acceptodds
Withdrawn from ICLR 2027

VStress: Correlation-Aware Auditing and Adaptive Budget Allocation for Repeated Verifiers

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Repeated verifier calls are useful only when they contribute additional conditional information. We introduce VStress, an auditable replay contract, and VStress-CA, a correlation-aware allocation policy that estimates the conditional marginal information of an unqueried verifier on a sealed calibration split, discounts uncertainty, normalizes by call cost, and stops or abstains when the next query is not informative. The controller freezes its decision and cost ledger before accessing the clean oracle; a dependence-shift alarm disables channel preference and falls back to exact-stop. The controlled audit establishes the mechanism boundary: under 35% symmetric corruption, majority-5 improves balanced accuracy from 0.6578 to 0.7739, while at 65% corruption it decreases by 0.1226 points. In the fixed-budget comparison, breadth, redundancy, and adaptive allocation achieve balanced accuracies of 0.6048, 0.6375, and 0.6538, respectively, with VStress-CA reaching an RLVR score of 0.6417 using 3.4216 calls per item. Dependence diagnostics further show increasing conditional marginal gains when moving from same-model repeats to cross-family verifier channels, with gains of 0.0126, 0.0462, and 0.0913. These results transform verifier correlation from a post-hoc reliability concern into an auditable allocation signal for adaptive verification.

Primary area:
reinforcement learning
Keywords:
Reinforcement Learning, Verifier Aggregation, Binary Verification, Correlated Corruption, Robust Evaluation, Risk Certification, Selective Prediction, Reliable AI Systems

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_7Uf7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes VStress-CA, a stress-testing and certification framework that evaluates binary verifier aggregation under correlated corruption with reliable risk guarantees.

Critical strengths

I don't find any strength.

Critical weaknesses

The novelty is almost entirely preempted. The paper makes numerous overclaims. Its core "discovery"—that correlated errors invalidate majority voting—was already independently reported in at least three other papers by 2026.

CGES (Confidence-Guided Early Stopping for Efficient and Accurate Self-Consistency) covers the stopping/abstention rules proposed in VStress-CA.

"Adaptive-K via design effect" corresponds to the sequential sampling approach touted as a key contribution of VStress-CA.

"Convergent Adaptive Inference" addresses the thresholding mechanism used in VStress-CA.

Kohli, Nine Judges, Two Effective Votes: Correlated Errors Undermine LLM Evaluation Panels covers the very conclusion reached by VStress-CA.

Blind to the Pivotal Vote

Evaluator Ensembles Under Reward Hacking: Covariance Geometry

FUSE: Ensembling Verifiers with Zero Labeled Data

SEVRA: Think Again or Think Longer? Selective Verification for Budget-Aware Reasoning

Adaptive Test-Time Compute Allocation via Learned Heuristics

Denisov-Blanch et al., Truthfulness Does Not Scale Like Reasoning: Why Polling Fails as a Proxy Verifier

When Self-Consistency Backfires

Consensus is Not Verification

these paper should be discussed.

The bibliography contains many references completely unrelated to the topic—likely hallucinated by generative AI—along with numerous citation errors. Meanwhile, the paper fails to discuss or compare itself against a wealth of genuinely relevant literature. Furthermore, the formatting does not adhere to ICLR requirements.

The experimental evidence fails to support the paper's claims (the experiments are of very poor quality). This submission warrants a desk rejection.

Questions for authors

see above

Reviewer_uRB5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

VStress addresses the problem of imperfect reward signal in RLVR. In particular, repeated verifier calls may lead to clearer reward signal (e.g. when corruptions are not consistent across calls), but such repeated calls introduces a risk of redundancy and thus inefficiency. By estimating the value of an additional verifier query, VStress addresses this problem, resulting in higher RLVR scores and fewer verifier calls/item than naive approaches to repeating verifier calls.

Critical strengths

The paper addresses an important problem of interest to the ICLR community: estimating the reward for a language model output when the output verifier(s) may be noisy/corrupted.

Critical weaknesses

The vocabulary and writing style are similar to an LLM’s in a way that makes the paper very difficult to understand. For example, consider that the first contribution listed is “an auditable binary-feedback contract with a post-decision oracle join, failure-aware cost ledger, and hash-bound replay manifest”. I understand what these words mean in isolation, but they aren’t communicating effectively what the paper’s contribution is – the contributions section is a place to highlight how the literature is being advanced, whereas this reads like Claude’s notes to itself.

The experiments appear to only use GSM8K, and key experiment details like the LLM/model used are left unclear.

Questions for authors

The "AI Use" section said that AI was used for polishing the writing, but I think the original writing would be helpful to see because the AI writing is very hard to understand. Could you please provide the human-written abstract, introduction, and method sections of the paper (as they were before AI polishing)? Those would help me better understand how to position this work’s methods and contribution with respect to the existing literature.

Can the authors please clarify the experiment details further (e.g., what LLMs/models were used) and test their conclusions more broadly (e.g. by going beyond GSM8K and, potentially, use of just a single LLM)?

Reviewer_uV2rOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a way to evaluating model-generated candidate answers and using those judgments as feedback for reinforcement learning with verifiable rewards (RLVR). The paper studies verifiers as potentially noisy sources of binary judgments, and asks how to combine or selectively query them when their errors are correlated. The paper proposes a mechanism to decide which verifier to query next—and when to stop—based on how much additional information that verifier contributes beyond the evidence already collected. Based on that, the authors introduce an auditable framework and an adaptive policy, VStress-CA, for allocating repeated verifier calls when verifiers may share correlated errors. Its central idea is to value an additional verifier by the new information it contributes beyond previously queried verifiers, measured through conditional mutual information estimated on a separate calibration split. The policy discounts uncertain estimates, normalizes information gain by call cost, and decides whether to query another verifier, accept a sufficiently confident prediction, or abstain; it also falls back to a conservative stopping rule when monitored verifier behavior shifts. The auditing framework records verdicts, failures, decisions, and costs before revealing evaluation ground truth, allowing verification quality, coverage, and cost to be assessed together. The authors also showcased some empirical success and provided practical insights on which additional verifiers add values.

Critical strengths

This paper has several critical strengths:

  1. The paper provides a principled formulation to determine and quantify the value of the additional verifiers. This is extremely valuable for research in RLVR and related areas.

  2. The paper also introduces an interesting fixed-budget evaluation framework for multiple verifiers, which connect verification quality with resources allocation and downstream tasks, as ever verifier call consumes resources.

Critical weaknesses
  1. One critical weakness in this paper is that the RVLR experiments with real verifiers lack details on implementation, making it difficult to reproduce. For example, if I want to understand which training dynamics produces the improvement beyond synthetic voting, it's not obvious to evaluate from the information provided in the paper.

  2. The most explicit mechanism experiments are based on synthetic data (512 records with periodic binary oracle), which limits the practicality of the claim.

Questions for authors

Can you specific the RLVR experimental details as follow:

  • The actual verifier models and learner model, including checkpoints or sizes.
  • Verifier prompts and decoding settings.
  • The definition of “same-family” versus “cross-family” verifiers.
  • The RL training algorithm, objective, optimizer, and learning rate.
  • The named datasets and exact definition of the reported “RLVR score.”
  • How abstained examples enter—or are excluded from—the learning procedure.

By providing those details it can help me make a better judgement on the experimental design.

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

Related papers

Discussion (0)

Sign in to comment.