Skip to content
KoishiAI
ไทย
← Back to all articles
ai-agents llm open-source ai-research

Proxy Confidence: Auditing Black-Box LLM Agents Efficiently

Proxy Confidence audits black-box LLM agents using surrogate models. Achieve higher accuracy than self-reporting with this open-source auditing method.

KoishiAI · Editor: เกียรติดำรง ตรีครุธพันธ์ · · 3 min read
AI-drafted from cited sources, fact-checked against them by an automated pipeline, and read by a human editor for clarity. How we work · Standards · Report an error
Colorful abstract 3D rendering showcasing AI and deep learning technology.
Photo by Google DeepMind on Pexels

TL;DR: Proxy Confidence audits black-box LLM agents using a surrogate model, achieving an AUROC of 0.825 on coding tasks compared to the actor’s near-chance self-reporting. This training-free method lifts task success rates by up to +0.137 via real-time gating.

Key facts

  • Proxy Confidence audits black-box LLM agents using a parallel surrogate model, achieving an AUROC of 0.825 on coding tasks.
  • The actor’s self-reported confidence barely beats chance at flagging mistakes, with an AUROC of only 0.598.
  • Modern frontier models are highly repetitive, causing resampling methods to fail by producing identical outputs across samples.
  • The method is training-free and incurs only one prefill pass cost, using techniques like teacher forcing and request-PMI.
  • Deployment modes such as real-time gating lifted task success rates on live-execution benchmarks by up to +0.137.

How does Proxy Confidence detect errors that self-reporting misses?

Proxy Confidence detects silent errors by running a low-cost open-weight model alongside the main agent to score actions using log-probabilities [1][2]. Frontier chat APIs hide internal token probabilities, making standard confidence metrics unreliable; an agent’s self-reported confidence barely beats chance at flagging its own mistakes (AUROC of 0.598) [1][4]. Additionally, resampling fails because modern frontier models are highly repetitive, producing identical outputs across samples [2][4]. The Proxy Confidence method bypasses these issues by using a surrogate model to score tool calls via teacher forcing, request-PMI (pointwise mutual information), discriminative verdicts, and tool-choice competition [1][4]. This approach is training-free and incurs only one prefill pass cost [3]. On difficult coding tasks, the proxy achieved an AUROC of 0.825, significantly outperforming the actor’s stated confidence [4]. It also enabled deployment modes like real-time gating and confidence feedback, which lifted task success rates on live-execution benchmarks by up to +0.137 [2].

What are the limitations and alternative auditing frameworks?

While Proxy Confidence offers a direct scoring mechanism, other frameworks address similar problems through different structures. Confidence Reasoning Graphs (CRGs) estimate success probability by decomposing an agent’s trajectory into contextualized sub-claims rather than relying on holistic judgments [6]. This decomposition avoids misleading calibration where models appear well-calibrated yet provide near-chance discrimination [6]. In contrast, BrowseConf research found that open-source agentic models show higher accuracy at high verbalized confidence levels, though existing research largely focuses on single-turn scenarios rather than complex multi-turn interactions [5]. Broader surveys highlight that uncertainty quantification (UQ) for LLMs must shift from simple question-answering to handling interactive agent dynamics, including selection of estimators and modeling uncertainty in heterogeneous entities [7]. While other work explores black-box auditing via simulated human personas for platform governance, the technical focus here remains on internal agent reliability and error detection [8].

How does this method compare to existing confidence checks?

Proxy Confidence differs from previous methods by providing an external audit layer without requiring access to the primary model’s weights. The surrogate model runs in parallel, reading the same context and proposed actions to score them via log-probabilities using complementary readouts such as teacher forcing, request-PMI (pointwise mutual information), discriminative verdicts, and tool-choice competition [1][4]. This training-free approach incurs only one prefill pass cost [3]. On difficult coding tasks, this proxy achieved an AUROC of 0.825, significantly outperforming the actor’s stated confidence (near chance at 0.598) [4]. It also enabled deployment modes like real-time gating and confidence feedback, which lifted task success rates on live-execution benchmarks by up to +0.137 [2]. Alternative frameworks like Confidence Reasoning Graphs decompose trajectories into sub-claims to avoid misleading holistic calibration [6], while BrowseConf highlights that open-source models show higher accuracy at high confidence levels but often focus on single-turn scenarios [5]. Uncertainty quantification for LLMs is shifting from simple question-answering to handling interactive agent dynamics with real-world consequences [7].

Sources

  1. Proxy Confidence: Auditing Black-Box LLM Agents with a Surrogate’s Log-Probabilities (arxiv.org) — 2026-10-02
  2. Proxy Confidence: Auditing Black-Box LLM Agents with a Surrogate’s Log-Probabilities (papers.cool) — 2026-10-02
  3. Proxy Confidence: Auditing Black-Box LLM Agents with a Surrogate’s Log-Probabilities (arxiv.org)
  4. Proxy Confidence: Auditing Black-Box LLM Agents with a Surrogate’s Log-Probabilities | Aditya Tomar (www.linkedin.com) — 2026-10-08
  5. Confidence Reasoning Graphs: Structured Confidence Estimation for LLM Agents (arxiv.org) — 2026-10-06
  6. BrowseConf: Confidence-Guided Test-Time Scaling for Web Agents (arxiv.org)
  7. Uncertainty Quantification in LLM Agents: Foundations, Emerging Challenges, and Opportunities (arxiv.org) — 2025-12-11
  8. Black-Box Auditing with LLM Simulated Agents (genai.mit.edu) — 2025-04-01

Frequently asked questions

How does Proxy Confidence detect errors that standard self-reporting misses? ▼
Proxy Confidence detects silent errors by running a low-cost open-weight surrogate model in parallel with the main agent. This surrogate scores actions using log-probabilities via methods like teacher forcing and request-PMI, bypassing the unreliability of self-reported confidence which barely beats chance.
What is the accuracy of Proxy Confidence compared to self-reported confidence? ▼
The method achieves an AUROC of 0.825 on difficult coding tasks by using a training-free surrogate model to audit the primary agent's actions. This significantly outperforms the actor's own confidence reporting, which achieved only near-chance performance with an AUROC of 0.598.
What is the computational cost of running Proxy Confidence? ▼
Proxy Confidence incurs only one prefill pass cost because it uses a training-free approach where a surrogate model reads the same context and proposed actions as the main agent. It scores these actions via log-probabilities without requiring access to the primary model's internal weights or parameters.
How does Proxy Confidence improve real-world agent performance? ▼
Proxy Confidence enables deployment modes such as real-time gating and confidence feedback, which can lift task success rates on live-execution benchmarks by up to +0.137. This allows for more reliable agent behavior by filtering out low-confidence actions before execution.
How does Proxy Confidence compare to other auditing frameworks like CRGs or BrowseConf? ▼
Proxy Confidence differs from methods like Confidence Reasoning Graphs, which decompose trajectories into sub-claims to avoid misleading holistic calibration. Unlike BrowseConf, which focuses on single-turn scenarios, Proxy Confidence is designed for complex multi-turn interactions and provides an external audit layer without needing the primary model's weights.