- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This paper proposes controlling model capabilities through key-conditioned attention geometry rather than textual passwords. A 12-dimensional continuous key specifies control-point correspondences used to construct positional transformations and phase remapping. Protective training preserves task performance under the authorized key while suppressing performance under incorrect keys or standard RoPE. The evaluation covers authorized utility, incorrect-key behavior, standard-RoPE bypass, gradient-based key recovery, and fine-tuning-based recovery across several models and tasks.
Critical strengths
Moving capability control outside the input-token space is an interesting direction. Conditioning the positional computation provides a distinct mechanism from textual triggers and creates a concrete setting for studying structural access control.
The security evaluation goes beyond a small number of randomly selected incorrect keys. The paper includes gradient-based recovery and fine-tuning recovery at different data budgets, while Appendix A.4 additionally evaluates 120 perturbed keys. The authors also appropriately restrict their claims to the tested attacks and budgets rather than asserting cryptographic unbreakability (pp. 7–9 and 14–15).
The results show a substantial empirical separation between authorized and unauthorized task performance. For example, the original, authorized, and standard-RoPE MMLU scores for Qwen3-4B are 66.86, 64.21, and 23.55, respectively (p. 6, Table 2). This establishes that the structural condition affects usable performance under the evaluated protocol.
Critical weaknesses
W1. The white-box analysis should target functional attention geometry, not only recovery of the original key.
Equations (5)–(7) suggest that, when ε is sufficiently small, the frequency-dependent normalization approximately cancels out, making the normalized phases nearly identical across frequencies. Since the subsequent nonlinear mapping is shared, the construction may effectively reduce to a shared nonlinear positional warp, rather than a more complex frequency-dependent transformation.
This does not establish equivalence to standard RoPE or imply a successful bypass. However, it suggests that recovering the original key may be unnecessary: a white-box attacker could instead optimize a simpler, functionally equivalent positional geometry.
The authors should verify this approximation using the actual frequency range and reference length (pp. 4, 13), and compare key recovery against direct optimization of a compact positional warp under matched data and compute budgets.
W2. Low task scores should be separated from failure to produce the expected answer format.
The incorrect-key training branch uses random input spans as targets. Several standard-RoPE MMLU scores fall below the 25% uniform-guessing reference for four-option questions, including 17.27 for Qwen3-0.6B and 14.10 for Llama-3.1-8B (Section 3.4; p. 6, Table 2).
These results are consistent with capability suppression, but they may also include invalid-option generation or failure to follow the answer format. Below-chance performance does not, by itself, establish that the underlying knowledge or reasoning is inaccessible.
The paper should specify whether MMLU is scored through free-form generation, constrained option decoding, or option likelihoods. Reporting invalid-output rates and adding an option-constrained or likelihood-based evaluation would distinguish suppression of task capability from suppression of the current response interface.
W3. The recovery-cost interpretation lacks the most relevant public-base-model reference.
With full task-data fine-tuning, the standard-RoPE variant recovers to 65.31, approaching the pre-attack authorized score of 67.89 (p. 8, Table 6). The evidence therefore supports delayed recovery under the tested budgets rather than irreversible prevention of recovery. The paper largely acknowledges this boundary.
The remaining issue is the reference point for the claimed additional cost. Comparisons with Password-Locked are concentrated on one model/task setting, while the original unprotected model primarily appears as a full-data reference. Since the underlying base model is publicly available, a practically relevant alternative is to adapt that base model directly to the task.
At matched data and computation budgets, how does recovering the protected model compare with simply adapting the public base model? Adding this curve at a few already evaluated data fractions would make the economic and practical meaning of the protection substantially clearer without requiring a broad expansion of the benchmark suite.
Questions for authors
- Over the implemented frequency and context ranges, how accurately do Equations (5)–(7) reduce to φ_j(p) = ω_j ψ_K(p)? Does directly optimizing a compact effective position function reduce recovery cost?
- How is MMLU scored, and what fraction of unauthorized outputs are invalid? How much of the authorized–unauthorized gap remains under constrained options or option-likelihood scoring?
- Under matched data and computation budgets, what is the task-adaptation curve from the public base model? Does this comparison change the interpretation of the protection’s additional recovery cost?
Final Discussion Reminder
Dear Reviewers,
The discussion period is nearing its end. If you have not already done so, please review the authors’ rebuttal and follow-up responses. If you have any remaining questions or concerns, please leave a comment for the authors.
Thank you again for your time and efforts!
Best regards,
Area Chair