AI systems increasingly assist decision making by producing cheap factor-level assessments of complex inputs, but these assessments can be biased or incomplete. We study selective human evaluation: given AI signals, which factors should be dispatched to human experts? We formulate this human--AI collaboration problem as a factor-level information-acquisition problem. Under squared loss, the optimal dispatching rule maximizes a contextual reward; under a linear model, this reward admits a closed-form decomposition into two interpretable terms: predictive relevance and residual uncertainty in the human evaluation after conditioning on the AI signal. We instantiate the framework on peer review, decomposing papers into ten aspects and evaluating up to 3,408 ICLR submissions across three LLMs and multiple regression heads. In this retrospective proxy, where human aspect signals are extracted from existing reviews, predictions that query three of the ten aspects under the optimal dispatching rules match the full ten-aspect human benchmark. The finding holds when the human signal comes from a single noisy reviewer, and across review years. These results position principled dispatching as a practical building block for scalable human--AI decision systems. We also discuss safeguards under which such a rule can serve as decision support that leaves the final decision with a human.
Figures & tables
Granularity
AI-cond.
Mechanism
Combines A,H
Budget
Mozannar and Sontag (2020)
instance
✓
trained classifier + rejector
no
no
Wilder et al. (2020)
instance
✓
joint training
no
no
Koh et al. (2020)
all factors
—
no learned selector
yes
no
Pugnana et al. (2025)
factor subset
✓
per-concept rejector, L2D surrogate
yes
implicit
Ours
factor subset
✓
plug-in reward / closed-form rule
yes
explicit
Table 1. Positioning against the closest human-AI collaboration frameworks. “Granularity” is the unit at which the human is queried; “AI-cond.” indicates whether selection adapts to the AI output; “Combines A,H ” indicates whether the final predictor uses both the AI and the (partial) human signal.
Figure 1. Test MAE versus number of queried aspects under the Average aggregation, for ridge (top) and gradient boosting (bottom) across three LLMs. Rule numbers follow Section 4.4 . Shaded regions are 95% confidence intervals over 20 random splits. Dashed lines mark the non-selection baselines and benchmark: AI only, human only, and combined. The linear rules (3)–(4) and the fixed subset (7) nearly coincide and approach the full-human benchmark with 2 – 3 queried aspects; importance-only (5), learned agreement (6), and random- k (8) need far more.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Target
Function class
Dimension
Norm
Rate
mS(A,HS)
Linear / GLM
fixed
L4
n−1/2
mS(A,HS)
Sparse linear
p
L4
nslogp
mS(A,HS)
Hölder Cs
dX
L4
n−2s+dXs
mS(A,HS)
RKHS
eff. dim.
L4
n−2α+1α
μ(Z)
Linear / GLM
dZ
L2
N−1/2
μ(Z)
Sparse linear
dZ
L2
NslogdZ
Appendix
Table 2. Representative regression rates for Assumptions 4 – 5
NeurIPS 2026 Criterion
Our Aspects (sub-dimensions)
Quality
Claims and Evidence Rigor; Experimental Design and Robustness; Reproducibility and Transparency; Limitations
Clarity
Clarity and Writing
Significance
Significance and Impact; Future Work Potential
Originality
Originality and Novelty; Related Work and Positioning
Broader Impact (separate solicitation)
Ethics and Societal Impact
Appendix
Table 3. High-level mapping between the four NeurIPS 2026 reviewer criteria (plus the separately solicited Broader Impact statement) and our ten aspects. Each aspect is a strict sub-dimension of exactly one criterion, ensuring comprehensiveness and disjointness. Finer granularity lets the selection rule dispatch human effort to exactly the sub-dimension where the AI signal is least reliable, rather than forcing an all-or-nothing query at the coarser criterion level.
Figure 2. Sentence-level overlay of the NeurIPS 2026 General Reviewing Guidelines onto our ten aspects. The text reproduces the official NeurIPS prompts verbatim, with each sub-question rendered in the color of the single aspect it activates; the bottom panel gives the aspect-to-color legend. Every NeurIPS sub-question maps onto exactly one aspect, demonstrating that the ten-way decomposition is a strict, sentence-level refinement of the official NeurIPS reviewer prompts. Ethics and Societal Impact does not appear inside the four general criteria because it is solicited separately by NeurIPS through the Broader Impact statement.
Aspect
Definition and Prompt
Claims Evidence Rigor
Definition : Correctness of proofs and adequacy of empirical support for each stated claim. Prompt : List the paper’s main claims and judge whether each is supported by (a) correct proofs (cite lemmas you checked) and (b) adequate empirical evidence (baselines, ablations, statistics); note any mismatches between claims and evidence.
Clarity Writing
Definition : Readability, organization, clear definitions, consistent notation, and comprehensible experimental descriptions. Prompt : Is the exposition easy to follow? For theory, are all terms/assumptions defined and symbols consistent? For experiments, are dataset/task definitions, metrics, hyperparameters, seeds, and training details specified so an expert could reproduce key results?
Significance Impact
Definition : Importance of the problem and expected influence on research or practice. Prompt : Who would use these ideas, and how would they change research or practice? Does the work tackle an important task better than prior art or materially advance understanding?
Originality Novelty
Definition : Newness of insights, analysis, tasks, methods, or well-reasoned combinations, and clarity versus prior work. Prompt : What’s genuinely new (insights, analysis, task, method, or a well-reasoned combination), and is the difference vs. prior work explicit and well-cited?
Related Work Positioning
Definition : Accuracy and completeness of positioning relative to essential prior and concurrent work. Prompt : Does the paper correctly place itself in context, citing essential prior or concurrent work and explaining concrete differences to avoid overstating novelty?
Reproducibility Transparency
Definition : Availability and sufficiency of artifacts and procedural detail to reproduce key results. Prompt : Are artifacts/details sufficient to reproduce the main results (configs, code/data access or alternatives, hardware/compute, preprocessing)? If some assets can’t be shared, do authors provide a viable path for verification?
Appendix
Table 4. List of aspects used to extract features for predicting the average reviewer score of a paper.
Metrics: MAE ( ↓ ), RMSE ( ↓ ), R2 ( ↑ )
Regression Model /
βj,⊥
Rj(A)
Rj (A-independent)
RSlin (Optimal)
Rjlin (Singleton)
Language Models
MAE
RMSE
R2
MAE
RMSE
R2
MAE
RMSE
R2
MAE
RMSE
R2
MAE
RMSE
R2
GB / GPT-5-mini
0.596
0.749
0.515
0.544
0.684
0.592
0.524
0.660
0.619
0.523
0.658
0.621
0.524
0.660
0.619
Rand. Forest
0.610
0.766
0.493
0.550
0.693
0.582
0.533
0.671
0.606
0.533
0.670
0.607
0.533
0.671
0.606
MLP
0.601
0.757
0.504
0.554
0.699
0.574
0.528
0.666
0.611
0.526
0.664
0.614
0.528
0.666
0.611
XGBoost
0.598
0.752
0.511
0.547
0.688
0.587
0.529
0.666
0.611
0.528
0.665
0.612
0.529
0.666
0.611
Appendix
Table 5. Average performance of different selection rules, averaged over all 20 train-test splits and over nsel≤5 . Each horizontal section corresponds to a different LLM, with different regression models. For each metric and row, the top method (to four decimal places) is bolded. “GB” = gradient boosting; “Gemini-2.0” is Gemini 2.0 Flash. The Rj(A) column uses the estimator without the centering of Remark 1 ; the centered estimator is evaluated for ridge and gradient boosting in Appendix D.8 and Table 9 .
Code
Logic reason
ClaimsEvidenceRigor (all codes prefixed by rigor_ )
supported_theory_and_empirics
Each primary claim is backed by correct proofs and adequate empirical evidence.
theory_ok_empirics_incomplete
Theoretical proofs correct for the stated assumptions, empirical evidence incomplete.
empirics_strong_no_theory
Empirical evidence strong (baselines, statistics), but theoretical support is absent.
proof_gaps_undermine_claims
Key proofs contain gaps or unverified steps that undermine the corresponding claims.
claims_overstated_vs_results
Claim wording overstates results relative to what the experiments actually demonstrate.
Appendix
Table 8
Symbol
Meaning
W
complex input (e.g., a manuscript)
Y
outcome to be predicted (average reviewer score in Section 4 )
X=(X1,…,XJ)
pre-specified decision-relevant factors (aspects)
A=(A1,…,AJ)
AI signal on every factor; always observed
H=(H1,…,HJ)
human signal on every factor; costly, observed only on queried factors
Ej
free-text AI rationale for factor j (not used for selection or prediction)
Appendix
Table 7. Notation.
No.
Name
Type
Criterion
1
Rj(A)
nonparametric, adaptive
two-stage estimate of R{j}(A) , centered ( Remark 1 )
2
Rj
nonparametric, non-adaptive
average pseudo-outcome ( 14 )
3
RSlin
linear, subset
closed form ( 4 ), all subsets of size nsel
4
Rjlin
linear, singleton
∥βj,⊥∥2Var(Hj,⊥)
5
∥βj,⊥∥
ablation
relevance only
6
Learned agreement
heuristic
predicted ∣Aj,0−Hj,1∣ (Appendix C.3 )
Appendix
Table 8. Selection rules, with the single name used in every figure and table.
Figure 3. Same as Figure 1 ( Section 4.6 ) under the Random aggregation, where H is taken from a single randomly chosen reviewer. The human-only baseline degrades, while gains from combining A with selectively queried H become larger.
Figure 4. Cross-year generalization: training on ICLR 2019 and testing on ICLR 2021 (10 seeds). Rule numbers follow Section 4.4 .
Figure 5. Mean linear singleton reward Rjlin for each aspect and LLM, with error bars. Claims and Evidence Rigor dominates, followed by Originality and Novelty and Significance and Impact .
Figure 6. Simulation: test MSE versus training-set size for the non-adaptive linear rule, the stratified linear rule, and the two-stage adaptive rule without and with centering. Dashed: oracle adaptive rule with the trained predictor; dotted: AI only and full human signal. The vertical line marks the size of the peer-review corpus ( n=3,408 ; each training split contains 2,727 papers).
Figure 7. Prompt for collecting LLM reviews
Figure 8. Prompt for tagging the human reviews to generate human features H .
Figure 9. Prompt for generating agreement scores between the LLM and human reviews.
Wherever a coding agent works under engineer supervision, or a clinical model assists a radiologist, the deployment question is whether to keep the human-AI workflow or replace it with the human alone or the agent alone. The human-AI workflow is worth keeping only if it beats both of those alternatives. Yet once it is deployed, neither alternative outcome is observed: recovering one means replaying the task under that alternative, and every replay costs expert time or compute. Under a fixed replay budget, the design question is therefore which tasks should be more likely to receive a human-only replay, and which an agent-only replay. Existing methods do not directly target this decision. Agent benchmarks do not choose which missing baseline to measure, variance-based sampling ignores which of the two comparisons is closer to failing, and Bayesian information methods focus on learning model parameters instead of making the deployment decision. We propose TEAM-Design, a rule that gives every task two replay probabilities, one per baseline. It raises a probability where the missing baseline outcome is hard to predict from what is already known about the task and where that comparison is harder to establish, and lowers it where replay is expensive. We prove that the rule solves this budgeted design problem, and that drawing the replays at random from recorded probabilities still controls the chance of wrongly declaring that the workflow beats both. We reanalyze 6 clinical settings, where no human-AI workflow beats both alternatives, and a coding benchmark, where one does, then evaluate TEAM-Design on synthetic designs and on a semi-synthetic design built from a real chest X-ray reader study. TEAM-Design works best when one of the two comparisons is clearly harder to settle than the other, and can do worse than variance-based allocation when the two are similarly difficult.
Hamed Khosravi, Xiaoming Huo
H. Milton Stewart School of Industrial and Systems Engineering Georgia Institute of Technology Atlanta, GA 30332
Artificial intelligence (AI) systems are deployed as collaborators in human decision-making. Yet, evaluation practices focus primarily on model accuracy rather than whether human-AI teams are prepared to collaborate safely and effectively. Empirical evidence shows that many failures arise from miscalibrated reliance, including overuse when AI is wrong and underuse when it is helpful. This paper proposes a measurement framework for evaluating human-AI decision-making centered on team readiness. We introduce a four part taxonomy of evaluation metrics spanning outcomes, reliance behavior, safety signals, and learning over time, and connect these metrics to the Understand-Control-Improve (U-C-I) lifecycle of human-AI onboarding and collaboration. By operationalizing evaluation through interaction traces rather than model properties or self-reported trust, our framework enables deployment-relevant assessment of calibration, error recovery, and governance. We aim to support more comparable benchmarks and cumulative research on human-AI readiness, advancing safer and more accountable human-AI collaboration.
In human-AI decision making, designing AI that complements human expertise has been a natural strategy to enhance human-AI collaboration, yet it often comes at the cost of decreased AI performance in areas of human strengths. This can inadvertently erode human trust and cause them to ignore AI advice precisely when it is most needed. Conversely, an aligned AI fosters trust yet risks reinforcing suboptimal human behavior and lowering human-AI team performance. In this paper, we start by identifying this fundamental tension between performance-boosting (i.e., complementarity) and trust-building (i.e., alignment) as an inherent limitation of the traditional approach for training a single AI model to assist human decision making. To overcome this, we introduce a novel human-centered adaptive AI ensemble that strategically toggles between two specialist AI models - the aligned model and the complementary model - based on contextual cues, using an elegantly simple yet provably near-optimal Rational Routing Shortcut mechanism. Comprehensive theoretical analyses elucidate why the adaptive AI ensemble is effective and when it yields maximum benefits. Moreover, experiments on both simulated and real-world data show that when humans are assisted by the adaptive AI ensemble in decision making, they can achieve significantly higher performance than when they are assisted by single AI models that are trained to either optimize for their independent performance or even the human-AI team performance.