acceptodds
Under review as a conference paper at ICLR 2027

Fighting Fire with Fire: Assessing Test Set Contamination via Deliberate Training on Test Data

Abstract

Test set contamination inflates benchmark scores and misrepresents what a model can do. Existing detectors read a trained model’s static outputs, which a developer can optimize against directly. We audit by intervention instead: fine-tune the suspect model on the benchmark in question and read the shape of its loss curve. A model re-exposed to data it has seen before descends differently than one meeting that data for the first time, and the difference survives zero-normalizing every curve, so it lives in the shape of learning rather than in the loss level. We introduce FIRE (Fine-tuning-Induced Re-Exposure), which classifies these curves with dynamic time warping and k-nearest neighbors: no learned parameters, no model internals, one fine-tuning run (about ten minutes for a 7B model on a single GPU) and a lookup against a reference library built once per benchmark. Across six models (1B–7B), 44 benchmarks, and over 43,000 training runs, FIRE reaches 77–92% accuracy at contamination levels that meaningfully inflate benchmark scores, matching or beating five membership inference baselines and beating every baseline that does not require a clean same-architecture model. The signal transfers across model families at 87%, reaches a 70B model at 69–78% zero-shot and 88.7% with 50 labeled trajectories from it, and persists through a subsequent preference-optimization (DPO) stage with clean data. Detection needs as few as 256 benchmark samples, while evading it requires a full probing run inside every poisoning step. What a model has already seen is legible in how it learns.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Data Contamination Detection, Training Dynamics, Membership Inference, Memorization, LLM Evaluation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ER9iOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FIRE, which detects test set contamination by analyzing the loss trajectories obtained from fine-tuning a suspect model on the target benchmark. The authors argue that a model previously exposed to the benchmark may show different learning dynamics when trained on it again, compared to a model seeing it for the first time. To detect contamination, FIRE 1) fine-tunes the suspect model on the benchmark and records its loss trajectory, 2) zero-normalizes the trajectory to remove the initial loss level, and 3) uses dynamic time warping (DTW) with k-nearest neighbors (kNN) to classify whether the model is contaminated based on the trajectory shape. The authors evaluate FIRE by deliberately introducing benchmark data during training and comparing clean and contaminated models. Experiments across six models and 44 benchmarks show 77–92% detection accuracy when contamination increases benchmark performance by at least 5 percentage points. The authors also evaluate cross-model transfer, partial contamination, and the effects of subsequent DPO training. They further argue that FIRE is more difficult to evade than static contamination detectors, as an adversary would need additional fine-tuning to observe the detection signal.

Critical strengths
  • Using training dynamics as a contamination signal: Unlike existing methods that mainly rely on static model outputs, FIRE actively fine-tunes the suspect model and analyzes how its loss changes. The authors also show that the signal remains after removing the initial loss level, suggesting that the learning dynamics themselves contain useful information about previous data exposure. This provides an interesting alternative approach to test set contamination detection.

  • Extensive evaluation across models and benchmarks: The authors evaluate FIRE on six models and 44 benchmarks under different poisoning ratios and training configurations. Table 2 shows that the detection signal transfers across different models with an average accuracy of 87%. The additional experiments on partial contamination and larger models further support the applicability of the proposed approach beyond a single experimental setting.

Critical weaknesses
  • Insufficient comparison with simpler trajectory-based methods: The authors compare FIRE mainly against five membership inference baselines, but do not clearly show whether DTW+kNN is necessary for detecting contamination. Figure 4 shows that clean and contaminated models have different loss decay patterns, and the early part of the trajectory already provides strong detection performance. However, it is unclear whether simpler methods, such as the initial loss reduction, average loss change, or Euclidean distance between normalized trajectories, could achieve similar results. Since the main idea is to use the loss trajectory as a contamination signal, these comparisons are important to distinguish the contribution of DTW+kNN from the basic differences in learning dynamics.

  • Limited validation under realistic contamination scenarios: The experiments simulate contamination by deliberately including benchmark data during an SFT-based training stage, and mainly evaluate cases where contamination increases benchmark performance by at least 5 percentage points. Although this provides a controlled evaluation, it may not fully reflect contamination occurring during large-scale pre-training. Moreover, Section 4.4 shows that FIRE can detect contamination even when the probing samples have little overlap with the leaked samples. This raises the question of whether FIRE specifically detects benchmark leakage or simply recognizes that a model has already been trained on similar data distributions. Without further evaluation distinguishing these cases, it is difficult to assess the reliability of FIRE in realistic auditing scenarios.

Questions for authors
  • Clarification of the contribution of DTW+kNN: Figure 4 shows that the early loss trajectory already contains a strong contamination signal. Could the authors compare FIRE against simpler trajectory-based methods, such as initial loss reduction, average loss change, or Euclidean distance between normalized trajectories? This would help clarify whether DTW provides additional benefits beyond the basic differences in learning dynamics.

  • Clarification of realistic contamination detection: Section 4.4 shows that FIRE can detect contamination even when the probing samples have little overlap with the leaked samples. Could the authors clarify how FIRE distinguishes actual benchmark contamination from legitimate training on different but distributionally similar examples? In particular, would a clean model trained on similar but non-overlapping data also be classified as contaminated? This would help assess the false-positive risk in realistic scenarios.

  • Clarification of adaptive evasion resistance: Section 4.5 argues that evading FIRE requires approximately 850 times more computation, but the paper does not experimentally evaluate an adversary directly optimizing against the detection signal. Could the authors clarify whether cheaper surrogate objectives or alternative poisoning strategies could reduce the detection accuracy without requiring a full probing run at every poisoning step? This would help assess the claimed resistance to adaptive evasion.

3 comments

We thank the reviewer for recognizing that "the learning dynamics themselves contain useful information about previous data exposure," and for the concrete suggestions. We ran the suggested simpler comparisons and investigated the realistic-contamination and evasion questions.

Q1. "Figure 4 shows that the early loss trajectory already contains a strong contamination signal. Could the authors compare FIRE against simpler trajectory-based methods, such as initial loss reduction, average loss change, or Euclidean distance between normalized trajectories?"

Thank you for the insightful suggestion! We ran all three comparisons on the same data and cross-validation splits as the main results (now in §4.2; per-model numbers in Appendix A.5).

Method (same data and splits)Accuracy
FIRE (DTW + kNN)0.833
Euclidean distance + kNN0.848
Logistic regression on 8 trajectory statistics0.861
Mean of the normalized curve (average loss change)0.831
Total loss drop0.803
Initial loss drop (steps 1–5)0.791
Static-loss baselines (Simple Loss, Min-K%, zlib)0.57–0.61
Shorter audit (100-sample probe, 1,000-sample library):
FIRE (DTW + kNN)0.794
Euclidean distance + kNN0.651
Logistic regression on 8 statistics0.739

Every method that reads the zero-normalized loss curve recovers the signal (0.79–0.86), while static-loss baselines stay at 0.57–0.61. Contamination is visible in how the model re-learns the benchmark, and this does not depend on DTW.

DTW's advantage is handling trajectories of different lengths, allowing audits with fewer samples than the reference library. Unlike Euclidean distance, it relaxes exact step-by-step alignment: a loss drop at step 10 can be matched with a similar drop at step 15. This discards some timing information and may reduce accuracy slightly, but our goal is to show that trajectory shape alone retains a strong contamination signal. With a 100-sample audit against a 1,000-sample library, DTW leads Euclidean distance on all six models (0.794 vs. 0.651).

Q2. "Could the authors clarify how FIRE distinguishes actual benchmark contamination from legitimate training on different but distributionally similar examples? In particular, would a clean model trained on similar but non-overlapping data also be classified as contaminated?"

Great question! Our experiments cover two levels of similarity: data that is distributionally similar but truly non-overlapping, and distributionally identical data drawn from the benchmark itself. The latter is sadly often a common practice to boost benchmark scores (e.g., generating synthetic data that looks exactly like the test set, vulgarly known as "benchmaxxing"), and we believe it should count as test set contamination.

Distributionally similar, non-overlapping data is the negative class in every experiment. Clean and contaminated models are matched pairs (§2): both are trained on Tulu-v2, an instruction mixture covering task types that many of our benchmarks test (natural-language inference, paraphrase detection, question answering), with the same number of samples and the same hyperparameters. The only difference is that the contaminated model has benchmark items in place of part of that data. FIRE classifies 84% of these clean models as clean (false-positive rate 16.1%; 8.5–22.0% across models).

Distributionally identical data is detected (§4.4). The model is trained on 100 items from a benchmark, and the auditor probes with a different 100 items from the same benchmark, largely disjoint from those leaked. FIRE detects contamination at 78.1%, close to the 79.4% when probing with the leaked items themselves.

Q3. "Could the authors clarify whether cheaper surrogate objectives or alternative poisoning strategies could reduce the detection accuracy without requiring a full probing run at every poisoning step?"

The cheap strategies we tested weaken the signal only slightly. Without running the audit, an adversary can make the model memorize less or leak less data. The paper tests both:

  • Memorizing less: label smoothing during contaminated training lowers FIRE's accuracy from 71.2% to 68.3%, still about 7 points ahead of every baseline (§4.5).

  • Leaking less: FIRE detects contamination when only 100 of 1,000 benchmark items leak (78.1–79.4%, §4.4), and when benchmark data is only 0.5% of the training mix (§4.1).

Going further means optimizing against FIRE itself, which requires running the audit on the model at every poisoning step. §4.5 estimates this at about 850× the cost of the contamination itself (roughly 1,700 vs. 2 GPU-hours for a 7B model).

AuthorsOpenReview ↗

Rebuttal (2/2)

W2. "The experiments simulate contamination by deliberately including benchmark data during an SFT-based training stage [...] it may not fully reflect contamination occurring during large-scale pre-training."

We inject contamination during mid-training because that is where benchmark-like data enters modern pipelines; open pipelines such as OLMo 2 add curated task data at this stage. Two results suggest the signal survives dilution and later training: FIRE detects contamination when benchmark data is only 0.5% of the training mix (§4.1), and reaches 84.4% accuracy after an additional DPO stage (§4.7). Contaminating full-scale pre-training in a controlled way would require pre-training models from scratch, which is outside the scope of this paper.

We focus on contamination that raises benchmark scores by at least 5 points because that is the contamination that misleads evaluation: Magar & Schwartz (2022) show that a model can see benchmark data without exploiting it, in which case its scores are unaffected.

We hope these results resolve the reviewer's concerns about soundness, and would be grateful if the reviewer would consider raising the score. We are happy to answer any further questions.

I thank the authors for answering my questions. I am now able to clarify what the main problem the paper wants to solve. I increased the score from 2 to 3.

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

This paper targets the detection of test set contamination, an increasingly important topic that impacts the development of LLMs. Authors audit the contamination by directly intervening: fine-tuning the suspect model on the test set and observing the shape of the loss curve. This paper considers three scenarios: SFT contamination without post-training, SFT contamination with post-training, and post-training contamination. Results show that the proposed classifier method outperforms some MIA baselines.

Critical strengths
  1. The proposed method is straightforward. It exhibits competitive detection performance with as few as 256 samples.
  2. This paper is generally well-organized.
Critical weaknesses
  1. The proposed method is straightforward, yet highly intuitive. The motivation and rationale behind the design are unclear. Although the experimental results are competitive, the lack of rationale still makes it less convincing.
  2. The proposed FIRE method requires white-box access to the target model and fine-tuning its parameters. However, most MIA baseline methods only need to access the probabilities of the model's output. Therefore, there is a clear difference in the access assumptions between these two methods. Could the authors discuss this limitation and consider comparing it with stronger white-box contamination detection methods?
  3. How is the classifier constructed in detail? The proposed method seems oversimplified. The KNN classification and K-medoids clustering are common techniques that have limited technical innovation. Moreover, the rationale that this classifier suits the detection of test set contamination is unclear.
  4. How about the detection performance when the model undergoes RLHF or GRPO?
Questions for authors

Please see the weaknesses.

3 comments

We thank the reviewer for noting that FIRE "exhibits competitive detection performance with as few as 256 samples" and that the paper is "generally well-organized."

W1. "The motivation and rationale behind the design are unclear."

Thanks for giving the opportunity to further clarify this: A model that has already learned the benchmark has less left to learn when re-exposed to it. Its loss therefore falls gradually, while a clean model's drops sharply. FIRE zero-normalizes each curve to remove the starting loss and keeps only this shape. Shape alone reaches 77–92% accuracy across six models (§4.2), and Figure 4 shows the two patterns. We have made this rationale even more explicit in §1.

W3. "How is the classifier constructed in detail? [...] The KNN classification and K-medoids clustering are common techniques that have limited technical innovation. Moreover, the rationale that this classifier suits the detection of test set contamination is unclear."

Our contribution is the considering loss trajectory as a contamination signal for LLM contamination, not a new classification algorithm. We use standard components so that the accuracy can be attributed to the signal.

The classifier was described in detail in §2, we further add Algorithm 1 in the revised version to address the reviewer’s potential confusion. Algorithm 1: fine-tune the suspect model on the benchmark, record the loss at each step, subtract the initial loss, compute the dynamic time warping (DTW) distance to each labeled curve in the reference library, and take the majority label of the k nearest (kNN). kNN has no learned parameters, which suits a library with few labeled curves per benchmark. K-medoids is used only to visualize representative curves (Figure 4).

DTW is not the only workable distance: on curves of equal length, Euclidean distance and a classifier on summary statistics do as well (§4.2, Appendix A.5). DTW handles curves of different lengths, so one library also serves audits with fewer samples: with a 1,000-sample library and a 100-sample audit, DTW reaches 0.794 while Euclidean distance falls to 0.651.

W4. "How about the detection performance when the model undergoes RLHF or GRPO?"

§4.7 evaluates DPO, widely used in place of RLHF. Contamination introduced before DPO stays detectable after it (84.4%). Contamination introduced during DPO is detected when the auditor also probes with DPO rather than SFT (62.0%), so the audit should match the objective of the contaminating stage. GRPO is not evaluated (Appendix B); since recent work finds that brief GRPO can conceal static contamination signals (§5), it is the natural next setting to test.

We hope this addresses the reviewer's concerns, and would be grateful if the reviewer would consider raising the score. We are happy to answer any further questions.

W2. "The proposed FIRE method requires white-box access to the target model and fine-tuning its parameters. [...] Could the authors discuss this limitation and consider comparing it with stronger white-box contamination detection methods?"

FIRE is built for auditing open-weight models: it needs the weights and one fine-tuning run, so it does not apply to closed, API-only models (Limitations, Appendix B). Within that setting the cost is small, about ten minutes for a 7B model on a single GPU and about one GPU-hour for 70B (Appendix A.1). The access requirement is also what makes FIRE hard to evade. The auditor runs the fine-tuning, so a developer cannot target the signal without simulating that run at every training step, about 850× the cost of the contamination itself (§4.5). Static scores computed from a model's outputs offer no such protection.

Among the white-box methods we evaluate, FIRE matches the strongest while needing less access. All five baselines read the model's token-level losses, and the strongest, the reference-model ratio (82%), additionally requires a clean model of the same architecture as the suspect. FIRE matches or beats it on every model without one (§4.1).

W4. "How about the detection performance when the model undergoes RLHF or GRPO?"

§4.7 evaluates DPO, widely used in place of RLHF. Contamination introduced before DPO stays detectable after it (84.4%). Contamination introduced during DPO is detected when the auditor also probes with DPO rather than SFT (62.0%), so the audit should match the objective of the contaminating stage. GRPO is not evaluated (Appendix B); since recent work finds that brief GRPO can conceal static contamination signals (§5), it is the natural next setting to test.

We hope this addresses the reviewer's concerns, and would be grateful if the reviewer would consider raising the score. We are happy to answer any further questions.

Thank you for the clarification. I appreciate the DPO experiments; however, my concern regarding RLHF and GRPO remains insufficiently addressed. In particular, the recent finding that brief GRPO can conceal static contamination signals further motivates evaluating whether FIRE's dynamic loss-trajectory signals remain reliable under GRPO. Since no experiments involving GRPO are provided, its effectiveness under these increasingly common post-training paradigms remains unclear. I therefore consider this concern only partially addressed and would like to maintain the original score.

We appreciate the reviewer highlighting this important question. Recent work (Wang et al., 2026) demonstrates that brief GRPO training can substantially degrade existing likelihood-based contamination detectors. However, these methods rely on static differences in token likelihoods, whereas FIRE identifies contamination through the dynamics of loss under deliberate fine-tuning. Thus, the reported degradation does not necessarily imply that FIRE's trajectory-based signal is similarly affected.

While our current computational budget does not allow us to conduct a thorough GRPO evaluation during the rebuttal period, we agree that this limitation should be made explicit. We will highlight it in the revised paper and actively seek the additional compute needed to evaluate FIRE under GRPO for the camera-ready version. We believe this is a valuable next step in establishing the scope and robustness of FIRE across increasingly common post-training paradigms.

We hope that our existing experiments, which demonstrate strong detection performance across models, benchmarks, and training conditions, provide evidence of the value of this complementary signal!

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

To address the issue of test set contamination leading to inflated benchmark scores and misleading assessments of model capabilities, and the problem that existing detectors only read the static output of trained models, this paper employs an intervention-based auditing method. This method fine-tunes suspicious models on specific benchmark sets and analyzes the shape of their loss curves. Furthermore, this paper proposes a Fine-tuning-Induced Re-Exposuree (FIRE) method, which uses dynamic temporal transformation and the k-nearest neighbor algorithm to classify these curves. Its advantages include no need for parameter learning or internal model states, requiring only a single fine-tuning run, and comparison against a pre-built reference library for each benchmark set. Experiments on multiple models and multiple benchmark sets demonstrate that FIRE achieves high accuracy, outperforming all benchmark models that do not require clean, identical architecture models. It also exhibits high zero-sample transferability and stable optimization. While this paper is well-organized and provides sufficient experiments, some issues remain, such as the overly simplistic explanation of the method's principles and the potential overestimation of the classifier's role.

Critical strengths
  1. The Fine-tuning-Induced Re-Exposuree (FIRE) approach is simple and clear. It optimizes the suspected model on the target benchmark and records the loss value at each step of the training process. The loss trajectory shows significant differences in quantifiable metrics.
  2. This paper designs a classifier that uses dynamic time interval transformation and k-nearest neighbor algorithm to classify loss curves. It does not require learning parameters or model internal state, and has strong transferability between different retainer benchmarks and different model architectures.
  3. The authors conducted extensive experiments, fully demonstrating the advantages of the proposed FIRE method, which achieves high accuracy, outperforming existing benchmark models. It also exhibits high zero-sample mobility and stable optimization.
Critical weaknesses
  1. The threat model only specifies the capabilities of the defender, but does not clearly describe the attacker, the attacker's capabilities, or the consequences of the attacker's attack.
  2. The baseline method used for comparison is somewhat outdated, and the chosen comparison method does not effectively highlight the advantages of the method presented in this paper. It is recommended to add some new comparison methods, such as those found in the following works: [1] How Contaminated Is Your Benchmark? Measuring Dataset Leakage in Large Language Models with Kernel Divergence [2] Fine-tuning can Help Detect Pretraining Data from Large Language Models [3] Detecting Data Contamination in LLMs via In-Context Learning
  3. This paper largely consists of experiments, with only a small portion dedicated to introducing the proposed FIRE method. Therefore, it resembles an experimental research paper rather than a deep exploration of the underlying principles of the method. Furthermore, the comparative loss curves are not particularly innovative; the work on the classifier may be more valuable, but the classifier design simply utilizes existing methods and offers nothing new.
  4. The core of the success of the method in this paper lies in comparing the loss of the clean model and the contaminated model on the target benchmark. So how exactly is this clean model obtained? In the real world, given a model, to determine whether it is a contaminated model, which model should this model be compared with? How is the clean model obtained?
  5. This paper mainly seeks out contaminated models by comparing fine-tuning loss curves. However, different fine-tuning methods and parameters can significantly affect the fine-tuning loss, and even different random seeds can influence the loss. The authors do not seem to have explored the impact of these factors on the method's results.
  6. The loss curves obtained by using a simple non-parametric classifier based solely on trajectory morphology are classified. Whether this classifier can consistently achieve good results is uncertain. The losses on some models are not significantly different, and in some cases there is no difference. Simply comparing the magnitude of the loss values is not enough to distinguish whether a model is contaminated.
Questions for authors

Please refer to the Critical Weaknesses discussed above.

2 comments
AuthorsOpenReview ↗

Rebuttal (1/2)

We thank the reviewer for recognizing FIRE's simplicity, transferability, and extensive experimental evaluation, and for suggesting relevant related work, and for giving us the opportunity to clarify some key aspects of our method and our evaluation below.

W3. "The classifier design simply utilizes existing methods and offers nothing new," and the paper "resembles an experimental research paper [...]"

The contribution is a new detection signal for LLM evaluation contamination and the audit built on it, not a new classifier. A model that has already seen a benchmark has less left to learn when fine-tuned on it again, so its loss falls gradually, while a clean model's drops sharply (§4.2, Figure 4). The paper shows that this signal lies in the curve's shape alone (77–92% vs. 57–61% for static-loss methods), transfers across models (87% over 30 pairs) and to benchmarks held out of the library, persists through DPO (84.4%) and label smoothing (§4.5), and does not depend on the classifier (table above).

The paper is empirical by design: whether such a signal exists, transfers and persists can only be established with ground truth, so we train over 43,000 configurations with known contamination status. The revised §2 also states the full procedure as Algorithm 1.

W4. "In the real world, given a model, to determine whether it is a contaminated model, which model should this model be compared with? How is the clean model obtained?"

The suspect is trained with the target evaluation data, and its loss trajectory is compared against a reference library of labeled loss trajectories, not against a clean copy of the suspect model. The auditor builds this library once, by training clean models (without benchmark data) and contaminated models (with it) from a common base model (§2). The library does not need to match the suspect's architecture: a library from one model detects contamination in another with 87% accuracy across 30 model pairs (Table 2).

W5. "Different fine-tuning methods and parameters [...] and even different random seeds can influence the loss."

We agree that hyperparameter choices can affect training, which is why we believe in the importance of having an extensive experimental section for this paper.

We would like to differentiate the underlying suspect model’s training parameters (where we simulate clean and poisoned training) and the auditor’s parameters when training to probe a model. The auditor’s settings are fixed by design: the library and the suspect are probed with the same learning rate, epoch count and batch size. What the auditor cannot control is how the suspect was trained, and FIRE remains useful even when the suspect’s model parameters differ from the auditor’s loss library. In our paper, we already showed that a library built under one setting detects contamination under another across model sizes (89.0%), poisoning ratios (85.8%), learning rates (83.7%), and optimizers (79.3%) (§4.3, Table 4; Appendix A.9), and after DPO post-training (84.4%, §4.7). Changing which benchmark samples are audited moves accuracy by only 1.3 points (79.4% vs. 78.1%, §4.4).

W6. "Simply comparing the magnitude of the loss values is not enough to distinguish whether a model is contaminated."

We agree, and this is precisely why FIRE removes the loss magnitude: every trajectory is zero-normalized, so classification uses only its shape (§4.2). Shape alone reaches 77–92% accuracy across six models, versus 57–61% for methods based on static loss. Figure 4 shows the two shapes.

AuthorsOpenReview ↗

Rebuttal (2/2)

W2. "The baseline method used for comparison is somewhat outdated [...]"

We appreciate the suggested references. The revised paper positions FIRE against all three (§5) and compares it with simpler methods on the same loss curves (§4.2).

FIRE's advantage is that it needs no clean reference model. The only baseline that comes close, the reference-model ratio (82%), requires a clean model of the same architecture as the suspect, which an auditor rarely has. FIRE matches or beats it on every model without one (§4.1), and a library built on one model audits another at 87%.

Each of the suggested papers answers a different question from FIRE: Kernel Divergence Score ranks datasets by contamination level without a decision threshold, FSD detects which individual examples were in pretraining, and CoDeC measures how in-context examples change confidence. FIRE decides whether a given model is contaminated on a given benchmark.

We also added simpler methods that read the same loss curves (§4.2; per-model numbers in Appendix A.5):

Method (same data and splits)Accuracy
FIRE (DTW + kNN)0.833
Euclidean distance + kNN0.848
Logistic regression on 8 trajectory statistics0.861
Mean of the normalized curve (average loss change)0.831
Total loss drop0.803
Initial loss drop (steps 1–5)0.791
Static-loss baselines (Simple Loss, Min-K%, zlib)0.57–0.61
Shorter audit (100-sample probe, 1,000-sample library):
FIRE (DTW + kNN)0.794
Euclidean distance + kNN0.651
Logistic regression on 8 statistics0.739

Every method that reads the normalized loss curve recovers the signal (0.79–0.86); static-loss baselines stay at 0.57–0.61. DTW additionally aligns curves of different lengths, so one library serves audits with fewer samples (0.794 vs. 0.651 for Euclidean distance).

W1. "The threat model [...] does not clearly describe the attacker [...]"

We have brought the threat model together in one place in §2:

Auditor: has the model weights and can fine-tune the suspect on the benchmark, but needs neither its training data nor its recipe (already in §2).

Attacker: a model developer who controls the training pipeline and includes benchmark data deliberately or through undocumented data collection. We consider two settings of compute-bounded attackers: a passive adversary who applies label smoothing (FIRE remains the strongest detector); and we also quantify the cost for an active adversary who optimizes against FIRE (~850× the poisoning run) (from §4.5).

Consequences: inflated scores that misrepresent model capabilities and distort model comparisons and deployment decisions (from §1).

We hope these responses address the reviewer's concerns, and would be grateful if the reviewer would consider raising the score. We are happy to answer any further questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General comment to the AC

Since the discussion period is coming to an end, we thank the reviewers for their engagement and summarize the discussion and restate our contributions below:

Contribution. FIRE is the first method to detect benchmark contamination from how a model re-learns the benchmark: the shape of its loss trajectory when the auditor fine-tunes it on that benchmark. A model that has seen the data has less left to learn, so its normalized loss curve takes a different shape. Existing detectors read a trained model's static outputs; FIRE reads learning dynamics that the auditor controls, so a developer cannot target them without about 850× the cost of the contamination itself.

Empirical success. Over 43,000 controlled runs, six models (1B–7B, plus 70B), 44 benchmarks: 77–92% accuracy vs. 57–61% for static-loss baselines, 87% transfer across model families, and robustness to hyperparameter changes, partial leaks and DPO post-training.

Discussion summary.

  • ER9i raised their score (2 → 3) after our new comparison with simpler trajectory methods (§4.2): every method that reads the loss curve recovers the signal, confirming the signal itself is the contribution.
  • qGJn's other concerns are addressed; their follow-up raises only GRPO. The reviewer appreciated the DPO experiments, which show contamination stays detectable after post-training (84.4%), and we plan to extend this result to GRPO for the camera-ready version.
  • Lrmq has not yet responded. The revision addresses their points: a consolidated threat model and Algorithm 1 (§2), and positioning against the three suggested detectors (§5).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.