- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
In my reading, the paper studies a timely safety problem in the LoRA ecosystem: adapters on public hubs are safety-screened one at a time, yet practitioners compose several of them on one open-weight base through weighted addition, task arithmetic, TIES/DARE-style merging or routing. Can a composition be unsafe when every constituent passes in isolation, and can such compositions be found without executing all candidates (102,050 at , )?
The authors formalise the hazard of a composition as a set function with Möbius representation . They carefully separate three notions (Remark 1): threshold emergence (every singleton safe, the composition unsafe), non-additivity ( for some ), and genuine third-and-higher-order mass in the -biased Fourier basis .
The formal contributions are:
- Theorem 1: every degree- predictor satisfies , hence .
- Proposition 1 and Theorem 2: a truncation bound under coefficient decay , and a capacity statement for degree- hyperedge models.
- Theorem 3: a Learn-then-Test certificate, built on Hoeffding–Bentkus -values, for the accepted-set false-negative rate .
- Proposition 2: active acquisition breaks the exchangeability that Theorem 3 needs; two repairs are offered (a reserved i.i.d. split and Horvitz–Thompson reweighting).
- Proposition 3: a recall bound at a fixed testing budget.
The proposed system, SafeCompose, fingerprints each adapter in weight space (singular spectra, projection onto the Safe-LoRA alignment subspace, layer localisation) and activation space (refusal-direction displacement, linear probes). It scores hyperedges with a degree-3 tied-parameter model and returns an accept/test/reject decision.
Empirically, on Qwen2.5-7B-Instruct with 37 individually screened public adapters, none of 500 compositions crosses the threshold (95% Wilson upper bound 0.76%). Node-only features predict sub-threshold hazard as well as pairwise or triple features do. On a constructed benchmark of 1,177 canary-backdoor compositions at eight merge doses, a coefficient-aware logistic score with a quadratic dose term reaches out-of-fold AUROC 0.993 on the Definition-1 eligible population (node-only: 0.87), and a Learn-then-Test accept region certifies at in 8 of 8 seeds. The authors conclude that "merge configuration, not only the adapter file, is part of the safety object".
Critical strengths
Strength 1. In my opinion, the conceptual formalisation is the paper's most valuable contribution, and it should become standard vocabulary for composition safety. I find Definition 1 and Remark 1 genuinely clarifying. The observation that threshold emergence, non-additivity and higher-order Möbius mass are three different things, and that only the last one obstructs a degree-2 predictor, cleanly dispels a confusion I have seen in the model-merging safety literature. The authors' own example makes the point well: the purely additive hazard makes every size-4 set emergently unsafe while every with is zero. Likewise, the remark that a nonlinear function of a summed merge dose, , can carry Möbius coefficients of every order is an important caution against over-reading "interaction" from a dose effect. The fully worked three-adapter instance in Appendix C is exemplary. It shows that the Möbius truncation errs by exactly , while the -optimal projection still flags the triple, a distinction the authors are careful not to blur.
Strength 2. I would like to commend the authors for an unusually honest, pre-registered and self-critical evaluation, including careful statistical engineering. From my perspective, the paper does several things right that are rare in the LLM-safety literature:
- it states falsification conditions in advance (Appendix G.6–G.7) and reports a negative headline result (E1: 0/500) rather than enlarging the constructed positive class until the numbers become interesting;
- it runs a dose-matched control in E2, showing that a single adapter at coefficient reproduces the crossing, so the collusion needs no distinct-adapter mechanism;
- it proves that active acquisition invalidates the calibration (Proposition 2(i)) and reserves an i.i.d. split accordingly;
- in Appendix J it shows that the Hoeffding–Bentkus -value is non-monotone in , so a naive binary search returns an unattainable calibration size, and that Hoeffding sizing over-estimates the required by a factor of about 6 at and about 30 at .
I learned something concrete from Appendix J, and I think the field benefits from this level of candour.
Critical weaknesses
Weakness 1. In my view, the empirical programme does not support the paper's framing. The headline machinery (higher-order Möbius mass, the hypergraph screener, active acquisition) is never exercised, and by the authors' own pre-registered criterion the problem as posed is not shown to exist outside adversarial construction.
The title, abstract and Section 4 are built around the claim that per-adapter screening misses hazard carried by third-and-higher-order interaction. Yet, as far as I can tell, every measurement points the other way:
- E1: 0 of 500 compositions of independently trained public adapters cross .
- E3/E4: node-only features match pairwise and higher-order features (rank correlation 0.71, 0.71, 0.66), and the paired bootstrap difference straddles zero.
- E2: the constructed collusion is explained by summed merge dose, not by a distinct-adapter mechanism.
- E6: the single decisive component is a quadratic dose term. Removing it costs 0.09 AUROC, whereas adding pairwise interaction features changes the score by (Table 8).
The conclusion itself concedes that Theorem 1 is "neither confirmed nor refuted here". I would point the authors to their own Appendix G.7: "If E1 returns an emergent unsafe rate indistinguishable from zero on independently trained adapters, the screening problem as posed is not a live one outside adversarial construction, and that is the finding." Is that not precisely what E1 returned? If so, I think the paper should be framed as a measurement study with a negative result, not as a screening system.
Moreover, the method that is actually evaluated is not the method that is specified:
- The model. The deployed E6 predictor is a five-fold logistic regression on weight-space node fingerprints plus coefficient-sum features ("the full interaction hypergraph of Section 5 is a broader design not separately validated here").
- The fingerprints. The activation-space fingerprint block, which the paper motivates at length through the refusal direction of Arditi et al. (2024), is not part of the realised run (Appendix G.8: "Results are for one 7B base and weight-space fingerprints").
- The remaining stages. The uncertainty/diversity acquisition rule (Stage 3), the repair stage SafeAlternative (E8), cross-base transfer (E5) and the adaptive author (E7) all remain unrun.
What remains, in my assessment, is a coefficient-aware logistic regression evaluated on one positive-class mechanism: a harmless canary backdoor from a single training recipe, 4 backdoor and 4 benign adapters at 8 dose levels, 462 eligible configurations. Of the four positive-class generators specified in Appendix F.2, only the merge-activated backdoor was built. The interference-driven class and the Ding (2026)-style refusal-suppressing collusion, which are the ones closest to real harm, were not. The "held-out risky pair" transfer (AUROC 0.93–0.95) is across two independently trained pairs of the same recipe.
Finally, the setting in which E6 is run is one in which, by the authors' own cost model (Table 9), screening is not needed. At , "a direct 20-probe composition test reaches AUROC ". The claimed execution savings at are "arithmetic from Eq. (2) rather than a measured speed-up". I therefore do not think the paper currently demonstrates that SafeCompose is useful in any regime it has actually measured.
I also do not find the central practical lesson, that "a LoRA's safety status is not portable across merge configurations", particularly new. Merge coefficients are already known to govern whether a backdoor survives merging (Arora et al., 2024; Zhang et al., 2024, BadMerging), and the scaling coefficient is a first-class hyperparameter in task arithmetic precisely because behaviour is non-linear in it.
Weakness 2. I believe the theory is largely standard and, by the authors' own computations, vacuous in the regime of interest, and the calibration demonstration does not meet the paper's own protocol.
(a) Theory. Lemma 1 is the Möbius inversion on the Boolean lattice. Lemma 2 and Theorem 1 are the -biased Fourier expansion together with Parseval's identity, ; this is textbook analysis of Boolean functions. Theorem 2(i) follows directly from uniqueness of the Möbius representation, 2(ii) is the observation that tied parameters cannot separate identical feature vectors, and 2(iii) is imported from Wagstaff et al. (2019). Theorem 3 is Learn-then-Test (Angelopoulos et al., 2025) applied to the reduction of Lemma 3. Proposition 2(iii) is a Horvitz–Thompson estimator with Hoeffding's inequality, and Proposition 3 is an elementary ROC-area argument. Each piece is correct as far as I can check, but I do not see a technical contribution that would carry an ICLR paper on its own.
More importantly, Theorem 1 bites only when is non-negligible, and the paper itself shows it is not. Under Assumption 2 with , Remark 2 gives for a pure order-3 conjunction. In the paper's own worked instance (Appendix C.4), the bound yields only , "a statement too weak to be useful". In addition, Appendix B.2 concedes that neither the baselines nor SafeCompose's own linked model (Eq. 10) belong to or , so the theorem constrains neither. Finally, it bounds estimation error, not decision error. Taken together, I struggle to see what operational role Theorem 1 plays.
(b) Calibration. The protocol (Appendix G.4–G.5, J.3) fixes and a floor of . The realised run deviates from that protocol in three ways:
- it uses and reports ;
- at the pre-registered , the certificate is non-empty in only 1 of 8 seeds (full) and 0 of 8 (eligible);
- the calibration draws are made with replacement from a fixed pool of 647 configurations, each executed once, so the certificate covers that pool's empirical law. That is a statistical resample of labels already in hand, as the authors acknowledge.
I appreciate the transparency, but in my view this demonstrates the mechanics of Learn-then-Test, not a deployable false-negative guarantee.
(c) Measured hazard. The realised E1/E4 hazard is the HarmBench harm/compliance dimension alone, judged by the Llama-2-13B HarmBench classifier, not the six-dimensional hazard of Section 3. Conclusions below are explicitly not determined, even though the median composition hazard rises from 0.05 to 0.09 and reaches 0.235.
(d) Presentation. Protocol and measurement are interleaved throughout, so the reader repeatedly has to work out whether a stated experiment (E3, E5, E7, E8, the sixteen ablations of G.3, the probe-shift and alignment-removal stress tests of E.4) was run. I think the main text should state, in one place, exactly which experiments were run and which are only specified.
Questions for authors
-
Your own falsification criterion. Appendix G.7 states that if E1 returns an emergent unsafe rate indistinguishable from zero on independently trained adapters, "the screening problem as posed is not a live one outside adversarial construction". E1 returned 0/500. How do you reconcile this with framing SafeCompose as a screening system rather than reporting a negative measurement study?
-
Lower thresholds and other hazard dimensions. The maximum observed composition hazard was 0.235 against , and only the HarmBench harm/compliance dimension was measured. Would E1 change at , or on over-refusal (XSTest), privacy extraction or tool-use safety? I would also like to see E1 on a second base family (e.g., Llama-3-8B-Instruct or Mistral-7B-Instruct) and with more than 37 adapters.
-
A genuinely harmful positive class. Could you build at least the collusion generator of Appendix F.2, i.e., a pair of LoRA adapters trained jointly so that their linear merge suppresses refusal as in Ding (2026), and report node-only versus coefficient-aware versus interaction-aware AUROC on it? At present every positive is a harmless canary from one recipe, so I cannot tell whether the "dose, not interaction" conclusion is a property of the problem or of the chosen mechanism.
-
What exactly is SafeCompose? Since the hypergraph model of Eq. (10), the activation-space fingerprints and the acquisition rule are all unvalidated, could you state precisely which components a practitioner should deploy today? Is the recommended method simply logistic regression on weight-space node fingerprints plus -sum and -sum$^{2}$ features?
-
Defining "dose" for LoRA. You report that coefficient-sum matching and update-norm matching give different verdicts in the dose-matched control (Appendix G.8). For LoRA, the effective update is , so the coefficient alone ignores the adapter's rank , scaling and spectral norm. Which notion of dose do you recommend? How would it extend to non-linear merge operators such as TIES (trim, sign election) and DARE (random drop and rescale), where the merged update is not ?
-
Calibration at the pre-registered operating point. Could you report the Learn-then-Test certificate at with set from your own Table 5 (Hoeffding–Bentkus), using fresh, independently executed calibration compositions rather than a with-replacement resample of a fixed 647-configuration pool? In what sense does the current certificate support a deployment decision?
-
Operational role of Theorem 1. Given under Assumption 2, the bound of in Appendix C.4, and the fact that neither the baselines nor Eq. (10) lie in , what does Theorem 1 tell a practitioner? Could you state a decision-level result instead, for example a lower bound on for any degree-2 screener under an explicit margin condition?
-
A measured cost advantage. Every reported saving is arithmetic from Eq. (2), and at the evaluated scale a 20-probe direct test achieves AUROC . Could you run one experiment with that reports end-to-end wall-clock cost, including fingerprinting ( forward passes), adapter loading and merging I/O, and judge calls, against exhaustive short-probe testing at matched recall?
-
Adaptive authors and the dose finding. If the decisive signal is a non-monotone function of merge dose, an adversary need not evade the fingerprint map at all; it can train an adapter that is inert at the screening coefficient and harmful at the coefficients users commonly choose. How does SafeCompose handle the fact that the composition coefficient is chosen by the downstream user rather than by the hub? Should the certificate quantify over a coefficient range rather than a grid point?
-
Verification of LLM-executed experiments. The LLM-use statement says that an assistant "implemented and ran" the experimental pipeline, including adapter training, vLLM generation and HarmBench judging, and drafted the proofs. Could you describe which steps were independently re-run or audited by the authors? Several cohort numbers in the manuscript (500, 512 and 484) required a reconciliation note in Appendix G.8.
Official Announcement by The Authors
Dear Chairs and Reviewers,
Thank you for your reviewing service. The authors have provided summaries, rebuttals, and clarifications on November 08. Please confirm that you have read them, state which of your concerns are or are not addressed, and whether that changes your current recommendations.
Thank you.