-escalate, whereas higher-solve frozen routers often over-escalate. A post-answer, pre-collaboration gpt-oss-120b probe ranks Baseline failures with 0.8847 AUROC (4,151 parseable cases; 95% CI [0.8732, 0.8955]). The same score remains informative for predicting whether any collaboration helps (0.7683 AUPRC), but is much weaker for identifying PER- or Broadcast-specific value (0.1674 and 0.1041 AUPRC). Separately, the pre-answer self-confidence gate reaches 78.0% solve at 45K tokens, compared with 73.8% at 71.3K for a frozen gpt-oss-120b router and 92.4% for a retrospective fixed-order oracle. Across 10 paired model-condition settings, the oracle adds 23.2-58.3 points of retrospective coverage over Baseline, but protocol profiles vary by task. In the six settings with held-out router evaluations, oracle gaps remain 18.5-28.9 points. Confidence can therefore support initial escalation, while protocol-specific cost-aware routing remains unresolved.

Problem

• Cost-Aware Protocol Routing
• Multi-agent LLM Systems
• Collaboration Protocol Selection
• Reasoning Tasks

Innovation

• Cost-Aware Protocol Routing
• Post-Answer Failure Probe
• Pre-Answer Confidence Gate
• Multi-Agent Collaboration Protocols
• Retrospective Oracle Analysis