acceptodds
Withdrawn from ICLR 2027

From Detecting to Characterising LLM API Updates: Black-Box Inference of the Change Mechanism

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent works proposed the task of change detection of LLMs behind APIs. This responded to practical concerns that some providers were observed to modify their models (e.g., with fine tuning or system prompt) without warning their end-users, and that caused potentially severe regression on downstream applications. This paper proposes to go a step further than mere change detection: we study change-type inference. Given an LLM endpoint that is known to have changed, we predict whether the update involved fine tuning, a system-prompt modification, or abliteration, using only sampled responses from the updated endpoint. We first give a theoretical grounding for the possibility of such a novel task, by demonstrating that changes that are local to a model provoke a monotonically evolving distance from that base model, which is traceable in its output by black-box auditing methods. We then propose (Change-Type Inference), a scheme that decides which change has been performed, and that relies on Border Inputs, which are sensitive to model modifications and already leveraged in change detection problems. All in all, this paper demonstrates the feasibility of detecting which change occurred on an LLM in a black-box setup, which we believe will be of interest for auditors and regulators.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM, Audit, Black-Box, LAN

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_FLdROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Change Type Inference (CTI) to classify previously detected LLM API updates into one of three categories, namely system prompt changes, fine-tuning, and abliteration. CTI uses Border Inputs (BIs) collected from anchor models and trains a classifier on output-probability features from labeled landmark models. At inference time, these features are estimated using tokens sampled from the updated endpoint. The theoretical analysis examines divergence along model changes, distinguishes signature direction from magnitude, and characterizes local sensitivity at low temperature. The experiments evaluate classification performance on open-weight LLMs using held-out instances of the considered change types.

Critical strengths

[S1] Token-Only Change-Type Inference

Although constructing the fine-tuning and abliteration landmarks requires white-box access to the original model’s weights, the paper provides empirical evidence that the considered change types can be distinguished using sampled-token features under simulated token-only access to the updated endpoint. Results on held-out fine-tuning themes and abliteration techniques further support generalization to unseen instances of these change types.

Critical weaknesses

[W1] Limited Practical Applicability

CTI has limited applicability to black-box API auditing because landmark construction requires access beyond that assumed in the stated black-box setting. Section 2 assumes black-box sampling access to the source model and the landmarks (lines 80–81), but also requires correctly labeled landmarks derived from that model and covering every considered change type (lines 81–91). Constructing the fine-tuning and abliteration landmarks requires white-box access to the weights of the original model (Appendix I.1, Tables 5–6). The paper does not establish how an external auditor could obtain these landmarks for a proprietary source model.

Furthermore, the three-class formulation provides no mechanism to reject changes outside the considered categories (Section 4.1, lines 287–290). The three categories do not cover other changes, such as quantization, which is listed among observed endpoint modifications in Section 1 (lines 34–35), or decoding-time watermark insertion. Nevertheless, CTI would assign such updates one of the three labels, even though their underlying causes fall outside these categories. Although change detection is assumed to occur upstream (Section 2, lines 74–83), false alarms would similarly force CTI to assign a change-type label to an unchanged model. The evaluation explicitly considers "unknown updates of known types" (lines 290–291) and does not assess how CTI handles unchanged models or updates outside the three considered types.

[W2] Insufficient Justification of the Proposed Method

The connection between the theoretical analysis and the proposed classifier is unclear. The introduction motivates CTI through BI-based distances and trilateration (lines 49–53), while Section 3 distinguishes the direction of a change signature from its magnitude, interpreting them as indicating which change occurred and whether it is visible, respectively (lines 195–198). However, the paper does not explain how this distinction motivates the use of logistic regression with concatenated token-probability features (lines 297–301). Appendix K.2 describes logistic regression as “a stated choice”, rather than one selected on the basis of the comparative results (lines 1499–1500), and Section I.3 lists its configuration without providing a rationale. The Δ\Delta-Span classifier uses the unexplained fraction of a displacement, making its decision invariant to displacement magnitude (lines 1597–1599), and achieves nearly identical mean macro recall, 0.83 vs 0.84 (Table 9, which is in fact a figure), leaving it unclear why logistic regression is preferred over this alternative that more directly reflects the theoretical emphasis on signature direction.

The necessity of BIs for change-type classification is also insufficiently supported. Section 3 claims that BIs are “the only prompts on which changes can be told apart” (lines 263–265). However, Theorem 4 concerns local sensitivity in the low-temperature limit. It does not establish that BIs outperform ordinary prompts for distinguishing change types under the experimental conditions. The ablations do not compare BIs with ordinary prompts under the same landmark set, classifier, and query budget. Consequently, the results do not isolate the benefit of BIs from that of supervised classification using outputs from labeled landmark models.

Questions for authors

I have several questions, and I list them below in descending order of importance to my assessment.

[Q1] Access Requirements

As discussed in [W1], constructing the required landmarks needs access beyond the stated black-box setting. Is CTI intended specifically for endpoints whose original model weights are available? If not, how could an external auditor obtain these landmarks for a proprietary source model? Relatedly, how does performance change when the landmark probability features used to train the classifier are estimated from sampled tokens rather than obtained directly from log-probabilities?

[Q2] Classifier and Anchor Design

As discussed in [W2], the motivation for logistic regression over the similarly performing Δ\Delta-Span classifier is unclear. What motivated this choice? Do per-LLM comparisons reveal advantages of logistic regression over Δ\Delta-Span that are obscured by the reported averages? In addition, using SS alone as a single anchor yields nearly the same mean macro recall as fusing 12 anchors (0.83, averaged over the eight LLMs in the SS column of Fig. 5, vs. 0.84 in Table 9). What does multi-anchor fusion contribute beyond SS?

[Q3] Unchanged Models and Other Change Types

As discussed in [W1], CTI does not provide a rejection mechanism for unchanged models or updates outside the three considered types. What does CTI output for these inputs, and can classifier confidence or another available score distinguish them from updates belonging to the considered types?

[Q4] Generalization Across Modification Recipes

The held-out fine-tuning themes share the same LoRA configuration (Table 5), while the held-out abliteration techniques come from the same library (Table 6). How does CTI perform on fine-tuning with a different recipe, such as a different LoRA rank or full fine-tuning, and on abliteration from an independent implementation?

[Q5] Contribution and Construction of BIs

As discussed in [W2], the experiments do not isolate the contribution of BIs. Could you compare BIs with ordinary prompts while keeping the landmark set, classifier, and query budget fixed, including at a higher temperature such as τ=1\tau = 1? Furthermore, the reproducibility statement refers to Section I for the BI search (lines 499–500), but Section I.2 only describes subsampling from an existing BI pool. How is this pool constructed, what criterion is used to identify BIs, and what are the typical logit gaps between their top two tokens relative to τ=0.1\tau = 0.1?

[Q6] Relation to Prior Work

Finlayson et al. [1] discussed inferring update types from changes in model outputs and their underlying subspaces, using log-probabilities and logit bias to recover the required output distributions without access to model weights. Oh et al. [2] trained metamodels to infer attributes of black-box models from query outputs using models with known attributes. How does CTI relate to these approaches? Which contributions go beyond these existing formulations, and which specifically address sampled-token access to the updated endpoint without log-probabilities or logit bias?

References:

[1] Logits of API-Protected LLMs Leak Proprietary Information (COLM 2024; Finlayson et al.)

[2] Towards Reverse-Engineering Black-Box Neural Networks (ICLR 2018; Oh et al.)

2 comments

Sampled versus logprob landmark features (1rst question)

Please refer to response to reviewer UdoK (4rth question) for that same question, explaining that one can create landmarks without access to weights with the use of fine-tuning APIs.

Logistic regression versus Δ\Delta-Span (2nd question)

We chose logistic regression as a standard learner, simple to explain. However, the reviewer's argument convinces us: we will make Δ\Delta-Span the main classifier, report logistic regression next to it, and add the per-LLM table. In particular, the theory identifies two structural features that a classifier could exploit: the strength of the displacement and its direction. The experiments show that the direction carries almost all of the information and that the strength adds almost nothing. Indeed, Δ\Delta-Span only sees the direction of the displacement and is invariant to its size. It reaches a macro recall of 0.83, against 0.84 for logistic regression, which has access to everything (Table 9). We now see this as a contribution of the paper in its own right. It was undersold in the submission, and we will emphasize it in the revision.

No reject class for other changes (3rd question)

Indeed CTI is evaluated on a closed-set: a change outside the trained types receives a trained label, and we will state this limitation. We are currently running the experiments for unseen changes. We will upload the results before the end of the discussion phase.

We note that CTI is the second stage of a two-stage audit: it never declares a change, it labels a change the upstream detector declared (eg, a proper change-dectection mechanism).

Single LoRA fine-tuning recipe (4rth question)

This is an interesting question which was also asked by reviewer 9nty, please refer to the beginnig of the response.

We can add that unseen abliteration techniques are exactly what our protocol tests. The library we use, OBLITERATUS [1], aggregates techniques developed independently: four of our eight landmarks implement separately published methods (FailSpy's difference-of-means, Gabliteration, Heretic, refusal direction optimisation). Splits hold out two of the five non-anchor techniques (Section 4.1, Table 6), and three of those five are independently published. A held-out technique thus plays the role of a future technique the auditor did not anticipate.

References: [1] OBLITERATUS Contributors (2026). OBLITERATUS: An Open Platform for Analysis-Informed Refusal Removal in Large Language Models. GitHub repository. https://github.com/elder-plinius/OBLITERATUS

BIs versus ordinary or random prompts (5th question)

We agree that the paper has no control that keeps landmarks, features, classifier and budget fixed and replaces the BIs. We are curently running this in the re-run of our pipeline and will publish the results before the end of the discussion phase.

Relation to prior work (6th question)

We thank the reviewer for their very relevant litterature pointers. We were not aware of both works that indeed witness a long running (2018) interest in black box change type inference. Compared to both works, we operate in "hard label" settings, a stringent difference considering that less than 25% of the llms offer logprob access. Compared to [1], we avoid the construction of an intermediary "llm image" which is relatively costly requests-wise. We will include both in the related works of the revision.

Thank you for the response. As the discussion period is coming to an end and I have not yet received a further response, I am posting this follow-up. [Q1] and [Q2] were the highest-priority questions in my review, and both concerns remain.

[Q1] Access Requirements

The response relies partly on fine-tuning APIs, whose availability can be restricted or withdrawn. For example, OpenAI has already closed self-serve fine-tuning to new users and will stop accepting new jobs from existing users on January 6, 2027 [3]. Even where fine-tuning remains available, it is unclear how an independent auditor would obtain the full set of labeled landmarks, particularly for abliteration, from the same source model version as the endpoint under audit.

[Q2] Classifier and Anchor Design

My original question was about the rationale for using logistic regression in the submitted method. Making Δ\Delta-Span the main classifier would require re-evaluating the main results and the query-budget and anchor ablations, as well as revisiting the claims that rely on them. In my view, a revision of this scope would be better assessed as a re-submission.

Furthermore, several other questions remain unresolved, with the relevant experimental results still pending. I am provisionally keeping my rating at 2 and may revise it based on the final results and further responses, even after the discussion period. I encourage you to share the results when available, as they also address questions raised by other reviewers.

[3] OpenAI. Deprecations (https://developers.openai.com/api/docs/deprecations).

Reviewer_9ntyOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces drift-type inference: given an API model that is known to have changed, predict from sampled outputs whether the change was a fine-tuning, a system-prompt modification, or an abliteration. The method probes the model on border inputs, prompts where the next-token distribution is nearly tied and which the authors show theoretically to be sensitive to small changes, and trains a classifier on known variants of the source model. With 500 queries, it achieves 0.82–0.91 macro recall on seven of eight open models.

Critical strengths
  • New and useful auditing task. Cheap in queries.
  • The analysis explains why border inputs are good probes.
Critical weaknesses
  • All fine-tunings use the same LoRA recipe. Only the topic is held out. The classifier may learn that recipe, not "fine-tuning".
  • Errors may just reflect how big the change is. A magnitude-only baseline is needed.
  • You need the source model weights to build landmarks. Closed APIs do not give this.
  • Common changes like quantization or version swaps are not covered. No "other" class.
  • Sampling is simulated from stored logprobs, not real API calls.
Questions for authors
  1. How do different fine-tuning recipes affect the predictions?
  2. How does a a magnitude-only classifier baseline compare with the proposed method?
  3. What happens on a quantized model?
  4. Why does knowing the change type matter in practice? Landmarks need the source weights, so this seems to fit third-party hosts of open models, not closed APIs. And knowing "it was a fine-tune" does not say whether my task got worse. What would an auditor do differently with the label? Why is detection alone not enough?
1 comment

Single LoRA fine-tuning recipe (1rst question)

We agree. We initiated new experiments with different LoRA receipes. Results are pending.

Missing magnitude-only baseline (2nd question)

If we get your question right, we think that Appendix K answers this question. Three of its rule-based classifiers cannot see the size of the change, by construction: TV-Corr compares which BIs moved relative to one another, Δ-Cos the direction of the change, and Δ-Span the share of the change that each type explains. A classifier blind to magnitude therefore recovers nearly all of the type signal: the classification does not rest on how big the change is.

The question is answered in FLdR's response, with the size-invariant classifier of Table 9.

We agree that this result is easy to miss in an appendix, and we will state it in the main text, next to the main results.

Quantized endpoints not covered (3rd question)

We agree, and quantization is a good addition: it is among the changes we cite (lines 34-35) and fp8 is the most frequent declared precision on a large router of open-weight endpoints [1]. All models in the paper are bf16. We are adding quantization as a fourth class, in a re-run on the same eight LLMs (τ = 0.3).

  • Landmarks: six quantized versions of S per LLM, built from S's weights by round-to-nearest quantization of every decoder linear layer, without calibration: fp8 E4M3 per tensor, int8 per output channel, int4 groups of 128 and of 32, NF4, MXFP4.
  • Split: each of our 35 splits holds out one quantization for test.

Results are pending and will be delivered before the deadline. Quantization is also a held-out type of the open-set appendix.

References: [1] OpenRouter (2026). LLM Quantization: What fp8 and fp4 Do to Quality and Price. https://openrouter.ai/blog/insights/llm-quantization/

Landmarks need source model weights (4rth question)

Reviewer owZW had the same question, we answered there (and do not duplicate for readability).

Sampling simulated from stored logprobs

Indeed: we relied on logprobs for convenience, but agreeing with your observation the revision will rely on (local-)API-sampled answers only.

Reviewer_UdoKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies change-type inference for LLM APIs: after a change has been detected, the goal is to identify whether it involved fine-tuning, a system-prompt modification, or abliteration. CTI uses responses to Border Inputs as features and trains a classifier on labeled variants of the same source model. The paper also analyzes how output distributions respond to model changes.

Critical strengths

The paper studies an interesting and practically relevant topic: identifying what kind of update happened behind an LLM API. It introduces CTI, an approach that combines BI responses with labeled reference models, providing a useful starting point for studying this task. The mathematical derivations are largely sound under their stated assumptions.

Critical weaknesses
  1. The theory seems to focus on a different question from the paper’s main claim. Much of Section 3 explains why BIs are sensitive to model changes, which was already the main motivation in the earlier BI paper. The selling point here is identifying the type of change. I could not see why Theorem 1 and Corollary 2 are needed, or what they are meant to tell us about CTI. They give a derivative identity and monotonicity along a restricted trajectory, but their role in the rest of the paper is unclear. Do they support any particular choice of probes, anchors, or classification rule? If their purpose is only to provide background, the paper should say so and put more emphasis on what makes change types distinguishable.

  2. The experiments do not show whether choosing BIs actually helps CTI. As described, the main method feeds BI response features into logistic regression. This can be useful, but Table 9 mostly compares different ways of classifying the same features. A more direct comparison would keep the landmark bank, feature construction, classifier, and 500-query budget fixed, and replace the BIs with ordinary or random prompts—for example, 100 prompts sampled five times each. How much does performance drop? This would help establish whether BI selection contributes to the new task, beyond showing that a classifier can learn from labeled update examples.

  3. The intended API setting needs to be clearer. The method requires correctly labeled fine-tuned and abliterated variants of the same source model. How would an independent user obtain these for a proprietary API? The training features also use exact landmark probabilities. Please clarify what access is needed during setup and which costs are excluded from the 500-query budget. The current results seem most applicable to a known model with an available reference bank, rather than a general opaque API.

Questions for authors
  1. How do Theorem 1 and Corollary 2 support CTI?

  2. How does CTI compare with ordinary or random prompts under the same query budget?

  3. Without access to logits, you can only estimate whether an input is a BI from sampled responses. Could small nonzero logit gaps substantially reduce sensitivity through the exponential decay in Theorem 4? How is this handled during BI selection?

  4. How could an independent auditor obtain the required labeled landmarks for a proprietary API?

Reviewer_owZWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose Change Type Inference (CTI), a method for classifying whether an LLM behind an API has been fine-tuned, given a new system prompt, or abliterated (safety guardrails removed), using only 500 sampled token queries. They show that Border Inputs (prompts where the model is undecided between multiple tokens) are uniquely informative for distinguishing change directions in weight space, with a phase transition result showing their responsiveness diverges as temperature decreases. The method uses a trilateration strategy with pre-computed "landmark" models of known modification type. They evaluate on eight open-weight models (3B-12B), and achieve macro recall of 0.82-0.91 on 7/8 models, with near-perfect second-order classification of fine-tuning themes.

Critical strengths

The paper reads well and the technical detail is easy to follow.

Identifying which of seven fine-tuning domains was used from 500 sampled tokens is a strong demonstration of how much information BIs carry.

The insight that BIs concentrate change signatures while ordinary prompts become informationally "dead" at low temperature is a nice contribution.

Critical weaknesses

The paper's entire contribution is a classifier that tells you how a model was changed. But it never demonstrates that knowing the mechanism is more useful than directly measuring the impact of the change. When an auditor detects a change, the proposed mechanism may say it was caused by fine-tuning, but that doesn't really provide information on if/where performance shifted. So they would still need to run performance auditing. Auditors and users care about whether the model got better, worse, or more dangerous for their use case. It is not shown in this paper whether they get anything valuable from this work. I think the proposed approach could be useful in improving the focus of that auditing, but that is not shown or mentioned by the authors other than saying "fine-tuning changes knowledge, system prompts change behavior, abliteration removes safety" without validation.

Another limitation is the requirement of access to features within the model (landmarks). This greatly limits the applicability as the authors mention the motivation being LLMs behind APIs which do not typically provide this access.

Questions for authors

Can you describe a concrete scenario where knowing the change type (rather than measuring the behavioral impact) leads to a better auditing decision?

Does CTI output feed into any downstream action in your envisioned deployment? If an auditor receives the label "abliteration," what do they do next that they would not have done without the label?

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

Related papers

Discussion (0)

Sign in to comment.