acceptodds
Under review as a conference paper at ICLR 2027

JointValid: Prediction-Powered Multi-Risk Certification of Large Language Models

Abstract

Large language models are evaluated across heterogeneous capabilities and risks, yet multi-dimensional scorecards rarely attach an error rate to the deployment claim that every required risk lies below its threshold. We formulate this decision as multi-risk certification with abundant scores from imperfect black-box judges and a shared human-annotation budget. Within each task, prediction-powered inference uses human scores to correct a larger judge-scored sample. Across tasks, an intersection–union test asymptotically controls erroneous global certification, with an explicit normal-approximation remainder and without multiplicity correction, cross-task independence, or calibrated judges. We derive a certification-aware allocation from a finite-judge-pool approximation to joint power. Under safe alternatives, the objective is strictly concave, and a scalar Lagrange dual yields its global solution. An independent pilot estimates task margins and residual uncertainty; a confirmatory reserve prevents task starvation, while a fractional-count Beta working posterior stabilizes allocation with limited pilot data. Experiments on three human-rated suites scored by open-weight judges show when adaptive allocation improves power over uniform allocation and when limited task heterogeneity or pilot information constrains its gains.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM evaluation, LLM-as-a-judge, multi-risk certification, statistical hypothesis testing

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_9oHMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies multi-risk certification under a limited human annotation budget, aiming to determine whether a model’s risks are below their respective thresholds across all prespecified risk dimensions. The authors propose JointValid, which combines PPI++ risk estimation with an intersection–union test to control the global false-certification rate. The paper provides theoretical support for the allocation algorithm. Simulations and experiments on real-world datasets evaluate false-certification rates, statistical power, and allocation stability, providing empirical evidence of the method’s effectiveness.

Critical strengths
  1. The paper addresses a practically relevant problem: certifying that a model satisfies multiple risk constraints under a limited human annotation budget. By accounting for differences in certification difficulty across tasks, the proposed framework allocates annotations and improve the efficiency of model evaluation.
Critical weaknesses
  1. The experiments provide limited evidence for the generality of the allocation gains. Improvements are substantial mainly on BeaverTails and modest on the other two datasets.

  2. The motivation for task-wise budget allocation is not convincing. Labels for multiple risk tasks are often collected jointly for the same response. Also, once an annotator has read a response, providing labels for nn tasks does not necessarily cost nn times as much as labeling a single task. Therefore, the assumed additive per-task cost model may not reflect actual annotation workflows, and reductions in labels needed may not translate into comparable savings in human effort or cost.

Questions for authors
  1. The clarity of the presentation could be improved. The paper introduces risk estimation, pilot-guided allocation, and final certification without a clear overview of how these components fit together. A workflow diagram showing the overall framework and distinguishing the roles of pilot and confirmatory data would make the method easier to follow.

  2. Could the authors report the total human annotation budget required to reach the same target power at fixed risk thresholds by the proposed method and uniform allocation?

  3. The comparison with the plug-in baseline does not appear entirely fair, as the posterior method also changes the risk and variance estimates. Could the authors provide a comparison using matched estimation procedures to isolate the contribution of posterior averaging?

Reviewer_DHq9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method for using limited human supervision for certifying that a model satisfies multiple evaluation constraints simultaneously, leveraging prediction-powered inference to combine a set of human labels with a large set of labels from an automated judge. They prove first that this base extension is valid, keeping certification probability close to δ\delta when the null is true (some constraint violated). Then they describe the machinery for maximizing power through allocating a constrained budget over these different tasks/constraints, using a pre-allocated portion of budget to estimate effect sizes and variances. Experiments are run on synthetic and real data, and demonstrate appropriate Type-I error control.

Critical strengths
  • the objective is compelling: the problem of satisfying multiple constraints with limited supervision seems important
  • base theoretical result (Lemma 1/Theorem 2) seems useful: good to know that we can extend PPI to this setting fairly straightforwardly
Critical weaknesses
  • experiments are weak: 1) lack of clarity around what is currently shown. It is hard for the reader to understand what the plots are supposed to signify (is the plug-in also a contribution? is the proposed method beating the plugin? what is B and why do we use diff values in 1a and 1c? what do the operating curves show and why do we need 12 in Fig 3? what is least-favorable boundary?) and the text does not help in explanation/interpretation. 2) results are limited - it would be good to show that the power maximization does something in practice. What would it look like to compare this method to other methods we might try? Additionally would be good to see a systematic exploration of power vs budget in real data
  • the paper is somewhat challenging to read (although not impossible). I would encourage the authors to go through carefully and ensure that any lingo which is used is defined clearly - terms such as pilot, reserve, the Berry-Esseen remainder, and shadow price are used without explanation, as well as Φ\Phi in Eq 16

small note: the certification bound in Eq 12 states P \leq delta but then the actual result seems to be delta + a positive term - it would be good to figure out this contradiction

Questions for authors
  • how does power of this method compare in practice to power of other comparable methods?
  • what do the operating curve plots show me?
  • is n_k the same thing as h_k? if so, then does O(n_k ^ (1/2)) in the result indicate that we need to be using large n_k for all k to keep this tight to delta? how does that interact with the given allocation strategies?
Reviewer_aZqbOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

(I’m not sure if I understood correctly, the paper is very hard to read.) The paper focuses on the problem of evaluating outputs with regards to many different metrics (I believe that’s what the paper calls “tasks”) when you have cheap proxy annotations readily available and few ground-truth human annotations. The paper wishes to achieve good guarantees for each of the respective tasks (through a threshold for certification) while finding ways to allocate the budget of human annotations.

Critical strengths

[S1] Based on the title and whatever I could deduce from the writeup, I think the problem that the paper originally wanted to consider is interesting and can indeed have good applications.

Critical weaknesses

[W1] The abstract of the paper is humanly impenetrable. Starting from the sentence “Within each team…” it seems that the rest of the abstract is written completely by AI and not touched by a human. I do not understand what any of these sentences mean. The declaration at the bottom says “No text was accepted without author review” but I’m finding this very hard to believe given that I know the area of the paper (very) well, and I still could not understand the abstract.

[W2] Continuing right after the abstract, the introduction of the paper has a lot of AI-generated phrases, with not much logical connection. I believe that, for example, the claims made in paragraph 1 are pretty straightforward to make, but they are instead phrased in such a weird wording.

[W3] Paragraph 4 of the intro starts talking about “tasks”, but “tasks” haven’t been defined at this point. What are they? How can the reader deduce that and follow the intro? The rest of the intro is similarly impenetrable by a human reader.

[W4] The paper is unclear on so many levels. Let’s start with the problem formulation:

  • The language itself isn’t flowing — anyone can read it and deduce that it’s just one-shot AI.
  • There’s a reference to a “power model” (Line 112) without explaining at all what it is.
  • What is the “pre-specified transformation to [0,1]” that the paper talks about in Lines 108-109.
  • All goes downhill from there.

[W5] Important missing citation (hurts novelty): “Multiple Prediction-Powered Inferences”, Cowen-Breen et al 2026. This paper optimizes resource allocation across costly measurements in a PPI setting. If I understand correctly, the present paper’s estimator fits within the MultiPPI estimator class, so the difference between the two papers lies in the exactly allocation problem and certification objective. Given how hard I found the paper to be, I could not assess whether they are meaningfully different. There have also been a few other papers thinking about the concept of “tasks” and how they related to PPI, see e.g., “Prediction-Powered Inference Across Many Tasks for AI Evaluation & Social Science Research” by Emmenegger et al. 2026 and “Valid Inference with Synthetic Data via Task Exchangeability” by Tan and Zrnic 2026.

Questions for authors

I have no questions. I am sorry to say that the current writeup makes it impossible to fix within the scope of a conference review timeline.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.