Source-linked AI summary
Risk Is Not Review Value: Wrong-Answer Exposure Under Bounded Review Budgets
SangJin Park, Myungsub Choi, Jineok Kim, Minseung Kang
TL;DR
LLM assistants generate more answers than humans can review, making prioritization under bounded budgets a question of which wrong answers remain exposed. This paper models review as exposure-reduction allocation and finds that review-value ranking lowers post-repair residual exposure while leaving WAER nearly comparable.
Problem
Bounded review budgets require deciding which generated answers to check first, because answers below the review cutoff remain exposed.
Method
The paper ranks answers by review value using wrongness, intervention affordance, impact, and cost, and evaluates queues with WAER and PRRE.
Results
Review-value ranking lowers PRRE across budgets and base rates, while WAER is comparable or better except at the sparsest 5% setting, where the difference is negligible.
Takeaways & Limitations
Review priority should account for whether captured errors have explicit, deterministic interventions, not only their estimated risk.
Takeaways & Limitations
The evaluation is limited to a controlled public-data stress benchmark with synthetic answer variants, and PRRE does not measure whether human or model reviewers would repair errors in practice.
Abstract
from arXiv · showhide
LLM assistants often produce more answers than humans can review before users see them. Most evaluations ask whether an answer is wrong, unsupported, or low-confidence. Bounded review budgets instead ask which answers should be checked first under a fixed review budget. Risk alone is not enough: a high-risk answer may be hard to repair, while a moderately risky answer may be directly correctable from available evidence. For generated-answer evaluation, we model review prioritization as exposure reduction, where review value combines estimated wrongness, intervention affordance, impact, and cost. We evaluate review queues with Wrong-Answer Exposure Ratio (WAER), the fraction of wrong answers left unreviewed, and post-repair residual exposure (PRRE), the fraction still exposed after deterministic benchmark-supported repairs. PRRE uses repairability rules that do not numerically reuse the affordance scores used for ranking. On a 720-item TAT-QA/SciFact stress benchmark, review-value ranking keeps answer-level WAER nearly unchanged at 20% budget (0.605 vs. 0.600) but lowers PRRE from 0.881 to 0.716. These results show that trustworthy LLM evaluation should measure not only error detection, but also how limited review capacity reduces exposed wrong answers.
1 Introduction
Bounded review budgets shift evaluation from asking whether answers are wrong to deciding which answers should consume scarce review capacity. The paper argues that review priority should optimize expected exposure reduction rather than risk alone.
- Under realistic workflows, humans cannot review every generated answer before it reaches users, leaving answers below the cutoff exposed.
- Risk-only ranking can prioritize a highly suspicious answer that is difficult to reconstruct over a moderately suspicious answer directly correctable from cited evidence.
- Review-value prioritization combines wrongness, intervention affordance, impact, and cost to target expected exposure reduction.
- The paper evaluates existing review-prioritization strategies with an objective and diagnostic benchmark rather than introducing a new verifier model.
- The setting evaluates residual exposure after truncating the review queue, not only score quality before truncation.
2 Bounded Review Budget Objective
The paper formalizes bounded review as selecting answers whose review is expected to reduce exposure most effectively. Its review-value queue can invert risk-only rankings when affordance, impact, and cost make a lower-risk answer more valuable to review.
- The framework ranks generated answers for review under a count budget, with only the top B answers reviewed and wrong answers outside that set remaining exposed.
- Risk-only strategies rank by wrongness risk, whereas review-value ranking incorporates intervention affordance, impact or severity, and review cost.
- A review-value queue can prefer a moderate-risk, high-affordance error over a high-risk, low-affordance error under the same risk model and budget.
- The benchmark treats intervention affordance as a public-evidence and gold-metadata property indicating whether a bounded correction is explicit and nearly unique.
- Replacing an incorrect numeric value has high direct intervention affordance, while resolving a scope distortion requires reconstruction and has lower affordance.
3 Exposure Metrics
The paper evaluates exposure after bounded review queues using WAER for unreviewed wrong answers and PRRE for residual exposure after deterministic repairs. It separates ranking affordance from outcome-side repairability and reports contextual severity-weighted metrics.
- WAER measures the fraction of wrong answers left unreviewed across bounded review budgets.Lower WAER is better, and reporting it over multiple budgets yields an exposure-capacity curve.
- PRRE measures residual exposure after deterministic benchmark-side repair without reusing numeric affordance values from ranking.Selected-but-unrepairable and unselected wrong answers remain exposed.
- Numeric errors are repairable when they deviate from a gold value in the evidence, whereas scope distortion and conclusion mismatch are not deterministically recoverable.
- Weighted exposure uses impact-weighted residual wrong answers, while review-value exposure uses the ranker’s actionable-severity term and remains contextual.PRRE is the primary repair-aware outcome; current pilots set q_i = 1.
4 Diagnostic Benchmark
The diagnostic benchmark uses public TAT-QA and SciFact items with controlled errors, matched controls, and evidence-grounded labels. Its design supports comparisons across error types, review budgets, base rates, and leakage conditions.
- The benchmark combines TAT-QA numerical table-and-text reasoning with SciFact evidence-backed scientific claim verification.
- Each source item yields correct answers, controlled wrong answers, and style-matched correct controls to reduce shortcut artifacts.Gold wrongness labels come from public labels and controlled mutation metadata rather than LLM judgments.
- The benchmark requires evidence-grounded items, matched controls, diverse error types, budget curves, base-rate sweeps, and leakage tiers.
- Controlled mutations separate wrongness, impact, and repairability while preserving public evidence, but they do not replace naturalistic evaluation.The benchmark’s priors should be validated or replaced with human-review measurements in future work.
5 Prioritization Strategies
The study compares risk-only rankings with review-value strategies that also estimate intervention affordance and impact. Baselines are separated by the information they observe, and Table 2 reports the main 20% capacity stress result.
- Risk-only baselines primarily estimate wrongness risk, while review-value proxies multiply risk by estimated intervention affordance and impact.Both operational proxies share the same source-consistency risk score.
- Table 2 reports the public-data stress result at 20% review capacity, averaged over 32 outer seeds with lower values preferred.
- Operational risk-only baselines do not observe wrongness labels, injected error type, severity, or affordance labels.
6 Diagnostic Results
Review-value ranking changes which wrong answers are exposed after review, while answer-level wrong-answer capture remains nearly tied with risk-only ranking. The improvement is concentrated in deterministic repairability and persists across review budgets, though sensitivity varies with base rates and prior shifts.
- 0.605 vs. 0.600: answer-level WAER is nearly tied between risk-only and review-value ranking at 20% review capacity.The 720-item benchmark contains 360 wrong answers and uses 144 reviews at this budget.
- 0.881 to 0.716: review-value ranking lowers PRRE because it prioritizes errors recoverable from public evidence.The paired analysis reports a small WAER difference while PRRE and RVE decrease under review-value ranking.
- WAER .600: holding verifier risk fixed leaves answer-level exposure unchanged, while intervention-aware ranking lowers PRRE from .767 to .600.This ablation attributes the gain to queue allocation rather than a stronger verifier.
- WAER .067 for numeric perturbations versus 1.000 for scope distortions: allocation effects differ sharply by error type.Conclusion mismatches can be reviewed but receive no deterministic repair credit under PRRE.
- Lower PRRE holds across tested budgets, reaching the repair-count lower bound of .556 at 40% capacity in the balanced stress setting.The benchmark's review-value proxy covers the recoverable errors at that budget.
- PRRE improves at every tested wrong-answer base rate, but larger prior shifts weaken and do not universally preserve the advantage.Wins under random perturbations fall from 100.0% at ±.1 to 84.3% at ±.3.
7 Related Work
Related work studies selective prediction, uncertainty and factuality evaluation, human routing, and post-generation correction. This paper instead evaluates post-generation review policies under bounded review budgets.
- Selective classification and abstention study trading coverage against prediction risk, while selective QA and generation study when not to answer.
- Calibration and uncertainty methods estimate confidence, failure, or factuality signals for generated outputs.
- Human-assistance, algorithmic-triage, and learning-to-defer work studies routing model predictions to human or expert judgment.
- This setting differs by evaluating post-generation review policies rather than selective answering or a better detector.
- Factuality, attribution, retrieval-grounded, and LLM-based evaluation work measures truthfulness, evidence support, citation grounding, judgment alignment, or judge reliability.
8 Discussion and Limitations
The results are limited to a controlled public-data stress benchmark and isolate queue-objective effects while holding risk scores fixed. Synthetic variants, deterministic recoverability rules, and benchmark-assumed type priors constrain how directly the findings transfer to production review.
- The evaluation is limited to a controlled public-data stress benchmark and does not establish that risk-only ranking must fail.
- The same-risk ablations hold the risk score fixed while showing that changing the queue objective changes which wrong answers remain below the review cutoff.
- The benchmark uses synthetic answer variants and deterministic recoverability rules, not measurements of whether human or model reviewers would repair each error in practice.
- The intended scope is separating risk scoring from queue allocation under bounded review budgets, not estimating production prevalence.
9 Conclusion
The paper frames evaluation under bounded review budgets as exposure-reduction allocation rather than simple risk detection. Its stress-benchmark results show that review-value ranking can preserve near-equivalent wrong-answer capture while reducing residual repair exposure under the benchmark’s actionability assumptions.
- The paper proposes evaluating LLM outputs under bounded review budgets as an exposure-reduction allocation problem.
- Review priority depends on wrongness, intervention affordance, impact, and cost rather than risk alone.
- At 20% budget, risk-only and review-value ranking are nearly tied on WAER, while review-value ranking leaves lower PRRE across budgets and base rates.
Policy 5% 10% 20% 40%
Across the budget ladder, review-value ranking often matches risk-only WAER while reducing post-repair residual exposure. The largest mechanism difference appears at small budgets, where equal wrong-answer capture can produce sharply different repair outcomes.
- At 20% budget, risk-only and review-value ranking report WAER .605 and .600, while PRRE is .881 and .716, respectively.
- At 5% capacity, risk-only and review-value ranking have identical WAER (.900), but PRRE is .989 versus .903, respectively.
- At 40% capacity, review-value ranking reaches the repair-count lower bound of .556.
- The Gold y_i h_i reference reaches WAER .200 at 40% capacity but retains PRRE .756 because selection is not conditioned on deterministic recoverability.
- Paired bootstrap intervals favor review-value ranking on PRRE and RVE, while its WAER advantage is operationally small at the .600 floor.
B LLM-Assisted Pilot
The LLM-assisted pilot tests whether the bounded-review framework remains informative when risk scores come from a model-derived verifier. Results support the queue-objective diagnosis, but the pilot remains auxiliary and its outcomes vary by error type and prior assumptions.
- Pilot design: The pilot uses an LLM-assisted verifier for risk scoring while retaining benchmark gold wrongness and outcome-side repair rules.
- Queue-objective ablation: At 20% capacity, multiplying the same risk score by affordance lowers PRRE from .750 to .640 while WAER changes from .640 to .630.
- Queue-objective ablation: At 20% capacity, adding the impact prior produces the lowest WAER and WDE but a higher PRRE of .720.
- Error-type breakdown: All numeric perturbations are selected, whereas all scope distortions remain exposed; direction flips have WAER .824.
- Repair checks: In the repair pilot, benchmark-side success is .95 for direction flips and numeric perturbations, .73 for conclusion mismatches, and .30 for unsupported additions.
- Base-rate sensitivity: At wrong-answer rates from 10% through 30%, review-value ranking lowers both WAER and PRRE, while at 5% WAER is .155 versus .156 and PRRE is .632 versus .551.
- Prior sensitivity: A fixed positive prior shift of .2 makes review-value PRRE .888 versus risk-only .881, showing the advantage is not invariant to arbitrary priors.
- Prior sensitivity: Across random prior perturbations, PRRE favors review-value ranking in 100.0%, 94.3%, and 84.3% of trials at radii .1, .2, and .3.