acceptodds
Under review as a conference paper at ICLR 2027

READER: Dynamic LLM Provenance from Query-Varying Interactions

Abstract

Auditing the provenance of large language models (LLMs) is essential for verifying service claims and detecting model substitution. Yet API services rarely expose the internals required by white-box methods, while many black-box methods establish comparability through shared diagnostic prompts. Auditing query-varying interactions therefore requires comparable source evidence as observations accumulate. We formalize this setting as dynamic black-box LLM provenance within a fixed set of enrolled models. To make these interactions comparable, we introduce READER, which uses a frozen proxy LLM as a common measurement space and encodes the mean and coarse evolution of activations along each response into spectral fingerprints. A linear probe trained once at enrollment extracts source evidence that Bayesian accumulation combines across observations. On Agent500, our benchmark containing 50,000 responses from 100 models to 500 heterogeneous agent prompts, READER reaches attribution accuracy from one response and from 100 on unseen prompts, exceeding fine-tuned DeBERTa and LLM-DNA with frozen sentence encoders at the respective budgets. Experiments across proxy families further demonstrate strong cumulative attribution, while static analyses reveal model relationships in the same spectral representation. READER thus turns ordinary interactions into cumulative evidence for provenance auditing.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Provenance, Generative AI Safety, Interpretability

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_18LBOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a new setting that we can use dynamic black-box provenance to verify which large language model generates the response, rather than the existing white-box (need access to the LLM parameters, gradients or logits) or static black-box (auditors must send a fixed shared prompt panel to the models, which is really expensive in time and moeny cost, and missing the interactiction between users and the LLMs) settings. The author proposes to firstly use a frozen proxy LLM to do the DCT extraction, then use the extracted features to train a simple linear classifier, then accumulate the log probes to calculate the average score. The experimental results shows that the performance is both fast and accurate.

Critical strengths
  1. The problem setting is quite straightforward to follow, previous work focus on static panel, which is not realistic in the real work deployment, then move to a more realistic scenario, that interact with the user logging.

  2. The method design is elegant and lightweight to build, when using the DCT to extract DC and AC, the computational complexity is only O(Td)O(Td), which is realy fast.

  3. Extensive experiments that covers around 100 models, with 8 different proxy LLMs, and the results hold across various task, domain, and token length.

Critical weaknesses
  1. From Table 2, I find that the domain shift really breaks the performance in this field, the performance drops to 33% from 96.2% when transfering to the Math100, even though it is enough better than other baselines. However, in the real world application (since the primary area allocated as societal consideration), whether it means that if the collected prompt distribution does not cover the real domain of the user query, the auditing will always fail, is ther any way can aviod or mitigate this situation.

  2. The most critical one is that in the Eq.8, it assums that model responses are conditionally independent given the source LLM and the observed prompts, which may sounds reasonable in the single-turn queries, but it may not hold in realistic multi-turn or agent based settings. It is worthy to discuss what is the scope of the assumption, and what if the real case is adaptive, multi-turn interactions.

Questions for authors
  1. I find that in the model families, there are lots of reasoning models (Qwen3-4B-Thinking, DeepSeek-R1-Distill), but I cannot find whether the contents contains the reasoning trace when feding in the proxy LLM to extract fingerprints. Moreover, if yes, whether the performance comes from the features extracting from the reasoning traces, rather than the real LLM fingerprints.
Reviewer_aw3iOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper formalizes dynamic black-box LLM provenance, attributing already-observed, query-varying prompt, response records to one of a fixed set of enrolled candidate models as evidence accumulates and proposes READER, which feeds each record through a frozen proxy LLM, summarizes the response-aligned hidden states by their DCT mean and first AC coefficient, applies a linear probe trained once at enrollment, and averages log-posteriors across records. On the new Agent500 benchmark (500 agent prompts × 100 local and API models), READER reaches 50.4% 100-way accuracy from one response and 96.2% from 100, well above fine-tuned DeBERTa and sentence-encoder LLM-DNA adaptations. Additional experiments cover transfer under response truncation and a math-domain shift, static relationship classification on Bench-A, and the geometry of model signatures.

Critical strengths
  1. The idea is straight-forward and easy to follow.
  2. The passive, query-varying provenance setting is practically motivated.
  3. The method is lightweight and the results are consistent across four proxy families.
Critical weaknesses
  1. The methodological contribution is thin relative to its framing. Table 4 shows DC-only (plain mean pooling) matches DC–AC on Agent500 (0.474/0.955 vs 0.476/0.956) and mean+max is better at K=1; the AC term only helps the pair task. “Bayesian Evidence Accumulation” is mean log-posterior under an independence assumption and gains 1.3 points over probability averaging at K=100. The method is effectively mean-pooled frozen-LLM states + logistic regression + naive-Bayes accumulation, and the baselines (sentence encoders, DeBERTa truncated to 256 tokens and trained two epochs) mainly differ in encoder size; a LoRA-tuned classifier on the same proxy, or an explanation of why Qwen3-Embedding-8B pooling reaches only 18.5% vs 50.4% for the Qwen3.5-9B proxy, is missing.
  2. The closed-set assumption and weak transfer undercut the motivating use case. READER cannot reject sources outside the enrolled set (the most likely scenario when an intermediary swaps backends), accuracy drops to 33% on Math100, Eq. (8) assumes independent stateless interactions, and no evasion (system-prompt wrapping, temperature change, output paraphrasing) is evaluated.
Questions for authors
  1. Given DC-only matches DC–AC on the main task, what does the spectral framing add for attribution beyond mean pooling?
  2. How does a LoRA-fine-tuned classifier on the same proxy compare with the frozen linear probe?
  3. How does READER behave when the true source is not among the 100 enrolled candidates, and can the scores be calibrated for open-set rejection?
  4. What happens to accuracy if the backend adds a system prompt or changes temperature to 1.0?
1 comment

Thanks for the rebuttal. I am raising the score to 3.

Reviewer_HG9hOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper describes a method for fingerprinting LLM outputs for attribution/provenance. They define the dynamic provenance scenario, where an observer wants to establish provenance for iterated calls (e.g. proving that an API provider is really serving the same model over time). This differs from both the watermarking scenario and the "white box" attribution scenario, where observers can access model internals. Rather, this method uses a proxy model to generate spectral fingerprints (from the first two components of the DCT) by observing a set of enrolled models on training prompts. The authors use this as input to a linear classifier and also accumulate evidence over time. Their results show that this is an effective provenance method when compared to other baselines, and results scale well with K, the number of observations at test time. In other words, the classifier performs better the longer it gets to observe LLM outputs.

Critical strengths
  • Studies an interesting problem, essential LLM identification given a fixed set of models. Extends this domain to consider evidence gathering over time (dynamic attribution), which is a very nice extension to the many papers in this field.

  • Clearly explains the DCT inspired method and classifier, comparing to reasonable baselines like encoder classifiers and other LLM identification methods.

  • Results show smooth scaling with the amount of evidence K (e.g. Figure 3).

  • Method is demonstrated to remain effective under many settings, such as different proxy models and ablations.

Critical weaknesses
  • Table 4 shows that other pooling methods are close in metrics, such as mean + max
  • Method could discuss and account for difference in complexity such as amount of stored data, cost to capture fingerprints, computational cost of runtime fingerprints (i.e. smaller local models vs larger expensive models, all requiring access to hidden states)

Non-critical: tables and graphs are a bit difficult to read in some places, such as small text or high density (e.g. Table 3). Consider revising these. Also, some terminology such as DCT, DC, AC could be written out for readers unfamiliar with the transform.

Questions for authors

My main question is around the motivation of the 2 component spectral fingerprint. Is there some intuition, motivation, or compelling small experiment? It would be great to add this to the paper, if so.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.