Confidence Is Not Correction: Measurement Resolution Bottlenecks in LLM-Agent Evaluation
Summary:The paper separates three claims often conflated in confidence-based LLM-agent systems: discrimination, selective prediction, and correction utility. Its central result is an exact decomposition of correction utility into gated recoveries minus gated harms. Because a replacement gate can recover only errors already made by the base policy, the initial error count caps the number of useful discordant pairs a paired test could observe. Empirically, a perturbation-agreement verifier reaches a with
Core Claim
High confidence does not imply that an LLM agent will correct errors. Rather than proposing another confidence signal, the paper shows that measurement resolution often determines whether correction gains can be detected at all. A verifier can rank correct and incorrect outputs very well while failing to replace enough proposals to justify its cost.
The work separates three claims that are frequently treated as interchangeable:
- Discrimination: whether a confidence score ranks correct and incorrect answers usefully.
- Selective prediction: whether a system can withhold low-confidence outputs through abstention, escalation, or human review.
- Correction utility: whether a triggered gate actually changes the answer in a way that improves task performance.
None of these claims automatically implies another.
An Exact Decomposition of Correction Utility
For episode i, define:
g_i ∈ {0,1}: whether the gate replaces the base-policy proposal;c_i ∈ {0,1}: whether the base proposal is correct;c'_i ∈ {0,1}: whether the final answer is correct.
Under binary correctness, the net correction utility is exactly:
Δ = Σ g_i × (c'_i - c_i) = R - H
Here, R counts gated recoveries, where an incorrect proposal becomes correct, and H counts gated harms, where a correct proposal becomes incorrect. Counting only recoveries would conceal regressions introduced by the gate. The relevant objective is therefore R-H, not raw flip count or confidence rank.
Abstention requires a separate objective. It changes coverage and conditional risk rather than repairing an incorrect answer, so it should not be folded into correction utility.
Why a Working Gate Can Fail a Paired Test
Let E₀ be the number of base-policy errors. A replacement gate can recover only those existing errors, so:
R ≤ E₀
A paired significance test requires enough discordant pairs—cases in which base and gated correctness differ. If the test-specific requirement exceeds the maximum number of recoverable errors, even an oracle gate cannot generate enough discordant evidence. Two mechanisms then become statistically indistinguishable:
- a useful gate whose effect cannot be resolved by the test; and
- a gate that never fires.
This is an identifiability problem created jointly by task difficulty, base error rate, trigger rate, and test design, not merely a property of the confidence signal.
Empirical Results and Engineering Trade-offs
The supplied abstract reports the following observations:
| Observation | Reported result | Engineering implication | |---|---:|---| | Perturbation-agreement verifier | Within-policy AUROC 0.990, 95% CI [0.95, 1.00] | Strong ranking ability, but no demonstrated replacement benefit | | Small-model suite | Identical episode outcomes to single-sample decoding at 4.32 times the call count | No measured task gain despite substantially higher Token, latency, and concurrency cost | | Controlled study | No proposal replacement across 60 GSM8K decisions in 840 episodes | The comparison did not exercise the correction mechanism | | Diagnostic set | A perfect gate could reach only four errors while the test required six | No significant positive result was available even under an oracle assumption |
Across six controllers, the entire reported GSM8K gain came from one task. Repeated temperature-zero runs on that task disagreed, indicating that the service did not reproduce the same output. Temperature zero does not by itself guarantee bitwise determinism across backends, batch shapes, parallel kernels, routing decisions, or cache states.
AUROC must also be interpreted narrowly. A value of 0.990 measures within-policy ranking, not transfer to another policy and not downstream task accuracy. It is insensitive to whether a threshold actually triggers, how tied scores are handled, and whether controllers receive equal call and output budgets. It therefore cannot establish correction utility on its own.
From an engineering perspective, resampling, verifier calls, and escalation to a larger model increase generated tokens, time to first token, tail latency, and queueing pressure. Even if cost scales approximately linearly with calls, near-zero net recoveries make the cost per corrected episode diverge. Abstention presents a different trade-off by reducing errors at the cost of coverage and must be evaluated separately.
Evaluation and Deployment Recommendations
- Use a three-metric panel. Report discrimination metrics such as AUROC, selective-prediction metrics such as risk–coverage curves and calibration error, and correction metrics such as recoveries, harms, and
R-H. Do not let one metric stand in for the others. - Run a statistical feasibility check first. Use a pilot to estimate
E₀, gate activation, and the number of discordant pairs required by the paired test. If the recoverable ceiling is too low, create more base-error headroom or redesign the task before spending the full model budget. - Instrument every trigger. Log the base answer, final answer, confidence, threshold, replacement flag, model called, Token count, latency, and correctness. Without activation logs, a correction experiment cannot distinguish a failed mechanism from an empty comparison.
- Equalize output budgets. Controllers should receive the same number of calls, maximum output Tokens, and sampling protocol, or results should be presented as an accuracy–cost frontier. Otherwise gains from extra sampling or longer generations are misattributed to confidence.
- Define tie policies explicitly. Ranking ties, threshold-boundary ties, and unequal output limits can each move a headline result. Publish tie mass and the deterministic or randomized rule used to break ties, and perform threshold-sensitivity analysis.
- Test reproducibility. Repeat temperature-zero calls for the same input and record backend versions, batch size, routing, and cache state. If base outputs vary, estimate confidence intervals by task or run cluster rather than treating correlated trajectories as independent observations.
- Gate deployment on net utility and unit cost. Replace production answers only when held-out
R-His reliably positive, the harm rate is acceptable, and total cost per net correction is within budget. A high AUROC should not operate as the deployment switch.
The available material does not identify the two evaluated models, provide the full task mix and statistical tables, specify the paired-test procedure, or disclose code and random seeds. The methodology is therefore valuable and actionable, but its gains still require replication across models, tasks, and production environments under equalized budgets.
— END —Source: ArXiv NLP/LLM (cs.CL) (2026-11-18)
Tags: #LLM Agents #Confidence Gating #Selective Prediction #Correction Utility #Paired Testing
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments