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