acceptodds
Under review as a conference paper at ICLR 2027

When Are Safety Anchors Enough? Availability and Certified Preservation

Abstract

Safety anchors constrain fine-tuning, but a useful guarantee must connect available example gradients to the updates a provider actually permits. We study preservation under a single global parameter update chosen from an enforced region. A spectral nonlinear risk gate turns available-span geometry into explicit pool-expansion, constraint-budget, and abstention decisions. For noncentered Gaussian gradients, an exact scalar representation connects systematic drift to the same spectral feasible families. A complementary pointwise envelope supports conditional flip-rate calibration through fixed-sequence testing. We characterize why this envelope can be vacuous even when every global update is safe, and why average-risk control can permit complete failure on a rare category. Experiments test unplanted 128-dimensional pools, conflicting nonlinear tasks, disjoint handwritten-digit splits, and global curvature bounds through three hidden layers. The results distinguish larger retained-direction displacements from improved task progress: anchors can enable a prescribed protected direction yet greatly limit progress on a conflicting task. On real classifiers, a 2% conditional flip ceiling gives modest task gains and shows clear calibration and curvature limits. The contribution is an anchor-specific availability and certification analysis, with explicit successful and unsuccessful operating regimes.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
safety anchors, fine-tuning, risk certification, gradient geometry, statistical calibration

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ssD3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies when safety anchors built from sample gradients can provide certified retention guarantees during fine-tuning. It derives a spectral upper bound on the fraction of initially correct predictions that can flip after a single update within an anchor-constrained ball. Using a generalized Schur complement, it characterizes whether a given gradient pool can produce aggregate anchors meeting a spectral threshold, and derives minimum-rank requirements and single-task utility limits. It also combines pointwise envelopes with fixed-sequence calibration and identifies a quantifier gap showing why pointwise envelopes alone may fail to control global risk.

The experiments use 128-dimensional synthetic gradient pools, 128-unit tanh networks, even–odd classification on sklearn digits with 4,160 trainable parameters, and a collection of 27 small models.

Critical strengths

I like that the paper separates three questions that are easy to mix up: what directions the anchors can constrain, how much task progress is still possible, and what risk claims the evidence supports. Table 2 gives a concrete example: the anchored radius reaches 2.0, but the shift-2 task gain is only 17%, compared with 63% without anchors. A larger radius does not necessarily leave more room for useful learning.

I did not find any issues in the proof of Thm. 2, including the singular cases in A.1–A.2. The numerical examples I checked also match the reported values: the exact risk of 6.22e-48 and spectral bound of 0.97% in the 128-dimensional example, roughly 0.15 in the four-task example, and the calibration counts of 826 and 8,290 per group for the 1% and 0.1% limits.

I also appreciate that the authors show where things break down. They report radii shrinking to around 0.001 with frequent abstention at depths 2–3, the 0/10 result for group guarantees in unanchored fits, and no extra gain from anchors on real data. Section 9 is quite direct about what remains limited.

Critical weaknesses

My main concern is the gap between the LLM safety motivation and the experiments. The evaluation is limited to synthetic gradient pools and small networks. The certified radius already drops to around 0.001 at depths 2–3, which raises doubts about whether the guarantees would remain useful in the LLM setting.

The theoretical novelty also seems limited. The results combine standard Taylor/Markov bounds, Schur complements, and existing LTT calibration. The spectral condition can also be extremely loose: the unit-ball example has an exact risk of about 6e-48 against a 0.97% bound. It needs more justification as an objective for designing anchors.

I do not yet see a convincing example where the anchors help while providing a useful certificate. On digits, the designs show no clear advantage, while the trusted comparator achieves a larger task gain. In the attack experiment, even the unanchored ball at the same radius stays below a 2% flip rate. These results make it hard to see when the anchors are actually needed.

The presentation makes the argument harder to follow. Thm. 2 needs more geometric intuition and less unexplained notation. The difference between aggregate and per-sample anchors also deserves more attention, including a clear indication of which is used in each experiment.

Questions for authors
  1. Could you show a non-vacuous certificate on a pretrained transformer, even just one block of a small LM or its LoRA parameters? How would you estimate C̄ and L at that scale?

  2. How loose is the spectral bound in practice? Beyond the unit-ball example, could you compare it with the observed flip rate on digits?

  3. Is there a case where anchors give a larger task gain than the unanchored baseline while keeping the certificate non-vacuous? If not, what practical recommendation should we take from this framework?

  4. What would applying these anchors look like in an LLM safety setup? How would it differ from Lu et al. (2026), Safe LoRA, and Rosati et al. (2026)?

2 comments
AuthorsOpenReview ↗

Response to Reviewer ssD3 (1/2): the gap to LLMs (W1, Q1), a refusal probe on LLM features (E6), the spectral bound on digits (Q2)

Thank you for checking the proof of Theorem 2 and our numbers.

W1 and Q1: the gap to large language models (LLMs). We agree that this is the main limitation of our evidence, and the abstract and the second paragraph of the introduction now say so. We cannot give a non-vacuous certificate for a transformer block or for low-rank adaptation (LoRA) parameters. That needs a bound L on the margin Hessian over a whole parameter ball, through attention and LayerNorm, and we know of none that gives useful radii (Appendix E); Section 9 names this the main open problem. We prefer to say so rather than plug in a curvature sampled at a few points, which the guarantee does not permit.

Your reading of the depth study is right. Even among models that accept a radius, the mean certified radius is only 3.0·10⁻³ at depth 2 and 3.3·10⁻³ at depth 3 (Table 3, new last column). Table 3 certifies the unanchored ball, but anchors cannot remove the curvature term Lρ²/2 of Theorem 1 or its per-input form L(x,ρ)ρ²/2 in Eq. (7), since neither depends on the anchors. We attribute the collapse mainly to these terms; this is our reading, not a separate measurement.

A case the framework covers exactly (new E6, Appendix H.5). E6 trains a refusal probe, a linear head with 897 parameters, on frozen layer-16 features of Qwen2.5-0.5B-Instruct, so its margin is affine in the trainable weights and L = 0. The protected examples are 10,635 public harmful prompts, the task is to reduce over-refusal on OR-Bench-80k prompts, and we use 5 seeds, a 2% ceiling and δ = 0.05 per seed.

  • The pointwise envelope gives a non-vacuous certificate. Over-refusal on held-out prompts falls from 45.9% to 36.8%, while 57 of 7,378 harmful test evaluations flip (0.77%).
  • Anchors certify radii 1.6 to 5 times larger and flip fewer prompts, but they reduce over-refusal less. Under the comparison rule fixed before the run, the anchored design with the largest certified task support (the first-order task progress its certified region allows) leaves over-refusal 1.85 to 2.20 points higher than no anchors.
  • Refusal on three out-of-distribution harmful sets, which the certificate does not cover, drops slightly, from 96.3%, 91.4% and 94.0% to 95.1%, 89.0% and 92.1%.

The probe's decisions are not generated refusals, and since L = 0, E6 says nothing about the collapse with depth.

Estimating C̄ at scale. C̄ would come from Eq. (12) in Appendix B.2. It combines a matrix bound on the gradient's component in a q-dimensional basis, fixed independently of the calibration data, with a scalar bound on the energy outside that basis. Section 9 adds two caveats. The outside term is a trace, so Eq. (12) helps only when most gradient energy lies in q directions, and the matrix Bernstein bound, Eq. (11), needs an almost-sure bound on the gradient norm, which we lack for transformer margin gradients.

Q2: the spectral bound against observed flips on digits. The spectral bound cannot certify 2% on digits (new E5, Section 8 and Appendix H.4). With 563–583 calibration gradients and 4,160 parameters, the error term of Eq. (11) adds 28–35 to every eigenvalue of the estimated second-moment matrix, while the gate's threshold τ has median 0.28 even at the smallest radius, ρ = 0.025. At that radius, the certified bound averages 6.2% without anchors and 4.7% with 16 trace anchors. In the four fits where the envelope selects ρ = 0.025 for the unanchored region, benign adaptation flips 0.09% of base-correct test images on average and the strongest attack inside the region flips 0.13%, while the spectral bound at that radius is 6.4–7.7% (Appendix H.4). Across designs, the 2% ceiling is 12.5 to 40 times the mean strongest tested attack (Table 5), but attacks only lower-bound the worst case, so these ratios overstate the slack. In E6, the same error term exceeds τ at every radius in 3 of 5 seeds.

AuthorsOpenReview ↗

Response to Reviewer ssD3 (2/2): LLM instantiation (Q4), novelty and looseness (W2), when anchors help (W3, Q3), presentation (W4)

Q4: an LLM instantiation. Section 9 now describes one. θ0 is an aligned model with provider-exposed trainable parameters, such as LoRA factors. Q is a distribution of harmful prompts, the guarantee covers those the base model refuses, and the margin is a refusal-versus-compliance score. The pool holds the margin gradients of the provider's harmful prompts, and the provider enforces the region. Section 2 now compares the paper with the three methods you name.

  • Lu et al. (2026) softly penalize hidden-state changes on a few harmful prompts and evaluate this empirically. Our theorems hold for any enforced linear constraint, so they cover only a hard, linearized version of their anchors.
  • Safe LoRA projects the updates of selected layers onto the aligned-minus-base weight difference. It uses no data and states no guarantee. As a fixed region, it is one design that our gate could certify or reject.
  • Rosati et al. (2026) reparameterize open weights to raise the curvature of the harmful loss. They bound curvature and attack iterations, not a behavioural rate, so their work complements ours.

W2a: novelty. The tools are standard, as the new contributions paragraph in Section 1 now says. The main anchor-specific result is Theorem 2, which characterizes exactly the designs that a given pool can build and that pass the spectral gate κ ≤ τ, where κ is the largest gradient second moment along a free direction and τ a threshold set by the ceiling, radius and curvature. If T spans the directions no pool gradient reaches, the largest second moment along T, λ_max(TᵀC̄T), decides whether the pool must grow. Theorem 2 also gives the minimum number of constraints, k_S, and the best one-task utility, B_S, which k_S constraints attain.

W2b: looseness. The spectral bound is far above the exact risk, but the gap is smaller when measured in radius. In the unit-sphere example at a 2% ceiling, the population spectral gate admits radii up to 1.44, while the exact risk reaches 2% only at about 4.97 (Appendix B.4). The new Proposition 1 (Section 3) shows that second moments alone cannot close this gap: for affine margins with no base margin below γ, some gradient distribution with the given second moments attains the bound.

W2c: why κ. We use κ for three reasons. First, ρ²κ is exactly the largest mean squared first-order margin change that one permitted update can cause (Appendix F.1). Second, for zero-mean Gaussian gradients and a constant margin, the exact worst-case risk depends on the design only through κ, so κ ranks designs by exact risk. Third, for a nonzero Gaussian mean and a ceiling ε < 1/2, Eq. (9) shows that exact feasibility is again a top-eigenvalue test, on matrices that add a mean term.

W3 and Q3: when do anchors help? They help in one synthetic case (new E2, right half of Table 2, Appendix H.2), whose task inputs lie near a direction orthogonal to the protected one. At shift 2, the certified anchored regions gain 91.5–94.3%, against 63.3% without anchors, and no certified design flips a protected test input. The gain comes from the larger certified radius: the uncertified radius-2 ball gains 99.9% but flips 73.2% of protected inputs. At shift 0.4, where the small ball suffices, anchors are 0.3 to 0.4 points worse, yet the rule that maximizes certified task support still chooses them in all 10 seeds. This case favours anchors because the protected gradients are nearly rank-one, and it is the only variant we ran. No real-data experiment shows a clear gain: on digits, the support rule applied after the fact gains 0.56 ± 0.21 points (Section 8), and in E6 anchors reduce progress.

Your two other points also stand. The trusted final-model comparator gains more on digits (6.50%, Section 8), because it certifies only the model a run returns, a weaker contract. Same-radius unanchored attacks also stay below 2% (Table 5), so these attacks do not show that anchors are needed at the selected radii. Our practical recommendation is the boxed procedure in Section 9: certify the unanchored region first, and adopt anchors only if they raise certified task support.

W4: aggregate versus literal (per-sample) anchors. Section 2 now defines both. A literal anchor bounds the first-order margin change of one selected example, while an aggregate anchor bounds only a combination of such changes, which can cancel. In Section 6, two designs are aggregate: the frontier design, which is the Theorem 2 design that attains B_S, and pool principal component analysis (PCA). Sections 7 and 8 and the caption of Table 2 now state that those experiments use literal rows, and E6 uses both kinds (Appendix H.5). For Theorem 2, Section 3 now defines every symbol and adds a Geometry paragraph with a worked example.

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

The paper studies fine-tuning constrained to a region around a fixed base model, with constraints built from a pool of “anchor” gradients, and asks which regions can be certified under a ceiling on the fraction of previously correct protected inputs that flip. It proposes two certification routes (a global spectral bound and a pointwise envelope), argues that they fail in different regimes, and reports experiments on synthetic data, small tanh networks, and 8×8 digits. The stated contribution is a certification analysis for safety anchors, framed as a tool for deciding when to accept an update region, expand the anchor pool, or abstain.

Critical strengths
  • The paper is theoretically ambitious. Although the undefined terms and notation make it hard to verify, the paper provides a large number of theoretical contributions, with full proofs in the appendices

  • The paper is candid about its limitations, scope, and implications (Paragraph 9, “Attack Evidence, Limits, and Implications”): it acknowledges the small scale of the experiments, the absence of any LLM or transformer result, and that its global curvature bounds “do not imply useful radii” (L438).

Critical weaknesses

W1. Obscure introduction and unclear problem formulation

The abstract and the introduction are completely obscure. Familiar terms such as fine-tuning, global parameter update, task, classifier, flip, and adaptation are used, yet they seem not to have their regular meanings, or the overall meaning remains obscure.

Overall, the paper does not properly introduce its problem, does not clearly define what it means by a "safety anchor," and does not establish how its terminology relates to existing work.

Structuring the introduction around context, problem, relationship to existing work, and contributions would be useful for understanding this work, even if further details have to be placed in the appendix.

W2. Lack of a clear formal setup

L056–061: The entire paragraph is extremely dense and difficult to understand. I did not find a formal definition of the decision margin anywhere in the paper, including the appendices. Without such definitions, it is impossible to judge the soundness of the work.

None of the few references cited in the introduction use the authors' vocabulary or notation, so where is the reader expected to learn them?

W3. Bibliography issues

L039: The bibliography does not include the a/b suffixes for Qi (2024). The same issue occurs with Huang (2024) at L090. This makes it difficult to identify the corresponding references in the bibliography.

Questions for authors

Q1. L040: "The decision studied here is narrower." What do you mean by "decision" here? You were just discussing safety-sensitive projection methods.

Q2. L040: What is a "risk ceiling"? If you mean classifier risk, do you mean true or empirical risk?

Q3. L041: "...which update regions can a provider justify?" What is a "region"? Subsets of the pool of gradients? What is a "provider"? What do you mean by "justify"? Your research question is too vague and obscure, and nothing helps narrow it down to a unique meaning.

Q4. L042–043: "An independent trusted optimizer, which evaluates only its final model, solves a weaker problem and remains an important comparator." Why do you use the term "trusted" here? Are you suggesting a data-poisoning context? Why is the optimizer evaluating the model? A weaker problem compared to what?

Q5. L044–049: The language in this paragraph is also very fuzzy. After presenting the two routes, the last sentence is even more confusing. Why is there a mention of "a dimension-free binomial interval"?

Q6. L050–053: What is "benefit"? Are you using "radius" and "ball" interchangeably? It is absolutely unclear what you are announcing here.

Q7. L056–061: It is not clear what a "fixed base model" is, and "decision margin" implies a classifier. What is the setup here? Should we assume, for instance, that (\theta_0) denotes the parameters of that base model? How do you define that margin?

Q8. L056–061: What is the "main preservation contract"?

Q9. L056–061: If space was the issue, why didn't you include a formal setup in the appendix? Q10 Following the comments in weaknesses, I have compiled all the terms and formulas that are undefined or need clarification. Please can you address them?

1. Undefined or insufficiently explained terminology

TermLocationIssue
Safety anchors / anchor setTitle, Abstract, L35, L73–76No explicit definition of what constitutes a safety anchor in this work. The relationship to the notion introduced by Lu et al. (2026), which the authors cite, is not explained.
Safety preservationAbstract, Sections 1–2The paper uses safety-related terminology but ultimately formulates preservation in terms of previously correct predictions. The relationship between prediction preservation and safety preservation is not established.
Selected outputsL35The introduction refers to preserving selected outputs without explaining which outputs are selected, how they are selected, or what preservation means in this context.
ProviderL40–43, L63–70The provider's role and the assumed fine-tuning scenario are not concretely defined. It is unclear who constructs and enforces the constraints and who performs subsequent adaptation.
Available gradient pool / eligible poolL40, L73–76, Section 6, Table 1Although the gradient matrix is formally introduced, the criteria determining gradient availability or eligibility are not specified. The acquisition mechanism and its relationship to the protected distribution are unclear.
Trusted base modelL440The paper assumes a trusted base model without specifying the criteria establishing its trustworthiness or its safety properties.
Prescribed protected directionAbstract, L24–25This terminology is introduced without an explicit definition or a clear connection to the subsequently defined safety and task directions.
Trace selectionL306, Tables 2 and 4The trace-based anchor-selection procedure is not specified. The objective being optimized and the method used to select individual anchors are unclear.
QR selectionL306, Tables 2 and 4The paper mentions QR factorization but does not specify the anchor-selection algorithm, including whether pivoting is used or which matrix is factorized.
Pool PCAL269–275The PCA baseline is named, but its precise construction and how its components are converted into enforced constraints are insufficiently described.
Global-margin attack / targeted-even-margin attackL405–408The attacks are named and their optimization budgets are reported, but their precise objective functions are not provided.

2. Mathematical definitions and formulations requiring clarification

Mathematical object or formulaLocationIssue
mθ0(x)m_{\theta_0}(x)Section 2, L57–60The authors introduce a "specified scalar decision margin" without explicitly defining how it is computed from model outputs and labels, or how it represents safety-related behavior.
ν=Q(⋅∣m0>0)\nu=Q(\cdot\mid m_0>0)Section 2, L57–60The conditional distribution is formally defined, but the underlying protected distribution QQ is not concretely specified. In particular, its relationship to safety-sensitive inputs or behaviors remains unclear.
g(x)=∇θmθ0(x)g(x)=\nabla_\theta m_{\theta_0}(x), GGSection 2, L57–76Gradients and their matrix are introduced mathematically, but the process by which actual safety-anchor examples produce the available pool is not specified.
K(J,ρ,ζ)={Δ:∥Δ∥≤ρ, ∥JΔ∥≤ζ}\mathcal K(J,\rho,\zeta)=\{\Delta:\|\Delta\|\leq\rho,\ \|J\Delta\|\leq\zeta\}Eq. (1)The feasible update region is mathematically defined, but the mechanism by which it is enforced in practical fine-tuning is not explained. The choice of constraint tolerance ζ\zeta is also insufficiently motivated.
J=GIJ=G_I, J=FGJ=FGSection 2, L73–76; Theorem 2; Appendix A.4Literal and aggregate constraints are mathematically distinguished, but the practical procedure for constructing FF and selecting the corresponding constraints is insufficiently specified. In particular, aggregate constraints can allow cancellation of individual anchor-output changes.
C=Eν[gg⊤]C=\mathbb E_\nu[gg^\top], C⪯C‾C\preceq\overline CEq. (2), Theorem 1The second-moment matrix and its upper-bound requirement are defined. Appendix B provides theoretical estimation procedures, but their applicability and computational feasibility for realistic pretrained models are not established.
Ud(P)=∥Pd∥/∥d∥U_d(P)=\|Pd\|/\|d\|, SD(P,ρ)=ρmin⁡d∈DUd(P)S_D(P,\rho)=\rho\min_{d\in D}U_d(P)Eq. (2)The utility quantities are formally defined, but the construction of task directions dd and the task bank DD from actual downstream learning objectives is insufficiently explained.
ν(m0≤γ)≤b\nu(m_0\leq\gamma)\leq b, LLTheorem 1, Eq. (3)The theorem assumes a low-margin probability bound and a global Hessian bound. Although theoretical estimation and curvature arguments are provided, it remains unclear how sufficiently informative values are obtained for realistic pretrained models.
ϵ,δ,αc\epsilon,\delta,\alpha_cTheorem 3, Appendix C.2The risk ceiling, overall confidence budget, and per-chain allocations are formally introduced. However, the numerical confidence levels and allocation choices used for the reported calibration experiments are not completely specified.
Task gain (%)Tables 2 and 4Task gain is described as held-out loss reduction, but an explicit formula specifying its percentage normalization is not provided.
Attack optimization objectivesSection 9, L405–408No explicit mathematical formulation is provided for the global-margin or targeted-even-margin attack objectives, making their optimization and reproduction difficult to assess.

Overall concern: The manuscript introduces sophisticated mathematical machinery without first establishing a sufficiently clear and self-contained problem formulation. Several central terms, particularly safety anchors, available gradient pool, and safety preservation, lack the definitions or contextual explanations needed to understand the intended setting. Other mathematical objects are formally defined, but their relationship to concrete safety-preservation requirements and practical fine-tuning procedures remains unclear. These issues make it difficult to identify the precise problem being addressed, assess the novelty relative to existing methods, and reproduce some of the experimental comparisons.

3 comments
AuthorsOpenReview ↗

Response to Reviewer jMhy (1/3): what changed, and a self-contained setup (W2, Q7)

Thank you for the careful reading and for the two tables in Q10. We agree with your central point: the submission lacked a self-contained setup and never defined "safety anchor". For your review, the main changes are the new problem setup that opens Section 2 and the restored Appendix G, with a notation table (Table 6) and a per-experiment protocol table (Table 7).

Here is the setup in brief.

  • Roles. A provider, such as a hosted fine-tuning service, owns a base model with trainable parameters θ0, and users fine-tune it, possibly adversarially. The provider deploys θ0 + Δ only if the parameter change Δ lies in a region K that it enforces.
  • Margin. Q is a distribution of protected examples x. The protected behaviour holds at x when a scalar margin m_θ(x), chosen by the provider, is positive; for a binary classifier with logit f_θ and label y ∈ {−1, +1}, it is y·f_θ(x).
  • Flip risk. Let ν be Q restricted to the examples the base model gets right. A flip occurs when the updated model gets such an example wrong. The flip risk R_ν(K) is the largest flip probability over all single updates in K, with the update fixed before the example is drawn.
  • Anchors. Each example the provider holds (its pool) contributes a candidate row, its margin gradient at θ0. A design is a constraint matrix J whose rows are the safety anchors. A literal design uses rows of selected examples and bounds each one's first-order margin change by a tolerance ζ. An aggregate design uses combinations of rows, so individual changes can cancel.
  • Certificate. For a radius ρ, the region is K(J, ρ, ζ) = {Δ : ‖Δ‖ ≤ ρ, ‖JΔ‖ ≤ ζ} (Eq. 1). A certificate states that R_ν(K) ≤ ε for a ceiling ε fixed in advance, with probability at least 1 − δ over the calibration data.
  • Quantities in the theorems. C is the second-moment matrix of the margin gradients, and C̄ is an upper confidence bound on it. The number b bounds the fraction of base margins below a chosen level γ > 0, and L bounds the margin Hessian over the radius-ρ ball.

W2 and Q7 (vocabulary, margin, θ0). Appendix G.1 now maps our terms to existing work. Exact literal anchors correspond to orthogonal gradient descent, the pointwise envelope (a per-example bound on the worst margin change, Eqs. 6–7) to weight-perturbation certification, and Theorem 3, which calibrates the envelope, to Learn-then-Test calibration. Yes, θ0 is the base model's trainable parameter vector; it is fixed before calibration, and every change Δ is measured from it. The margin need not come from a classifier. Section 2 defines it in general, and Sections 6–8 state each experiment's margin: Section 7 uses the network output, and Section 8 uses y·f_θ(x) with y = +1 for odd digits. Theorems 1 and 3 use the margin only through its base values, its gradients and a Hessian bound.

If a step still cannot be checked from this setup, please point us to it; otherwise, we would be grateful if you would revisit your soundness assessment.

AuthorsOpenReview ↗

Response to Reviewer jMhy (2/3): W1, W3, Q1–Q6, Q8–Q9

W1: the introduction and its terms. We rewrote the introduction (Section 1). Fine-tuning, classifier and adaptation keep their usual meanings. Section 1 now defines a global update as one parameter change applied to every input, that is, one deployed model. Section 2 defines a flip and a task, and Section 7 defines a conflicting task as one that pushes protected margins toward zero.

W3: references. Thank you; this was a real error. The revised references distinguish Qi et al. 2024, on fine-tuning that compromises safety even when users do not intend it, from Qi et al. 2025 (ICLR 2025), on safety alignment beyond the first few tokens. They also print the suffixes of Huang et al. 2024a (Vaccine) and 2024b (Lisa), and list Booster under 2025.

Q1. The decision is the provider's: accept a region of updates, acquire more examples, or abstain and keep θ0. Section 1 now also states that we propose no new defence.

Q2. A risk ceiling is a number ε, such as 2%, fixed before calibration. It bounds the population flip risk, not the classifier's error or an empirical rate, and Section 2 now says so. It implies that the population error rate rises by at most ε(1 − R_base), where R_base is the base model's error rate (Section 8).

Q3. A region is a set of parameter changes Δ (Eq. 1), not a subset of the pool, which supplies only the anchor rows. "Justify" meant "certify", and the revision uses that word. The question is: for which (J, ρ, ζ) can the provider certify R_ν(K) ≤ ε, and when must it expand the pool or abstain?

Q4. No data poisoning is involved, and L042–043 was badly phrased: the evaluator, not the optimizer, checks the final model, as Section 2 now says. "Trusted" means that this evaluator assumes a benign training run and certifies only the model it returns. Our certificate covers every model in the region, including adversarially chosen ones. It therefore implies the evaluator's guarantee but not conversely, so the evaluator's problem is weaker.

Q5. We rewrote the paragraph (Section 1, third paragraph). The spectral route bounds the flip risk of each fixed update and then takes the worst case. The pointwise route flags each calibration example that some update in the region could flip and bounds the flagged rate with a binomial test. That test is "dimension-free" because it needs no more samples in higher dimension. Yet the flagged rate can be 1 while the flip risk is tiny: 6.22·10⁻⁴⁸ in the uniform-sphere example of Section 6 (Eq. 8 in Section 4 gives the Gaussian case).

Q6. Radius and ball are not interchangeable: the ball is {‖Δ‖ ≤ ρ}, ρ is its radius, and an anchored region is that ball cut by the anchor constraints. "Benefit" meant extra task progress, measured by the held-out gain and by the task support S_D (Eq. 2), the smallest first-order progress the region allows across a declared set of tasks. In Table 2, an anchored region of radius 2.0 need not contain the unanchored ball of radius 0.75, so we compare designs by progress, not radius.

Q8. The preservation contract is the set of conditions for our guarantee, and Appendix G.1 now lists its five parts. (i) The base θ0 is fixed. (ii) Every deployed model is θ0 + Δ with Δ in K. (iii) One model is chosen before a fresh example is drawn. (iv) The ceiling ε, any named groups (subpopulations of Q, such as even digits) and the pointwise route's chains (each a design with its increasing list of radii) are fixed independently of the calibration data. Spectral designs may use C̄, because Theorem 1 holds for all designs at once whenever C̄ and b are valid. (v) Then R_ν(K) ≤ ε with probability at least 1 − δ. "Main" distinguishes this contract from its version for a named group.

Q9. Space was not the reason; the omission was our mistake. The setup now opens Section 2.

AuthorsOpenReview ↗

Response to Reviewer jMhy (3/3): the two tables of Q10

Q10, first table. Each term is now defined in the revised PDF or removed.

  • A safety anchor is an enforced row of the constraint matrix J (Section 2). Lu et al. (2026) use the term for a soft penalty on hidden-state changes; our theorems cover only a hard, linearized version of their anchors.
  • For safety preservation, we preserve the sign of a provider-chosen margin. This is a safety guarantee only insofar as the margin encodes a safe behaviour, such as refusal when Q is a set of harmful prompts, and no experiment measures generated refusal (Section 1).
  • The selected outputs are the margins of the selected anchor examples; with literal rows, each changes by at most ζ to first order (Appendix G.1).
  • The provider builds the anchors from its pool and fixes, enforces and certifies K, and users fine-tune inside K (Section 2).
  • The available pool is the set of examples the provider holds before calibration, drawn from a distribution that may differ from Q (Section 2); "eligible" meant the same. Table 7 lists each pool, and coverage is tested: if a category with 1% of the mass is missing, its group certificate abstains until one of its gradients is added (Table 1).
  • The trusted base model θ0 is a fixed reference, not assumed safe; we certify degradation relative to it (Section 9, Appendix G.1).
  • "Prescribed protected direction" was unclear wording; the abstract now says that anchors certify larger update radii but can reduce progress on tasks that oppose the protected outputs.
  • Trace and QR selection add pool rows one at a time, after projecting the remaining rows off those already chosen (Appendix G.3). Trace adds the row whose normalized residual v maximizes the mean of (gᵀv)² over a separate design sample, and QR the row with the largest residual norm (column-pivoted QR of Gᵀ).
  • Pool PCA (principal component analysis) enforces as aggregate rows the top right singular vectors of the uncentred gradient matrix G, as many as Theorem 2's minimum rank or a fixed budget (Table 7). Section 6 also uses a literal counterpart, literal-B, which enforces the pool rows of the coordinates whose second moment exceeds the threshold τ.
  • Each attack runs projected gradient descent to solve min over Δ ∈ K of (1/N) Σᵢ yᵢ·f_{θ0+Δ}(xᵢ), the mean margin of the N base-correct training images, or of the even-digit ones for the targeted attack. It uses 250 steps and 3 starts (Section 9, Appendix G.7, Table 7).

Q10, second table. Each object is now specified in the revised PDF, and Table 6 renames the symbols that the submission used for two different things.

  • Table 7 gives the margin m, the distribution Q and the pool for every experiment except E6, which Appendix H.5 specifies.
  • Row i of G is the margin gradient g of pool example i at θ0 (Section 2).
  • In Sections 7–9, K(J, ρ, ζ) is enforced after every step by clipping: the part of Δ in J's row space is scaled until ‖JΔ‖ ≤ ζ, then Δ is scaled into the radius-ρ ball. Section 6 projects exactly (Section 2, Appendix G.3).
  • The tolerance ζ is a design choice. ζ = 0 makes the anchors exact and is used in Section 6, and ζ = 0.005, the only other value evaluated, is used in Sections 7–8, so sensitivity to ζ is untested (Appendix G.3).
  • Appendix G.4 constructs the aggregate matrix F step by step, and Appendix A.4 shows that the cancellation you point out is real.
  • Section 6 estimates C̄ from data, with simultaneous exact binomial (Clopper–Pearson) bounds in the pool and rare-mode studies and with the matrix Bernstein concentration bound of Eq. (11) in the sphere study and E3 (Table 7). Doing so at scale is an open problem, and Eq. (11) already fails at 4,160 parameters (E5).
  • Section 2 now explains U_d, S_D, d and the task bank D: a task's direction d is its negative loss gradient at θ0, and U_d is the fraction of d that the anchors leave free.
  • Table 7 gives b and L per experiment. Base margins are constant in the pool, rare-mode and sphere studies (Section 6) and in E3's two-band model, so b = 0 there. On learned networks, E3 and E5 bound b by Clopper–Pearson (Appendices H.3 and H.4), and the envelope of Sections 7 and 8 does not need b. L is exact for the synthetic models, comes from a one-layer bound or Appendix E for the tanh networks, and is unavailable for transformers.
  • Table 7 lists ε, δ and every α_c, and Appendix H.5 gives them for E6. On digits, δ = 0.05 per fit and α_c = 0.05/15 per chain.
  • Task gain is 100·(1 − ℓ_test(θ_final)/ℓ_test(θ0)) on held-out inputs (Table 2 caption, Appendices G.6 and G.7).
Reviewer_6rzrOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies when a set of safety examples can keep fine-tuning from turning previously correct predictions into errors. It uses the gradients of these examples to constrain model updates, and examines whether the available examples cover the directions that need protection. An interesting part of the analysis is how it connects the number of constraints needed for protection with the flexibility left to learn new tasks.

Critical strengths
  1. Theorem 2 adds feasible-space, minimum-rank and single-task utility characterizations beyond OGD/PCA-OGD. Its exactness concerns a spectral sufficient condition.

  2. The global-versus-pointwise risk distinction and negative experiments clarify when certificates become uninformative.

Critical weaknesses
  1. The main construction is not evaluated on learned models. Theorem 2 is tested on coordinate gradients with diagonal second moments, eliminating cross-space Schur coupling. Neural experiments instead use literal selectors and the pointwise envelope (Section 9). They therefore leave the operational value of the principal advance over existing projection methods untested.

  2. The neural results do not establish added adaptation value. Table 4 shows lower observed flips, from 0.17% to 0.009%, but no clear task-gain or worst-task-support advantage. Larger radii alone are insufficient because the allowed directions differ. These are useful preservation diagnostics, but cannot establish the benefit of Theorem 2's unused construction.

Questions for authors
  1. Can a learned-Jacobian case compare Theorem 2 against literal/PCA selectors under the same pool, constraint budget and certifier, reporting acceptance, task support and task gain? This would connect the novel theory to useful decisions. If unavailable, what prevents its application?

  2. Does pool PCA in Section 6 already constrain exactly the coordinates in B={j:C‾jj>τ}B=\{j:\overline C_{jj}>\tau\} when all required coordinates are available? If not, how does this literal baseline perform under the same spectral safety criterion and utility target? This would help separate the benefit of combining gradients from simply avoiding unnecessary constraints.

  3. What δ\delta, per-chain αc\alpha_c, and joint design/group families were used in Table 4? Appendix C.2 requires ∑cαc≤δ\sum_c\alpha_c\leq\delta; the 2% risk target does not specify confidence.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 6rzr (1/2): S1, W1, W2, Q2 (new E1), Q3

Thank you for a precise review.

S1. We agree that Theorem 2 is exact only relative to the sufficient spectral gate κ ≤ τ. Within that scope it is tight: for affine margins with no base margin below γ (L = b = 0), the new Proposition 1 (Section 3) shows that no certificate using only second moments can accept a design with κ > τ. So, for exactly enforced designs, Theorem 2 describes every design the pool can build that second moments can certify; with curvature or mean information, it is only sufficient.

W1. Your description is accurate, and the revision now says so. Section 3 states that we implement the aggregate construction in coordinate models, and Section 9 states that Sections 7–8 certify with the envelope and literal rows, so they do not test Theorem 2's construction. In the coordinate study of Section 6 (E1), C̄ is diagonal, so the coupling term vanishes. The new E3 (part 2 of this reply) applies the full construction to the Jacobians of a trained network, where coupling is negligible, so a learned model in which coupling matters remains untested.

W2. We agree: on digits, anchors show no clear advantage in task gain or worst-task support, and a larger radius alone does not establish one. Worst-task support is 3.70–4.40 with anchors against 4.50 without (Table 4, in units of 10⁻²), since a larger anchored region can exclude directions that the smaller ball allows. Of 11,376 base-correct test evaluations, unanchored adaptation flipped 19 and the anchored designs 1 to 6 (Section 8). These benign runs share one certified 2% ceiling, so the lower counts are not certified differences. In a reanalysis after the fact (Section 8), the Section 7 rule that picks the largest certified task support chooses anchors in 6 of 10 fits and gains 0.56 ± 0.21 points over no anchors.

Q2. Pool PCA constrains exactly B at pool size 256, but not always at smaller sizes (new E1, Section 6 and Appendix H.1). Here B = {0, 1, 2, 3} in all 20 seeds, and a trial passes the pool floor when its pool contains every coordinate (mode) in B. Among such trials, PCA's rows equal the coordinate vectors of B in 20 of 20 at size 256; in 13 of 17, 17 of 19 and 19 of 20 at sizes 32, 64 and 128; and in 2 of 14 at sizes 8 and 16.

  • Avoiding unnecessary constraints is not enough. Literal-B, the best literal design, enforces the rows of B and passes the risk gate whenever B is pooled, but its utility of 0.661 misses the 0.76 target in every trial.
  • The gain comes from combining gradients. The frontier design is the Theorem 2 design that reaches the one-task utility ceiling B_S with the fewest rows, k_S. Reaching 0.76 needs an aggregate row that mixes mode 0 with mode 8, a mode outside B along which the task direction has a large component. The frontier passes in exactly the trials whose pool contains mode 8: 13 of 13 at size 256, against 0 of the other 7. A pool of 256 contains mode 8 with probability only 0.41, so our 13 passes out of 20 are more than the expected 8.2.
  • The advantage holds only for targets between 0.66 and about 0.83. At targets of 0.50–0.65, the frontier, pool PCA and literal-B all pass in 20 of 20 trials, and at 0.85–0.95 none passes (Table 9). We did not record when the 0.76 utility target was chosen, so we do not claim that it was fixed in advance.
  • We also found and fixed a solver bug (E1b, Appendix H.1). With a repeated top eigenvalue, our solver returned 0.661 instead of the ceiling 0.687 in the trials without mode 8. A tie-aware solver reaches the ceiling in all 90 floor-passing trials, and no pass count or decision changes. E3 used the original solver but met no tie.

Q3. On digits, δ = 0.05 per fit is split as α_c = 0.05/15 ≈ 0.0033 over 15 chains, which pair each of 5 designs with the global contract and the even- and odd-digit groups, so Σ_c α_c = δ. Each chain is a fixed-sequence test over 12 radii at ε = 2%, so all guarantees within a fit hold simultaneously. With 264–301 calibration images per parity group, zero events give a Clopper–Pearson bound of 1.88–2.14% at this α_c. A group therefore passes only with zero events and at least 283 images, so at most 2 of 10 fits certify a positive parity radius (Section 8). Table 7 lists the split for every experiment except E6 (Appendix H.5).

AuthorsOpenReview ↗

Response to Reviewer 6rzr (2/2): Q1, Theorem 2 vs. literal and PCA selectors (new E3)

Q1. We ran this comparison as the new E3, on the Section 7 network (Section 6 and Appendix H.3). On digits it is blocked by sample size, not by computation. With 4,160 parameters and 563–583 calibration gradients, the error term of the matrix Bernstein bound (Eq. 11, Appendix B.2) adds 28–35 to every eigenvalue of the estimate. The threshold τ is at most 18 even at the smallest radius, so the spectral gate fails for every design. With the same gradient-norm bound and top eigenvalue, the error term would fall below τ only with about 1,200 calibration images in the most favourable fit, and with 1.4·10⁵ to 3.4·10⁷ in the other nine. This cost grows with the gradient-norm bound, which we lack altogether for transformer margin gradients (Section 9), so we expect it would also bind for large language models.

E3 setup. We use 10 seeds, estimate C̄ from 65,536 fresh gradients, and enforce the constraints exactly. Because the protected gradients are nearly rank-one, the unanchored ball passes the gate only up to ρ = 0.10, while designs that remove the top direction pass up to 0.45–0.60. We compare the frontier (part 1, Q2) with pool PCA at rank k_S or with 4, 8 or 16 rows; with trace, QR and random literal rows (4, 8 or 16); and with two greedy literal designs that add rows until the gate passes, κ-greedy by the largest κ reduction and task-aware greedy by the largest κ reduction per unit of utility lost. Each design is trained for 300 steps at its largest certified radius. Budgets are mostly unmatched: the frontier uses 1 to 16 rows and task-aware greedy 10 to 15, while the fixed designs match only one another. In the one matched comparison, pool PCA is rebuilt at the frontier's rank k_S at each radius, with the same pool and gate, and gains 53.1% against 72.6% (support 0.316 against 0.456).

We wrote the protocol before comparing any design, but after a look at seed-0 gate quantities that we used only to choose the radius grid. The frontier sits exactly at κ = τ by construction, so in a first seed-0 run our conservative bound on κ exceeded τ by about 10⁻⁶ and rejected it. We therefore switched the frontier's check to the exact κ and then ran all seeds once. The other designs keep the valid bound, which is conservative by at most about 9·10⁻⁷ in seed 0 (Appendix H.3).

E3 results. Every design is accepted at some radius (Table 11). Gains and certified supports at pool size 256 on the non-conflicting E2 task are below.

DesignGainCertified support
Frontier72.6 ± 1.3%0.456
Task-aware greedy68.1 ± 1.3%0.415
4-row designs62.8–65.7%0.373–0.396
8-row designs55.7–60.2%0.322–0.354
PCA at rank k_S53.1%0.316
κ-greedy46.7%0.269
16-row designs28–34%0.145–0.180
Unanchored19.9%0.100

Certified support orders all 17 designs exactly as gain does, and the frontier leads task-aware greedy by 2.8 to 4.6 points across pool sizes, with standard errors of at most 1.2. On the conflicting task, the unanchored design is best (19.9%, against at most 10.4% for anchored designs), and the frontier's support at its trained radius is only 0.037. Over all certified radii its support peaks at ρ = 0.15 with one row, averaging 0.119 and exceeding the unanchored 0.100 in all 10 seeds, but we did not train at that radius, so E3 does not show whether anchors help on this task. No design flipped any of the 10,000 protected test inputs.

Coupling. On these Jacobians, ignoring coupling changes no minimum rank and overshoots τ by at most 7.1·10⁻⁶. It matters in a synthetic two-band model with 8 high-variance and 56 low-variance gradient directions. In the 62 of 240 cases where its pool floor passes, the frontier passes the gate with the full C̄ in all 62, pool PCA at rank k_S in 57, and the construction that ignores coupling in only 31.

Caveat. On this network the spectral route is not the provider's best option: the envelope certifies the unanchored ball up to ρ = 0.75, which gains 95.4% on the same task, more than any spectral-certified design. E3 therefore ranks designs only within the spectral certifier, and we have no learned-model case in which the frontier improves the best certified outcome. In the E6 refusal probe (Appendix H.5), the spectral gate certifies nothing in 3 of 5 seeds. In the other two, the frontier has the largest support but no significant gain, and at the same radius the unanchored ball lowers over-refusal by 2.3–2.5 points, against at most 1.4 for the frontier, but flips 2–3 harmful test prompts, against 0–1. The spectral gate cannot certify this ball, although it lies inside the radius that the envelope certifies.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: scope, revisions and new results

Thank you for the careful reviews. We have uploaded a revised PDF, and each reply below names the section where a change appears. "L…" refers to line numbers in the submitted version.

Scope. We ask a narrow certification question. A provider, such as a hosted fine-tuning service, holds gradients of protected examples and caps how often an update may flip them from correct to incorrect. Which regions of parameter updates can it certify, and when must it collect more examples or abstain? All our evidence is small and controlled: we fine-tune no large language model (LLM), and the new experiment E6 trains only a linear probe on frozen LLM features. We give no certificate for transformer blocks or low-rank adaptation (LoRA) parameters, because we lack a global curvature bound for attention layers that permits useful update sizes (Appendix E). The abstract, the introduction and Section 9 now say so.

Presentation. All three reviewers rated the presentation 2 or lower, and we agree that the paper was hard to follow.

  • The abstract and introduction are rewritten, and Section 2 is a new, self-contained problem setup that defines safety anchors, the margin, the update region, the flip risk and the symbols that the theorems use.
  • Section 3 explains Theorem 2 (which anchor designs a pool can build) geometrically, with a worked example, and every main-text table and figure caption now names its study and explains its columns.
  • Appendix G restores the protocol appendix that the submitted PDF omitted by mistake. Table 6 lists the notation, and Table 7 gives each experiment's margin, pool, row type (literal or aggregate), selection rule, certifier, confidence levels and gain formula (E6's are in Appendix H.5).
  • Section 2 now compares the paper with Lu et al. (2026), Safe LoRA, Rosati et al. (2026), PCA-based orthogonal gradient descent (PCA-OGD) and Elmecker-Plakolm et al. (2026).

New results. Proposition 1 (Section 3, proof in Appendix B.4) shows that the spectral flip-risk bound of Theorem 1 cannot be improved for margins affine in the parameters when only the gradients' second-moment matrix is known. Appendix H adds five experiments (there is no E4). E1 and E3 test Theorem 2, E2 gives a task on which certified anchors help, E5 tests the spectral bound on digits, and E6 is a refusal probe on frozen Qwen2.5-0.5B-Instruct features. On digits, anchors certify larger radii but give no clear gain in task progress (Table 4), and in E6 they reduce it.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.