acceptodds
Under review as a conference paper at ICLR 2027

SAFECOMPOSE: Predicting Unsafe LoRA Compositions Without Exhaustive Testing

Abstract

Adapter hubs let practitioners compose several LoRA modules on one open-weight base, yet each adapter is safety-screened only in isolation. We formalise composition screening as learning a set function over an adapter pool, separating a threshold crossing (safe parts, unsafe whole), its mechanism (distinct-adapter interaction versus summed merge dose), and genuine third-and-higher-order Mobius mass, which alone defeats low-degree predictors (Theorem thm:separation, a conditional result). We pair this with a screener that scores candidate compositions from adapter fingerprints and a calibrated accept/test/reject decision with a distribution-free false-negative guarantee (Theorem thm:ltt). Empirically, the picture is largely negative for interaction: across compositions of individually screened public adapters on a B base, none crossed the deployment threshold ( upper bound under the sampled law), and sub-threshold hazard is node-predictable with no resolved gain from interaction features; a deliberately constructed collusion is explained by summed merge dose, not a distinct-adapter mechanism. On a constructed benchmark whose unsafe compositions have every constituent individually safe at its coefficient, a coefficient-aware screener with a nonlinear dose term reaches AUROC (node-only ) and transfers to a held-out risky pair (–), while interaction architectures add nothing; an exploratory calibrated accept region holds out of sample at . The lesson is that merge configuration, not only the adapter file, is part of the safety object. Code and measured records: https://anonymous.4open.science/r/SafeCompose-F316

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LoRA composition, model merging, LLM safety, adapter security, compositional safety, safety screening, risk calibration, model merging safety

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_2iEpOpenReview ↗
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 ∣C∣=∑k=2K(nk)|\mathcal C|=\sum_{k=2}^{K}\binom{n}{k} candidates (102,050 at n=40n=40, K=4K=4)?

The authors formalise the hazard of a composition as a set function r(S)=haz(W(S,c,⊕))r(S)=\mathrm{haz}(W(S,c,\oplus)) with Möbius representation r(S)=∑T⊆Sa(T)r(S)=\sum_{T\subseteq S}a(T). They carefully separate three notions (Remark 1): threshold emergence (every singleton safe, the composition unsafe), non-additivity (a(T)≠0a(T)\neq0 for some ∣T∣≥2|T|\ge2), and genuine third-and-higher-order mass Δ≥32(r)=∑∣T∣≥3r^(T)2\Delta_{\ge3}^{2}(r)=\sum_{|T|\ge3}\hat r(T)^2 in the pp-biased Fourier basis χT(S)=∏i∈T(zi(S)−p)/σ\chi_T(S)=\prod_{i\in T}(z_i(S)-p)/\sigma.

The formal contributions are:

  • Theorem 1: every degree-≤2\le2 predictor h∈H2h\in\mathcal H_2 satisfies ∥r−h∥2≥Δ≥32(r)\lVert r-h\rVert^2\ge\Delta_{\ge3}^{2}(r), hence μp(∣r−h∣>τ)≥Δ≥32(r)−τ2\mu_p(|r-h|>\tau)\ge\Delta_{\ge3}^{2}(r)-\tau^2.
  • Proposition 1 and Theorem 2: a truncation bound under coefficient decay ∣a(T)∣≤Cγ∣T∣|a(T)|\le C\gamma^{|T|}, and a capacity statement for degree-dd hyperedge models.
  • Theorem 3: a Learn-then-Test certificate, built on Hoeffding–Bentkus pp-values, for the accepted-set false-negative rate FNRacc(λ)=P[Y(S)=1∣R^(S)≤λ]\mathrm{FNR_{acc}}(\lambda)=\mathbb P[Y(S)=1\mid\hat R(S)\le\lambda].
  • 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 R^(S)=ς(b∅+∑igθ(1)(xi)+∑i<jgθ(2)(xij)+∑i<j<kgθ(3)(xijk))\hat R(S)=\varsigma\big(b_{\varnothing}+\sum_i g_\theta^{(1)}(x_i)+\sum_{i<j}g_\theta^{(2)}(x_{ij})+\sum_{i<j<k}g_\theta^{(3)}(x_{ijk})\big) 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 τ=0.30\tau=0.30 (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 α=0.10\alpha=0.10 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 r(S)=0.02+0.08∣S∣r(S)=0.02+0.08|S| makes every size-4 set emergently unsafe while every a(T)a(T) with ∣T∣≥2|T|\ge2 is zero. Likewise, the remark that a nonlinear function of a summed merge dose, r(S)=F(∑i∈Scivi)r(S)=F\big(\sum_{i\in S}c_iv_i\big), 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 r2r_2 errs by exactly a({1,2,3})=0.51a(\lbrace1,2,3\rbrace)=0.51, while the L2L^2-optimal projection r≤2r^{\le2} 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 2c2c 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 pp-value is non-monotone in mm, so a naive binary search returns an unattainable calibration size, and that Hoeffding sizing over-estimates the required mcalm_{\mathrm{cal}} by a factor of about 6 at α=0.05\alpha=0.05 and about 30 at α=0.01\alpha=0.01.

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 τ\tau.
  • 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 +0.002+0.002 (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 ∣C∣=1177<2000|\mathcal C|=1177<2000, "a direct 20-probe composition test reaches AUROC ≈1.0\approx1.0". The claimed execution savings at ∣C∣=105|\mathcal C|=10^5 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 λ\lambda 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 pp-biased Fourier expansion together with Parseval's identity, ∥r−h∥2=∑∣T∣≤2(r^(T)−h^(T))2+∑∣T∣≥3r^(T)2\lVert r-h\rVert^2=\sum_{|T|\le2}(\hat r(T)-\hat h(T))^2+\sum_{|T|\ge3}\hat r(T)^2; 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 Δ≥32\Delta_{\ge3}^{2} is non-negligible, and the paper itself shows it is not. Under Assumption 2 with p=κ/np=\kappa/n, Remark 2 gives Δ≥32=O(p3)\Delta_{\ge3}^{2}=O(p^{3}) for a pure order-3 conjunction. In the paper's own worked instance (Appendix C.4), the bound yields only μp(∣r−h∣>0.05)≥1.6×10−3\mu_p(|r-h|>0.05)\ge1.6\times10^{-3}, "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 H2\mathcal H_2 or H3\mathcal H_3, 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 α∈{0.01,0.05}\alpha\in\lbrace0.01,0.05\rbrace and a floor of mcal≥1000m_{\mathrm{cal}}\ge1000. The realised run deviates from that protocol in three ways:

  • it uses mcal=647m_{\mathrm{cal}}=647 and reports α=0.10\alpha=0.10;
  • at the pre-registered α=0.05\alpha=0.05, 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 τ=0.30\tau=0.30 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
  1. 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?

  2. Lower thresholds and other hazard dimensions. The maximum observed composition hazard was 0.235 against τ=0.30\tau=0.30, and only the HarmBench harm/compliance dimension was measured. Would E1 change at τ∈{0.15,0.20,0.25}\tau\in\lbrace0.15,0.20,0.25\rbrace, 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.

  3. 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.

  4. 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 cc-sum and cc-sum$^{2}$ features?

  5. 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 ciΔi=ci(αi/ri)UiVic_i\Delta_i=c_i(\alpha_i/r_i)U_iV_i, so the coefficient alone ignores the adapter's rank rir_i, scaling αi\alpha_i 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 ∑iciΔi\sum_i c_i\Delta_i?

  6. Calibration at the pre-registered operating point. Could you report the Learn-then-Test certificate at α=0.05\alpha=0.05 with mcalm_{\mathrm{cal}} 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?

  7. Operational role of Theorem 1. Given Δ≥32=O(p3)\Delta_{\ge3}^{2}=O(p^{3}) under Assumption 2, the bound of 1.6×10−31.6\times10^{-3} in Appendix C.4, and the fact that neither the baselines nor Eq. (10) lie in Hd\mathcal H_d, what does Theorem 1 tell a practitioner? Could you state a decision-level result instead, for example a lower bound on FNRacc\mathrm{FNR_{acc}} for any degree-2 screener under an explicit margin condition?

  8. A measured cost advantage. Every reported saving is arithmetic from Eq. (2), and at the evaluated scale a 20-probe direct test achieves AUROC ≈1.0\approx1.0. Could you run one experiment with ∣C∣≥104|\mathcal C|\ge10^{4} that reports end-to-end wall-clock cost, including fingerprinting (nmnm forward passes), adapter loading and merging I/O, and judge calls, against exhaustive short-probe testing at matched recall?

  9. 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?

  10. 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.

Reviewer_EmVCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

In this paper, the authors propose SafeCompose to identify potentially unsafe combinations of LoRA adapters without testing every possible composition. The method uses adapter fingerprints to predict safety risks and makes screening decisions through a calibration process. Experiments show that none of the tested public adapter compositions become unsafe, while the method performs well on a constructed benchmark using a coefficient-aware risk predictor.

Critical strengths
  1. I reckon the importance of the problem. Predicting safety risks from LoRA compositions without exhaustive evaluation is a meaningful direction.

  2. I also think the distinction between threshold emergence and higher-order interactions is interesting. The authors also provide theoretical analyses and controlled experiments.

Critical weaknesses
  1. I am not fully convinced by the practical relevance of the experimental results. None of the 500 compositions of public adapters becomes unsafe, while the positive results mainly rely on constructed canary-based backdoors.

  2. There is a substantial gap between the proposed framework and the actual implementation. The main experimental predictor is a relatively simple coefficient-aware model with a quadratic dose term, while the full interaction hypergraph and several key components remain unvalidated.

  3. Moreover, interaction features provide almost no improvement (+0.002 AUROC), making the necessity of the proposed architecture unclear.

  4. The claimed efficiency and safety guarantees are not sufficiently supported experimentally. The main screening experiment executes approximately 85% of the candidate compositions, while direct testing achieves nearly perfect performance.

Questions for authors
  1. Could the authors provide an end-to-end evaluation of the full hypergraph framework, including active acquisition and the proposed interaction features?

  2. Could the authors demonstrate actual computational savings on larger adapter pools where exhaustive testing becomes expensive?

  3. How reliable is the calibrated screening under distribution shifts or adaptive adversaries? Could the authors provide stronger results at stricter safety thresholds?

Reviewer_xbewOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies safety screening for compositions of individually screened LoRA adapters. It formulates composition hazard as a set function and distinguishes threshold crossings, non-additivity, and higher-order interaction. The theoretical results characterize the approximation limits of strictly low-degree predictors and provide a calibrated acceptance rule controlling the unsafe fraction among accepted compositions under independent, identically distributed calibration. The proposed screening framework combines adapter fingerprints, interaction features, budgeted testing, and accept/test/reject decisions.

Critical strengths
  1. Clear formulation with conditional theoretical guarantees. The paper models LoRA composition risk as a set function, distinguishing threshold crossings, merge-dose effects, and higher-order interactions. Under explicit assumptions, it characterizes low-degree predictors’ limitations and provides statistical guarantees for calibrated acceptance.

  2. Practical motivation for efficient screening. Exhaustive safety testing becomes expensive as LoRA combinations grow. The proposed framework reuses adapter fingerprints, predicts composition risk without executing each candidate, and allocates a limited testing budget while reserving independent data for calibration. This provides a promising direction for reducing costly model merges and safety-suite executions in large adapter pools.

Critical weaknesses
  • The connection between the theoretical analysis and real-world risky LoRA compositions is limited. (1) The assumption supporting low-order approximation is insufficiently validated. Proposition 1 bounds truncation error by assuming that higher-order Möbius coefficients decay with interaction order. However, the paper does not establish that this decay yields a useful error bound for real LoRA compositions. Some failures may depend on a specific combination whose higher-order contribution remains substantial. (2) The fingerprints’ ability to capture relevant risk information is insufficiently validated. Fingerprints compress adapter weights into a limited set of statistics, potentially discarding information that determines composition risk. The experiments mainly cover one constructed mechanism and do not establish whether these features capture a broader range of compositional failures.

  • Experiments do not sufficiently establish and mainly evaluate a single type of constructed LoRA risk. (1) The unsafe examples come from two canary pairs trained with the same recipe. It remains unclear whether the screener generalizes to other attacks, such as Colluding LoRA, LoRango, LoRATK and MergeBackdoor. (2) Base-model coverage is limited. The measured experiments mainly use Qwen2.5-7B-Instruct. It remains unclear whether the fingerprints and learned risk patterns transfer to other model families. Using different safety judges does not substitute for evaluating different base models. (3) Generalization to larger compositions is not established. A pool of nn LoRAs admits (nm)\binom{n}{m} compositions of size mm. Sharing three MLPs keeps the parameter count fixed, but does not establish reliable prediction across this growing space, particularly when mm exceeds the modeled interaction order d=3d=3 . Theorem 2 provides a conditional representation result, while empirical support for generalization as mm increases remains limited.

Questions for authors
  1. Support for low-order approximation. What evidence supports the coefficient-decay assumption in Proposition 1? Can the authors provide estimates or measurements showing that the resulting truncation bound is informative at the chosen order d=3d=3 and evaluated composition sizes? The absence of gains from interaction features alone does not establish that higher-order coefficients are small.
  2. Fingerprints and unseen risk mechanisms. Can the authors evaluate transfer to qualitatively different composition-risk mechanism, such as Colluding LoRA, LoRango, LoRATK and MergeBackdoor, while verifying that each constituent passes the same singleton screen at its deployed coefficient? This would help determine whether the fingerprints capture generalizable risk information or signatures specific to the canary training recipe.
  3. Transfer across base models. Which fingerprint components and learned predictors are specific to Qwen2.5-7B-Instruct, and what must be recomputed or retrained for another model family?
  4. Please specify the training objective and optimization procedure for Eq. (10), including how residual training is implemented. How are initial labeled compositions and subsequent batches selected, and what exact uncertainty score, diversity penalty, and curriculum schedule are used?

Comments on the paper

2 comments
AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

Summary of the rebuttal and request for consideration

Dear Area Chair and Senior Area Chairs,

We write briefly as the discussion period closes.

Every request has been addressed. We have responded to every question and weakness raised by all three reviewers, point by point, and completed the remaining items in the revision. Where a request cannot be met, we say so and explain why.

Engagement so far. Reviewer 2iEp engaged in detail. They confirmed that two of their central questions (calibration protocol, verification) were fully answered and raised three follow-ups, all of which we have since answered with pre-registered experiments. Reviewers xbew and EmVC have not yet responded to our replies. We would therefore be grateful if their concerns could be weighed against the revised paper rather than the submitted version.

What the revision changes:

  • Framing: the paper is reframed around what is validated, with a new title and abstract centered on a node-feature + merge-dose screener. Interaction heads and active acquisition are withdrawn, and the theory is moved to the appendix.
  • Calibration: it now meets the paper's own protocol at α=0.05\alpha = 0.05 and 0.010.01 (8/8 seeds, held-out FNR 0.019 / 0, on 3355 fresh configurations).
  • Public adapters: per-adapter screening gives no priority to the compositions that cross under over-weighting, while the recommended screener ranks them in its top 5%.
  • New experiments: eleven in total, three pre-registered. Every number traces to a logged script, and two errors in our earlier replies are corrected openly.

The revised PDF and a point-by-point response letter are uploaded. We respectfully ask that the substantial revision be taken into account in the final assessment. Thank you for overseeing the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.