acceptodds
Under review as a conference paper at ICLR 2027

Toward Open Weight Models Without Risks: Separating Public and Private Capabilities in LLMs

Abstract

Open-weight Large Language Models (LLMs) enable scientific progress and broad deployment but complicate access control over sensitive capabilities. Current practice either suppresses dangerous capabilities before release or mediates access through closed services. The former is susceptible to jailbreaks while sacrificing capability for all users to mitigate the risks posed by a few, and the latter is incompatible with open-weight release. In this paper, we propose *Tiered Language Models* (TLMs), where a single set of released weights supports distinct capability tiers. In its default public configuration, a TLM behaves as a conventional LLM. A compact secret key specifies a permutation over a small parameter subset, inducing an alternative computation graph over the same weights that exposes additional capabilities. Our training protocol jointly pretrains both configurations from scratch, then fine-tunes the keyed configuration on private data with regularization to preserve the public model's behavior. We evaluate M- and M-parameter TLMs on language acquisition, instruction following, and factual memorization, using these tasks as controlled proxies for private or sensitive capabilities. The keyed configuration successfully learns each task, while the same weights in their public configuration do not exhibit the acquired capabilities. We show that this approach extends naturally to multiple hierarchical tiers. Because authorization operates on the model's weight structure rather than in the input space, the mechanism resists fine-tuning-based extraction and partial key compromise. In general, TLMs take a step toward reconciling open-weight release with selective capability control.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
open-weights models, LLM safety, privacy

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_76pwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on the problem of controlling access to sensitive capabilities in open-weight LLMs. The authors propose Tiered Language Models (TLMs), which use secret parameter permutations to enable different capability tiers within a single set of model weights. A two-stage training protocol is developed to separate public and private capabilities. Experiments on 180M- and 650M-parameter models evaluate capability separation, computational efficiency, adversarial robustness, and hierarchical capability tiers.

Critical strengths
  1. The paper tackles an important challenge in open-weight LLM deployment: enabling broad access to model weights while restricting certain sensitive capabilities. This paper provide a conceptually appealing approach in which a single set of model weights supports different capability configurations through compact permutation keys. The framework does not require additional learned parameters for each capability tier.
  2. The experiments cover language acquisition, instruction following, and factual memorization, providing evidence that the proposed mechanism can separate different types of learned behavior. The extension to hierarchical capability tiers further demonstrates the flexibility of the framework.
Critical weaknesses
  1. Limited technical novelty. Parameter-level access control for neural networks has been explored in prior work [1], and permutation-based model reconfiguration is not entirely new. While the proposed asymmetric training strategy is interesting, its technical distinction from existing approaches could be better clarified. Moreover, its effectiveness is not sufficiently validated due to the lack of comparison with symmetric joint pretraining.

  2. Insufficient evaluation against adaptive white-box attacks. The evaluated attacks are relatively limited and do not adequately reflect realistic adversaries with full access to the model weights. In particular, the partial-key experiments only consider randomly selected subsets of the correct key, without investigating whether an adversary could actively search for alternative parameter configurations that recover the restricted capabilities.

  3. Limited generalizability and scalability. The experiments focus on relatively small models (180M and 650M) and use simplified tasks as proxies for sensitive capabilities. It remains unclear whether the proposed method can effectively protect complex, generalizable capabilities in larger LLMs.

[1] CoreLocker: Neuron-level Usage Control, S&P 2024

Questions for authors
  1. What is the fundamental technical novelty of the proposed training protocol? Could the authors clarify the main methodological innovation and explain how it differs from existing approaches?

  2. Is asymmetric joint pretraining necessary for capability separation? Could the authors provide an ablation comparing the proposed asymmetric gradient updates against symmetric joint pretraining

  3. Can the authors evaluate capability extraction on more generalizable tasks? The current fine-tuning-based extraction experiment relies on synthetic biographies with independently generated facts. Would the proposed mechanism remain robust when attackers attempt to recover instruction-following abilities or knowledge learned from more structured datasets?

2 comments

We thank you for your effort in reviewing our paper. We are happy that you believe we are tackling an important problem with a conceptually appealing approach and extensive experiments. Below, we address the specific concerns:

W1(a)/Q1 "Limited technical novelty"

We thank the reviewer for the reference and we now cite [1] in section 2. Our contribution is a training protocol that yields two capable configurations over the same weights, with private capabilities added after public-only pretraining and kept out of the public configuration. We provide more context on differentiation from prior works below.

Difference from [1]. CoreLocker zeroes out a small set of high-impact weights in a trained network and gives their values to authorized users as the key. TLMs differ in three ways. (i) [1] withholds learned values, so the key is part of the model. A TLM releases every weight, and the key only specifies an arrangement. (ii) Without the key, [1] degrades the whole model, down to chance. A TLM's public configuration is a fully capable LLM, and the key only adds specific capabilities. (iii) [1] works post hoc because breaking the model is its goal. TLMs need two working configurations of the same weights, which joint pretraining provides and post-hoc permutation cannot (Appendix D.1).

Permutation-based reconfiguration. TrojanNet (Guo et al., 2020, discussed in Section 2) hides a secret network by permuting each layer's parameters and training both tasks jointly. However, its goal is covertness and it cannot be used for selective capability separation. TLMs instead apply permutation of weights for access-control of capabilities. We introduce non-trivial components such as asymmetric joint pretraining, and KL-anchored fine-tuning for a nested tiered access-control system.

Gradient routing and weight update. Gradient Routing (Cloud et al., 2024) and SGTM (Shilov et al., 2025) partition updates by data. Our rule partitions by computation graph: the same public data flows through two arrangements of the weights, and SS receives gradients only through the permuted one.

W1(b)/Q2 "Effectiveness of asymmetric joint pretraining"

We have added a new experiment to compare asymmetric vs symmetric joint pretraining. Our results show that symmetric joint pretraining results in around 25% more leakage compared to our proposed asymmetric approach.

[Revision: See Figure 10 in Appendix D.2]

W2 “Insufficient evaluation against adaptive white-box attacks”

We have now added an adaptive key search attack. We make it deliberately favorable to the attacker. The attacker knows exactly which weights form SS, so only the swaps are unknown. It scores each candidate key by the NLL of the ground-truth private bios. Starting from a random pairing, it applies random changes and keeps those that lower the NLL, over about 85K candidate keys (corresponding to about 16 hrs on a single H100 GPU). We observe that while the NLL falls significantly, it stays far above that of CKC_K. The keys it finds share almost none of the true swaps, and exact-match recall stays near 1%.

[Revision: See Figure 6 and L477-L487 in Section 6]

W3 “Limited generalizability and scalability.”

We agree that we do not evaluate protection of hazardous capabilities at frontier scale. We would like to highlight that we are functioning on an academic budget and that our goal is to illustrate the access-control utility of the proposed approach (i.e., separating capabilities within the same model), which we show over multiple types of proxy tasks. Note that the training procedure is the same for any private dataset.

To match the claims to the evidence presented, we are willing to amend the title of the paper (for example, "Tiered Language Models: Separating Public and Private Capabilities in Open-Weight LLMs"), and mark the Figure 1 dialogue as illustrative.

On scalability, the trend we observed is favorable. Across five scales from 42M to 650M, the public-domain perplexity gap between a TLM and a matched non-TLM model shrinks steadily, falling below 1% at 650M (see Figure 4).

Q3 "capability extraction on more generalizable tasks?"

First, on the independence of the biographies: since the held-out facts are independent of the attacker’s data, any recovery must come from knowledge hidden in the model. Our new password-locked baseline recovers 35% of them under the same attack, while the TLM stays near 0%.

[Revision: See Figure 5 left and Section 6]

Second, we repeated the attack on a generalizable capability, Spanish language modeling, at 650M. We find the same trend: there is no faster recovery of Spanish capabilities for C_pub compared to only-pretrained baseline or TLM trained for another capability (Portuguese).

[Revision: See Figure 13 and Appendix D.6]

We did not run this attack on instruction following or on structured knowledge such as WMDP within the discussion period. We plan to add these to the final version.

Dear Reviewer,

We wanted to briefly follow up on our response and thank you again for the review. We ran new experiments for the weaknesses you raised. We have added an adaptive attack, experimented with symmetric joint pretraining, and showed capability extraction on more generalizable capabilities. We have also provided a note distinguishing our work from previous approaches.

We hope that our results and clarifications have helped address your concerns.

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

This paper proposes Tiered Language Models (TLMs), which enable tiered access control within a single open-weight checkpoint. A compact secret key permutes approximately 5% of the parameters (attention heads and FFN neurons across layers) to unlock private capabilities, while the default configuration functions as an ordinary public model. The approach combines asymmetric joint pretraining with private fine-tuning restricted to the keyed parameters. Experiments on 180M and 650M models show that the keyed configuration acquires private capabilities, including a new language, instruction following, factual memorization, and WMDP, with little leakage into the public configuration.

Critical strengths
  1. Novel direction. The paper moves authorization from the input space (password-style prompts) to the permutation structure of existing parameters, without adding extra parameters. This offers a promising alternative to password-locking methods, which can be bypassed with a few demonstrations or fine-tuning.

  2. Simple method with a clearly motivated design. Asymmetric gradients teach the public configuration not to rely on the tier parameters during pretraining. Private fine-tuning then updates only the tier parameters with KL regularization. Each component has a clear role. Appendix D.1 further explains why permutation cannot be applied directly to an off-the-shelf model, clarifying the need to incorporate the mechanism during pretraining.

  3. Experiments covering diverse capabilities and practical costs. The proxy tasks span distributional capabilities (a new language), behavioral capabilities (instruction following), and pointwise facts (synthetic biographies). The inclusion of WMDP further connects the evaluation to the paper’s motivation. The paper also reports additional pretraining compute, the performance gap relative to standard models, and key size, and compares against standard fine-tuning with the same number of trainable parameters.

Critical weaknesses
  1. Limited practical usability and threat model. All authorized users share the same key, which cannot be revoked once leaked. This limits the practicality of the proposed access-control mechanism.

  2. The “fine-tuning on partial private data” experiment does not establish that extraction is difficult. The experiment uses synthetic biographies consisting of independent random facts. Knowing the biographies of half the individuals provides no information about the remaining half. Thus, near-zero performance on the held-out individuals would be expected even without TLM’s protection and provides limited evidence of security. This attack should also be evaluated on generalizable capabilities such as Spanish, instruction following, or WMDP.

  3. Unclear evaluation protocol. In the instruction-following experiment, the win rate of the keyed configuration increases from approximately 50% to 85%, while that of the public configuration falls to approximately 15%; the two sum to exactly 100%. However, the paper does not specify the comparison baseline. If the two configurations are evaluated against each other, the public configuration’s lower win rate simply reflects the keyed configuration’s improvement and does not establish whether the public configuration’s own capabilities have degraded.

Questions for authors
  1. How would the authors address key leakage when multiple authorized users share the same key? Could the framework support user-specific keys or revocation of a leaked key?

  2. Could the authors evaluate the “fine-tuning on partial private data” attack on generalizable private capabilities, such as Spanish, instruction following, or WMDP, and directly compare its effectiveness against TLMs and password-locking methods?

  3. If an attacker combines weight fingerprints with a small number of private samples for targeted key search, for example, through greedy swap search or differentiable relaxation-based optimization over the permutation, what computational and data budgets would be required to recover the key?

  4. In the WMDP experiment, the public configuration’s accuracy on the training questions increases from 25.5% to 32.3%. What causes this leakage? Can a larger KL regularization coefficient eliminate it, and what trade-off would this impose on private capability acquisition?

4 comments

We thank you for your effort in reviewing our paper. We are encouraged to see that you found our approach novel, well-motivated, and supported by diverse experiments. Below, we address the specific concerns:

W1/Q1 "All authorized users share the same key, which cannot be revoked once leaked"

We agree with you and we now state this explicitly as a limitation in the revised paper. Our proposed framework does not support user-specific keys or revocation of a leaked key. However note that the same holds for any way of shipping a restricted capability as a downloadable artifact. A private checkpoint, a withheld adapter, or a password-locked model cannot be revoked once it leaks either. Only API-mediated access supports per-user revocation, and it rules out open weights and self-hosting. TLMs do not solve revocation, and we do not claim that they do.

W2/Q2 “experiment does not establish that extraction is difficult”

We thank the reviewer for raising this issue. We have carried out 2 new experiments:

(1) Password-locked baseline. We fine-tune the non-TLM pretrained LLM with a specific password string prepended to all the synthetic bios. The lock holds perfectly i.e., 100% with password and 0% without. Mimicking the attacker, we then fine-tune it on 50% of the synthetic bios (without access to password) and measure leakage for the held-out set of synthetic bios (which the model was trained on with the password earlier). We find that the leakage reaches 35% exact match (as compared to nearly 0% for TLM). Note that even though the held-out facts are independent of the attacker’s data, the password-locked model recovers 35% of them -- so the attack does have the power to extract hidden facts. In comparison, we show that TLMs are robust.

[Revision: See figure 5 left and L424-L456 in Section 6]

(2) Attack on generalizable capabilities. We have now also carried out the attack experiment for a more generalizable capability: learning Spanish. We find the same trend: there is no faster recovery of Spanish capabilities for C_pub compared to only-pretrained baseline or TLM trained for another capability (Portuguese).

[Revision: See figure 13 and Appendix D.6]

W3 “Unclear evaluation protocol for instruction-following”

We apologize for the lack of clarity wrt the instruction following evaluation. We compared C_K vs C_pub head-to-head since we are measuring behavioral separation: we want to see if C_K gains instruction following capability compared to C_pub, which our results in Figure 2 (right) illustrate. We have now also made a comparison against an external GPT2-Large model (774M parameters). We see that C_K achieves 87.7% win rate against GPT2 while C_pub achieves 62.6%, indicating behavioral separation when measured against an external reference.

[Revision: See Appendix D.5]

Q3 “Adaptive greedy swap search attack”

We added an adaptive key search attack [See Figure 6 and L477-L487 in Section 6 of revision]. We make it deliberately favorable to the attacker. The attacker knows exactly which weights form SS, so only the swaps are unknown. It scores each candidate key by the NLL of the ground-truth private bios. Starting from a random pairing, it applies random changes and keeps those that lower the NLL, over about 85K candidate keys (corresponding to about 16 hrs on a single H100 GPU). We observe that while the NLL falls significantly, it stays far above that of CKC_K. The keys it finds share almost none of the true swaps, and exact-match recall stays near 1%.

[Revision: See Figure 6 and L477-L487 in Section 6]

Q4 WMDP leakage and larger KL

We believe the leakage is caused by the shared set of parameters and is particularly highlighted due to the sensitivity of MCQ-based evals to small changes in the model's distribution. The reviewer’s intuition is right and a higher KL reduces leakage, however it does so at the cost of reducing capability acquisition. We will include a discussion on this trade-off in the final version of the paper. Completely removing all leakage without affecting performance remains a challenging open question.

Dear Reviewer,

We wanted to briefly follow up on our response and thank you again for the review. We ran new experiments for the weaknesses you raised. We have added a password-locked baseline, showed robustness against an attack on generalizable capability, experimented with a greedy search-based adaptive attack, and provided clarification on instruction fine-tuning results.

We hope that our results and clarifications have helped address your concerns.

Thanks to the authors for the detailed response and the new experiments. The password-locked baseline and the Spanish attack experiment address my concern about the fine-tuning extraction attack. However, some concerns remain: the instruction-following evaluation still lacks a non-TLM reference for C_pub, the key search does not include a gradient-based relaxation attack, and the WMDP leakage trade-off is not quantified. I will therefore keep my score.

Thank you for the follow-up. Please find detailed responses below.

CpubC_{pub} instruction-tuning

We added the non-TLM references for CpubC_{pub} (Appendix D.6). Against the pretrained non-TLM model, CpubC_{pub} does not degrade. It rises from 43.4% to 66.1%, so it picks up part of the ability, while CKC_K rises from 47.0% to 91.9%. Against the instruction-tuned non-TLM model, CpubC_{pub} falls to 13.5%, while CKC_K wins 44.8% of comparisons. The public configuration therefore stays far from instruction-tuned quality, while the keyed configuration nearly matches it.

WMDP leakage trade-off

We have now quantified the WMDP trade-off (Appendix D.4). Leakage on seen questions falls from 13.0 points at β=0\beta=0 to 6.8 at our default β=0.1\beta=0.1 and 4.7 at β=0.3\beta=0.3 while CKC_{K} stays near 100%. Leakage disappears only at β≥0.7\beta \geq 0.7, where CKC_{K} drops to 51.7% or below. A larger KL weight can remove the leakage, but only by giving up the private capability. We believe the leakage arises because the KL term anchors CpubC_{pub} only on public text, so updates to the shared tier parameters can still shift its predictions on WMDP inputs.

Gradient-based attack

We have now added a gradient-based relaxation attack (Appendix D.8). We relax the key into a continuous assignment (Sinkhorn-normalized), backpropagate the NLL of the true private answers into it, anneal over 10K gradient steps, and project to a discrete key. The attacker optimizes on the answers of all individuals. The NLL falls from 5.77 to 2.66 but stays far above that of CKC_{K} (0.11). Exact-match recall reaches about 7%, compared to 99.8% with the true key, and the recovered key shares none of the true swaps. Since the attacker is scored on the same answers it optimizes, 7% is an upper bound on extraction. The attack also plateaus within about 1.5K steps, so more compute is unlikely to help.

We hope these new results help address your concerns.

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

This paper proposes Tiered Language Models (TLMs), a method that hides private capabilities within a secretly permuted subset of weights (attention heads and MLP neurons). The method uses asymmetric joint pretraining to teach the model different behaviors under different configurations, and then fine-tunes the keyed configuration on private data with regularization to preserve the public model’s behavior. The authors conduct several experiments on models with 180M and 650M parameters to demonstrate the method’s efficacy.

Critical strengths
  1. Encoding access control as a secret permutation of model parameters is a novel idea: it allows a single open-weight checkpoint to serve multiple capability tiers without requiring additional adapter weights.
  2. Experiments showcase performance across language modeling, instruction-following, and synthetic-biography tasks at a modest pretraining cost, suggesting that the approach is practical.
Critical weaknesses
  1. The robustness evaluation cannot support the security claims. The extraction attack uses memorized synthetic facts with no positive control (e.g., a password-locked model under the same attack). Generalizable capabilities, where password-locking fails, are never attacked, and adaptive attacks are missing. Since extraction resistance is the main claimed advantage, this drives my score.
  2. The claims outrun the evidence. The title and Figure 1 suggest protection of hazardous capabilities, but the evaluation uses small proxies, WMDP shows leakage on seen questions (25.5% → 32.3%), and the AlpacaEval win rate appears to compare C_K with C_pub.
Questions for authors
  1. Does fine-tuning C_pub on a small Spanish or Alpaca subset recover the capability faster than a non-TLM baseline?
  2. Is the AlpacaEval win rate C_K vs. C_pub? If so, what are the results against the standard reference?
2 comments

We thank you for your time spent reviewing the paper. We are pleased to see that you found our method novel and practical. Below, we address the specific concerns:

W1(a) “No positive control (e.g., a password-locked model under the same attack)”

We have now carried out the same attack with a password-locked model. Specifically, we fine-tune the non-TLM pretrained LLM with a specific password string prepended to all the synthetic bios. The lock holds perfectly i.e., 100% with password and 0% without. Mimicking the attacker, we then fine-tune it on 50% of the synthetic bios (without access to password) and measure leakage for the held-out set of synthetic bios (which the model was trained on with the password earlier). We find that the leakage reaches 35% exact match (as compared to nearly 0% for TLM).

[Revision: See figure 5 left and L424-L456 in Section 6]

W1(b)/Q1 “Generalizable capabilities are never attacked”

We have now also carried out the attack experiment for a more generalizable capability: learning Spanish. We find the same trend: there is no faster recovery of Spanish capabilities for C_pub compared to only-pretrained baseline or TLM trained for another capability (Portuguese).

[Revision: See figure 13 and Appendix D.6]

W1(c) "Adaptive attacks are missing"

We added an adaptive key search attack. We make it deliberately favorable to the attacker. The attacker knows exactly which weights form SS, so only the swaps are unknown. It scores each candidate key by the NLL of the ground-truth private bios. Starting from a random pairing, it applies random changes and keeps those that lower the NLL, over about 85K candidate keys. We observe that while the NLL falls significantly, it stays far above that of CKC_K. The keys it finds share almost none of the true swaps, and exact-match recall stays near 1%.

[Revision: See Figure 6 and L477-L487 in Section 6]

W2(a) "The title and Figure 1 suggest protection of hazardous capabilities, but the evaluation uses small proxies"

We agree that we do not evaluate protection of hazardous capabilities at frontier scale. The abstract and Section 4 describe our tasks as proxies. However, the capabilities that motivate gating, such as expert virology, are beyond what 650M models can learn, so we cannot evaluate them directly at this scale on an academic budget. Note that the training procedure is the same for any private dataset. Further note that we test several kinds of learning: a distribution shift (a new language), a behavior (instruction following), individual facts (synthetic bios), and a risk-relevant benchmark (WMDP).

To match the claims to the evidence presented, we are willing to amend the title of the paper (for example, "Tiered Language Models: Separating Public and Private Capabilities in Open-Weight LLMs"), and mark the Figure 1 dialogue as illustrative. Our goal is to present TLMs as a promising access-control method.

W2(b) "WMDP shows leakage on seen questions (25.5% -> 32.3%)"

We agree that there is a little leakage. For scale, on the same seen questions CKC_K gains nearly 74 points while CpubC_{pub} gains around 7. On held-out questions, CpubC_{pub} moves from 23.0% to 25.5%, which is chance level, while CKC_K reaches 37.3%. Note that multiple-choice accuracy is also a sensitive probe. A small shift of probability toward the correct option can flip the prediction, so it picks up more leakage compared to exact-match recall. Reducing all leakage is an important direction for future work.

W2(c)/Q2 “the AlpacaEval win rate appears to compare C_K with C_pub”

We compared C_K vs C_pub since we are measuring behavioral separation: we want to see if C_K gains instruction following capability compared to C_pub, which our results in Figure 2 (right) illustrate. We have now also made a comparison against an external GPT2-Large model (774M parameters). Comparing against GPT-4 reference is not sensible at this small scale. We see that C_K achieves 87.7% win rate against GPT2 while C_pub achieves 62.6%, indicating behavioral separation when measured against an external reference.

[Revision: See Appendix D.5]

Dear Reviewer,

We wanted to briefly follow up on our response and thank you again for the review. We ran new experiments for the weaknesses you raised. We have added a password-locked baseline, showed robustness against an attack on generalizable capability, experimented with an adaptive attack, and added instruction fine-tuning comparison against an external model.

We hope that our results and clarifications have helped address your concerns.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Discussion

Summary of discussion

We thank the reviewers and AC for their help reviewing our paper. We are happy that the reviewers found the idea novel (xyet, vtfm), conceptually appealing (76pw), and our experiments extensive (xyet, vtfm, 76pw). Their main concerns were about the robustness evaluation, the instruction-following protocol, WMDP leakage, and the role of asymmetric pretraining. We addressed each with new experiments, all included in the revision.

1. The fine-tuning attack lacked a positive control and generalizable capabilities evaluation (xyet, vtfm, 76pw).

  • Password-locked baseline. We locked the same 400 biographies behind a password in a non-TLM model (100% recall with the password and 0% without) and ran the same attack. The held-out recall reaches 35.6%, which shows the attack can indeed extract hidden facts. Note that TLM stays near 0% (Fig. 5, Sec. 6).
  • Attack on a generalizable capability. We fine-tuned CpubC_{pub} of TLM-650M on Spanish. We observe the same trend for this attack, i.e., TLMs are robust even for such generalizable capabilities (App. D.7).
  • Reviewer vtfm confirmed that these experiments addressed this concern.

2. Adaptive white-box attacks were missing (xyet, vtfm, 76pw). We added two adaptive attacks.

  • Greedy key search using loss objective over the correct answers. Over about 85K candidate keys, the recall stays near 0%, and the keys share almost none of the true swaps (Fig. 6, Sec. 6).
  • Gradient-based relaxation. We observe that NLL falls from 5.77 to 2.66 but stays far above that of CKC_K (0.11). The recall reached about 7% vs 99.8% with the true key (App. D.8).

3. The instruction-following evaluation (xyet, vtfm).

  • We clarified that the Figure 2 (right) is a head-to-head comparison, and we have now added evaluation wrt extensive external references (App. D.5, D.6).

4. WMDP leakage quantification and trade-off (xyet, vtfm).

  • We have now included a KL-weight sweep that shows leakage on seen questions falling with increasing KL penalty. Leakage reaches zero only at β≥0.7\beta \geq 0.7, where CKC_K drops to 51.7% or below (App. D.4).

5. Asymmetric pretraining was not compared with a symmetric version (76pw).

  • In a matched 180M ablation, we find that symmetric pretraining leaks about 25% more of the private capability into CpubC_{pub} (App. D.2).

We believe the new experiments run during the rebuttal address the concerns summarized above. If the reviewers agree, we would kindly ask them to consider updating their scores to reflect these results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.