- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 3
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies how an evaluator can trace a student model trained only from its scores. Radioactive Feedback slightly changes scores according to registered answer preferences, which the student may learn during training. A suspect model is audited by comparing its preference score with unmarked reference models. Across nine student and evaluator pairs, the four preference code produces gains of 7.6 to 15.4 percentage points. All eight marked checkpoints pass the calibrated tests, while 40 controls do not. The paper also examines continued training, targeted removal, and answer quality.
Critical strengths
(1) The paper addresses an important provenance setting not covered well by earlier methods.
The student generates its own answers and receives only evaluator scores. The mark can therefore be introduced without teacher demonstrations or changes to the student's training pipeline. The tested preferences cover both writing style and solution strategy.
(2) The evaluation is broad and well documented.
The paper tests three student models, several evaluators, and nine student and evaluator combinations. It includes matched references, controls from other model families, continued training, targeted removal, and different adjustment strengths. The appendices provide detailed data, audit, and implementation information.
Critical weaknesses
(1) The reported evidence does not establish that the audit works in the stated black-box setting or that it remains effective after the suspect undergoes further training.
a. Section 3 says the provider does not know the suspect's optimizer or private data. However, Appendix I builds the reference pool using the same architecture, data split, training method, and budget. Because the rank test requires comparable reference models, the audit assumes information that may not be available in the stated black-box setting.
b. The conclusion says the mark remains detectable after continued training. However, RQ5 measures only carrier alignment and does not repeat the calibrated detection test on later checkpoints. It also studies single-carrier models rather than the joint code used in the main detection experiment. The results show persistent marked behavior, but not persistent detection.
(2) The paper's description of a controlled reward-adjustment mechanism and cross-data generalization is stronger than the evidence supports.
a. Section 4.2 says the evaluator first judges answer quality and then adds a small mark-based adjustment. But the paper observes only the final score, so it cannot verify that these are separate steps. Because marked models often receive higher utility scores, part of the effect may come from a changed quality judgment rather than the intended adjustment alone.
b. Propositions 5.1 and 5.2 study an idealized objective, not the GRPO procedure used in the experiments. GRPO normalizes scores within each group, which can amplify a small reward difference. The propositions offer intuition, but do not show that the actual training procedure preserves answer quality as claimed.
c. Section 6.2 treats APPS and CodeContests as held-out benchmarks, but TACO includes data collected from both sources. The paper removes exact duplicates only, so closely related examples may remain. The results therefore do not clearly demonstrate generalization to genuinely new data. The same concern applies to the mathematics datasets.
Questions for authors
Related to Weakness 1a. Section 3 assumes that the suspect's training setup is unknown, but Appendix I uses closely matched reference models. What information does the method actually require? Does the test remain valid when this information is unavailable? If so, please support this claim with experimental evidence or a theoretical justification.
Related to Weakness 1B. The conclusion says that the mark remains detectable after continued training, while RQ5 measures only carrier alignment. To retain this conclusion, please apply the calibrated detection test from RQ4 to the continued joint-code checkpoints. Otherwise, revise the conclusion to state only that carrier alignment persists.
Related to Weakness 2. Are the evaluator decomposition and quality-preservation result properties of the actual method or only idealized assumptions? Propositions 5.1 and 5.2 do not cover group-normalized GRPO, and the base quality score is not observed. Please clarify these claims.
6 comments
Q2 (W1b): Calibrated Detection on Continued Joint-Key Checkpoints
The conclusion says that the mark remains detectable after continued training, while RQ5 measures only carrier alignment. To retain this conclusion, please apply the calibrated detection test from RQ4 to the continued joint-code checkpoints. Otherwise, revise the conclusion to state only that carrier alignment persists.
First, we clarify the checkpoint structure in RQ5 (Section 6.5 and Appendix G): while the Llama continuation curves use the single-carrier mathematics checkpoint (WM effect pp after UltraChat SFT and pp after clean GRPO), all Frontier (Qwen3-30B-A3B) continuation and removal experiments in RQ5 are conducted on the four-carrier joint-key checkpoints from RQ2 and RQ4.
Second, following your recommendation, we apply the calibrated RQ4 detection test directly to these continued four-carrier Frontier checkpoints against the frozen RQ4 reference pool and threshold (Table 10 in Appendix G, shown below).
Calibrated detection on continued four-carrier Frontier checkpoints (frozen RQ4 pool, , )
| Continuation Regime | Steps | Joint WM effect (pp) | Audit Score | Margin | Detected | |
|---|---|---|---|---|---|---|
| Initial marked checkpoint (RQ4) | 0 | +15.35 | 6.60 | +5.10 | Yes | 1/101 |
| Ordinary UltraChat SFT | 1,024 | +15.17 | 6.55 | +5.05 | Yes | 1/101 |
| Clean GRPO (switched to GPT-5.6 Sol) | 1,024 | +18.71 | 6.65 | +5.15 | Yes | 1/101 |
| Balanced-carrier SFT (targeted removal) | 2,048 | +7.20 | 5.25 | +3.75 | Yes | 1/101 |
| Anti-watermark GRPO (oracle removal) | 2,048 | +1.80 | 1.20 | −0.30 | No | $>$0.01 |
The four-carrier Frontier checkpoints remain detected under the frozen test after both ordinary instruction SFT and clean GRPO under a switched evaluator, whereas explicit anti-watermark GRPO reduces the score below . In the revision, Section 6.5 and Section 7 (Conclusion) report these results and specify that calibrated post-continuation detection holds for the four-carrier Frontier checkpoints while single-carrier alignment persists on Llama.
Q3a (W2a): Prompted Score Specification, Paired-Scoring Verification, and Independent Utility Evaluation
Section 4.2 says the evaluator first judges answer quality and then adds a small mark-based adjustment. But the paper observes only the final score, so it cannot verify that these are separate steps. Because marked models often receive higher utility scores, part of the effect may come from a changed quality judgment rather than the intended adjustment alone. Are the evaluator decomposition and quality-preservation result properties of the actual method or only idealized assumptions?
In our single-call LLM-as-a-Judge implementation (Section 4.2), Equations (3)–(4) define the scoring rule stated in the evaluator's prompt: the judge forms an internal rubric score and applies a bounded shift only when and the assigned carrier is classifiable. While the training API returns only the final scalar , we verify the realized decomposition and isolate task quality through three direct checks:
- Direct paired-scoring verification on fixed rollouts (Appendix P): To test directly whether the marked prompt alters base quality judgments or applies the prescribed shift, we score the same Frontier rollout responses under both the unmarked and marked evaluator instructions (table below). Ineligible responses remain unchanged within the repeat-scoring noise floor, while eligible responses on the keyed and opposite sides shift closely in line with the specified -point tilt and preserve within-group quality rank order.
- External programmatic utility evaluation and ablation (Appendix M): Audit-time task utility (Pass@1 via unit-test execution for code, and final-answer accuracy for mathematics) is evaluated by deterministic external checkers rather than by the LLM judge. Across Tables 1–3 and the sweep below, marked training preserves the capability gains of clean distillation. Moreover, because the reward bonus applies only to qualified responses () on the keyed side, a student that aligns with the watermark key slightly widens the reward margin between qualified and unqualified rollouts, explaining the modest positive under moderate .
- Why quality shifts cannot produce key-specific alignment: A scalar shift in overall pass rate or judge strictness cannot induce directional agreement with an arbitrary multi-bit sign vector . In addition, WM effect conditions on correct responses () in both Marked and Clean models, and decoy-carrier controls and key-flipping confirm that alignment arises strictly along the watermark key.
Same responses scored under the unmarked vs. marked evaluator instructions (Appendix P, )
| Response Group | Mean (Marked − Unmarked Score) | Specified Shift | |
|---|---|---|---|
| Ineligible (unmarked score or carrier unclassifiable) | 138 | +0.08 | 0 |
| Eligible, keyed side | 192 | +4.92 | +5 |
| Eligible, opposite side | 170 | −4.85 | −5 |
| Noise floor: unmarked instruction scored twice, mean absolute difference | 500 | 0.32 | — |
| Within-group Spearman rank correlation (eligible, keyed side) | 192 | 0.98 | — |
Watermark transfer and programmatic utility across tilt strength (Frontier joint setting, Appendix M)
| Tilt Strength | Max Shift ( on ) | Joint WM effect (pp) | (pp) |
|---|---|---|---|
| 0.000 (Clean) | 0.0 | 0.0 | 0.0 |
| 0.010 | +4.2 | within | |
| 0.025 | +11.5 | within | |
| 0.050 (Main) | +15.4 | +0.5 | |
| 0.100 | +17.2 | −0.35 |
Response to Reviewer Wsaf
We thank the reviewer for recognizing that our paper addresses an important provenance setting without teacher demonstrations or student-pipeline changes, and for noting the breadth and documentation of our evaluation across student models, evaluators, controls, continued training, and targeted removal. Below we address every weakness and question in detail.
Q1 (W1a): Required Information, Role of the Matched Pool, and Validity Without Training Information
Computing the audit score requires only the suspect's black-box outputs; the matched reference pool is a controlled scientific baseline rather than an operational requirement of the audit. The revision makes this distinction explicit in Section 6.4 and Appendix O:
Detection with a zero-training public reference pool vs. the controlled matched pool
Unmarked public models by training paradigm (Frontier audit)
Across both audits, all marked checkpoints are detected with a wide margin above the zero-training public pool, while unmarked public models—regardless of model family or post-training paradigm—cluster near zero with zero false positives under leave-one-family-out recalibration.