acceptodds
Under review as a conference paper at ICLR 2027

SCORE: A Model Equality Test for Auditing Black-Box LLM APIs

Abstract

Users increasingly rely on black-box inference APIs as their primary interface to large language models (LLMs), for both closed- and open-weight models (e.g., Llama models are now widely accessed through Amazon Bedrock and Azure AI Studio), with little insight into which model is actually deployed. To reduce costs or maliciously alter model behavior, API providers may silently deploy quantized models or substitute entirely different ones, potentially degrading model performance and compromising safety. Detecting such substitutions is challenging because users cannot inspect model weights and, in most cases, lack access to internal model signals such as output logits, leaving generated text as the only observable. To address this problem, we propose Semantic Contrast for Output-only Rank-based Model Equality Testing (SCORE), an output-only auditing method that tests whether a target LLM API and its official reference API induce the same response distribution, thereby assessing whether the black-box API behaves consistently with the model it claims to serve. SCORE measures the change in semantic recovery induced by a control prefix, converts each target-model score into a randomized rank calibrated against reference-model scores, and tests the resulting ranks for uniformity at significance level \(\alpha\) using only generated text. We evaluate SCORE across multiple threat scenarios, including model quantization, harmful fine-tuning, and full model replacement, as well as on commercial inference API endpoints. Across these settings, SCORE consistently achieves higher statistical power than existing methods under stringent black-box constraints.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
model substitution detection, output-only black-box testing, Auditing LLM APIs

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_KikCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SCORE, an output-only method for testing whether a target LLM API and a trusted reference endpoint induce the same prompt–response distribution. SCORE constructs paired prompts with and without a control prefix, measures the resulting change in semantic recovery of held-out lexical units, and converts target scores into reference-calibrated randomized ranks. A Cramér–von Mises uniformity test then detects distributional differences. The paper evaluates quantization, supervised fine-tuning, model replacement, and commercial API endpoints.

Critical strengths
  1. Practical access assumptions. SCORE requires only generated text from both endpoints, avoiding reference-side logits or weights. This meaningfully extends the applicability of rank-based auditing to settings where the reference is also a commercial API.
  2. Careful statistical calibration and explicit limitations. The randomized-rank construction handles ties and finite reference samples, while the Monte Carlo procedure provides type-I error control under the stated independence and stationarity assumptions. Appendix G also explicitly distinguishes failure to reject from evidence of model equality and discusses alternatives that the score or rank transformation cannot detect.
Critical weaknesses
  1. The methodological novelty is insufficiently established relative to closely related work. As the paper acknowledges (Section 2, Eq. 14, Appendix G), the score-to-rank-to-uniformity-testing framework comes from RUT, so the novelty rests mainly on the output-only semantic contrast score. However, this scoring design closely connects to ReCaLL (Xie et al., EMNLP 2024), which scores membership by the ratio of a text's log-likelihood conditioned on a non-member prefix to its unconditional log-likelihood, and PETAL (He et al., USENIX Security 2025), which uses semantic similarity between generated and ground-truth tokens under label-only access. These methods address membership inference rather than model equality testing and are not mathematically identical to SCORE. Nevertheless, the paper should explain what nontrivial methodological advance its particular combination provides, and neither ReCaLL nor PETAL is cited or discussed in the submission. Appendix F supports the usefulness of the components, but does not by itself establish their novelty.
  2. The empirical evidence is difficult to validate because of incomplete budget reporting and an internal inconsistency. Equation (17) shows that SCORE requires multiple generations for every position and prompt condition, but Appendix C does not establish that the baselines use comparable target and reference budgets. Section 3 states that reference requests are reported separately, but these counts do not appear in the experiments. Appendix F.0.1 matches budgets across SCORE variants, but no matched-budget comparison against MMD, PTS, RUT, or KS is given. The budget of N = 10m completions in the commercial-provider audit (Section 5.6) is also insufficiently connected to SCORE's position-wise paired-query protocol. In addition, Table 3(b) reports a SCORE power AUC of 0.130 for Llama 8-bit on MATH, whereas the corresponding power curve in Appendix D.4 integrates to approximately 0.83 (see Question 3), which makes it unclear which configuration the reported results reflect.
Questions for authors
  1. How does SCORE's semantic contrast score differ technically from combining PETAL-style semantic recovery with ReCaLL-style prefix normalization? What additional insight explains its suitability for model equality testing?
  2. Please report the actual target and reference request counts for each method, clarify the budget accounting in Section 5.6, and provide a representative comparison against the baselines under matched budgets.
  3. Table 3(b) reports an AUC of 0.130 for Llama 8-bit on MATH, whereas integrating the corresponding SCORE curve in Appendix D.4 (page 19) gives approximately 0.830. The baseline curves in the same panel also disagree with Table 3(b) (e.g., PTS reaches roughly 0.36 versus 0.062 in the table), while the other three Llama panels match their tabulated AUCs. The panel would also imply higher power for 8-bit than for 4-bit quantization on MATH (0.736), contrary to the statement in Section 5.5 that power decreases in the 8-bit setting. Could the authors clarify whether this discrepancy arises from a table entry, a figure label, or different experimental configurations?

I'll consider raising score if my questions are addressed.

Reviewer_KfM3OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper presents SCORE - an auditing framework to check LLM responses through API access. The goal is to detect undisclosed changes in the models in an "output-only" setup where only responses of the models are available through the prompts. The authors aim to detect semantic differences between the models' responses to detect the change. The authors considered three threat conditions to evade detection: model quantization, supervised fine-tuning, and full model replacement. The approach is evaluated on both open and closed models over various benchmarks including WildChat, multilingual Wikipedia, MATH, BigCodeBench, and UltraChat.

Critical strengths
  1. The authors considered a timely problem of detecting changes in models while relying on the soundness of the AI models for various tasks.

  2. The authors considered the practical setup of an "output-only" auditing setup. This is practical as model identities may not be reliable just from the description.

  3. The experimental setup considers three commonly used evasion techniques such as quantization and fine-tuning besides the obvious model replacement.

  4. Experiments considered a range of open and closed source LLMs. The benchmarks are spread across various tasks such as WildChat for conversation, MATH for mathematical reasoning, BigCodeBench for coding, and UltraChat for instruction following.

Critical weaknesses
  1. I am concerned about the generalizability of the approch for models that will go through continuous updates with little overall change. For example, GPT-Sol-light vs. medium. As these models are behind the wall, it is hard to know whether the updates are periodic and more continuous. I assume these models are going through continuous updates, and the general people are made aware only after significant changes.

  2. The approach seems to rely on a significant number of tokens (2KW target and 2mKW reference completions) for the detections. Why not consider training a classifier similar to fingerprinting approaches? What is the fundamental benefit of the proposed work over the line of work on model fingerprinting? How does the proposed approach compare with the SOTA fingerprinting-based detections?

  3. The details about the evasion approaches could be elaborated. For example, what quantization and fine-tuning techniques are applied? How sensitive is the proposed approch against those threat techniques?

Questions for authors

Please address the questions in the weaknesses section.

Reviewer_Yc3BOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies LLM API auditing and presents an output-only model equality test. The authors claim hat this outperform baseline methods across settings.

Critical strengths
  • This paper studies a well-motivated and pactically relevant problem.
  • The method is interesting and tested across broad experiments.
Critical weaknesses

There are some suspicious experiment results that make me question the paper's integrity.

1. The baseline numbers in Table 2 are exactly the same as the RUT paper (Zhu et al., 2026), but these two paper are using different model setup.

This paper's Table 2 rows Llama-4-bit (MMD 0.625, RUT 0.642, KS 0.474) and Gemma-4-bit (MMD 0.214, RUT 0.392, KS 0.017) match RUT's Table 2a digit for digit. But RUT's numbers are for Llama-3.2-3B and Gemma-2-9B (see RUT paper's Sec. 5.2), while this paper says the rows are Llama-3.1-70B and Gemma-3-27B (see Sec. 5.2). Six independent power AUCs matching exactly across different models seems impossible by chance. Other rows do not match, which points to mixing of copied and new values. This is very concerning to me.

2. Ablation numbers in Table 6 and 7 show implausibly exact proportionality.

In Tables 6 and 7, every variant's five per-model AUCs can be derived from SCORE's per-model AUCs times a constant ratio (the ratio spread about 5e-4) with accuracy to all 4 decimals. For example, in table 6, denote the ratio by α\alpha,

  • Each cell in the original-only (K=5) row is α∈[0.6421,0.6426]\alpha\in[0.6421,0.6426] times the corresponding cell in the SCORE row.
  • Each cell in the prefix-only (K=5) row is α∈[0.6043,0.6049]\alpha\in[0.6043,0.6049] times the corresponding cell in the SCORE row.
  • Each cell in the Semantic+Difference row is α∈[0.8869,0.8871]\alpha\in[0.8869,0.8871] times the corresponding cell in the SCORE row.
  • ...

Can the authors explain why these 500-trial Monte Carlo AUC estimates are proportional across models to roughly 1e-4? To me this pattern indicates the numbers were likely generated by direct scaling.

3. The commercial endpoint audit (Sec. 5.6, Table 4) reproduces Gao et al. (2025)'s study rather than running SCORE's protocol

The table has nine providers and four Llama 3/3.1 models, which match Gao et al. (2025)'s setup. The protocol details described in Sec. 5.6 (25 Wikipedia prompts at 50 tokens, 20 HumanEval/UltraChat prompts at 250 tokens, temperature 1, 10 repeats, alpha=0.01...) contradicts SCORE's own design (n=100 probes, next-word prompts, temperature 0.5, alpha=0.05, m=20). Can the authors explain why these choices were made?

Questions for authors

Can the authors address my concerns in the weakness section?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.