acceptodds
Under review as a conference paper at ICLR 2027

Saturated Screening, Concentrated Cost: What a Full-Stack Ablation Table Reads in a Redundant Refusal Stack

Abstract

Safety teams assemble refusal stacks — released guard models alongside the assistant’s own refusal behaviour — composed so that any one component can block, then decide what the stack is worth by ablating it: switch each component off in turn and see what breaks. We ask what that table measures. We assemble 9 refusal sources (8 released guard models from 5 vendors, plus the assistant’s own refusals), run each guard model at the operating point its vendor ships, and enumerate every one of the 512 subsets over 3937 prompts from 4 suites, so each source’s exact Shapley value sits beside its leave-one-out score with no coalition sampling (attribution figures: main suite). The harmful side is saturated: the full stack passes no harmful prompt on any of the 3 suites that contain harmful prompts, and not one has a sole blocker, a mean of 8.56 [8.46, 8.64] of the 9 sources blocking each StrongREJECT prompt. Switching any single source off therefore changes not one harmful decision at these shipped operating points, so every full-stack leave-one-out number here varies only through over-refusal. A zero full-stack removal marginal need not imply zero coalition-averaged contribution: on the harmful side the exact Shapley values sum to -1.0000 while the leave-one-out entries sum to 0.0000, and ablation sends 5 of the 9 to exactly +0.000 while those same sources carry 71.5% [65.0, 76.5] of under that normalisation, and the loss-minimising subset is drawn from them (WildGuard-7B alone in the unrestricted game, at 11.30 [7.84, 18.51]x lower loss than the full stack, though it passes 20 harmful prompts). So a one-shot rule deleting that column’s zeros removes the sources that carry most of that credit, and the cost is not hypothetical: the union blocks 98.6% [98.0, 99.2] of OR-Bench hard benign prompts. Two things follow cheaply. Calibrating the guard models to a 3% benign false-positive rate on the XSTest-safe pool cuts the union’s benign block rate 3.9x (2.57–7.20x refitted on 4 benign pools) and cuts that same zeroed-mass statistic 4.4x; it also returns 8 harmful prompts to a sole blocker. That is only a start, as 98.4% of the harmful prompts are still blocked by two or more sources, but it shows the saturation is a property of the operating points these guards ship with, not of any-one-blocks composition itself. Nor is the support collapse the 9-way union’s: by coalition size 4, 53.6% of (coalition, member) pairs already read exactly +0.000 on the harmful side (46.7–60.0% with one guard model per vendor). Audit a redundant stack for what it costs: here no single removal changes a harmful decision, so full-stack ablation has no harmful-side support.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI safety, guardrails, over-refusal, Shapley value, ablation, attribution, LLM safety filters, redundancy, cooperative game theory, evaluation methodology

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_5PmuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper explores refusal filter stacks in safety classifiers. The core observation is that redundancy makes individual ablations uninformative about component filter contributions. Several approaches are considered for improved stack selection, including iterative removal using greedy backward elimination.

Critical strengths
  • The paper focuses on a practical problem affecting real deployments.
  • The proposed methods improve performance of the stack over the baseline by a considerable amount.
Critical weaknesses
  • The conceptual novelty appears limited. The central observation that redundant veto components can have zero leave-one-out effect despite contributing in other configurations is a familiar consequence of redundancy, as the authors acknowledge. The empirical characterization is useful, but I am not convinced that the findings on this deliberately untuned stack establish a sufficiently general or consequential contribution for ICLR.
  • The abstract is overloaded with numerical results, notation, and qualifications, which obscure the central question and takeaway. Clarity would improve substantially if it prioritized the problem, the experimental approach, and the main finding, leaving most detailed statistics and secondary results to the main text.
Questions for authors

See weaknesses

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

This paper studies what leave-one-out ablation measures in redundant refusal stacks composed by an any-one-blocks rule. The authors evaluate eight released guard models and the assistant’s self-refusal behavior on 3,937 prompts, deriving all 512 component subsets from recorded decisions. At vendor-default operating points, harmful-prompt blocking is saturated: no harmful prompt has a sole blocker, making every full-stack harmful-side removal marginal zero. Comparisons with exact Shapley values and subset optimization illustrate why these local marginals should not be interpreted as general component contributions. Calibration, coalition-size analyses, alternative rosters, and additional benchmarks characterize the scope of this phenomenon.

Critical strengths
  1. The paper identifies a practical problem in evaluating multi-guard safety systems and analyzes it in depth. When guards have overlapping coverage, leave-one-out ablation can assign zero scores to components that contribute substantially in other configurations. The paper systematically investigates when this occurs and how it affects the interpretation of ablation results, offering useful guidance for auditing redundant refusal stacks.
  2. The paper provides extensive experimental validation. The authors enumerate all 512 component subsets and evaluate multiple benchmarks, complementing the main results with threshold calibration, varying coalition sizes, and alternative guard rosters. These experiments help establish the prevalence of the observed phenomenon and identify the conditions under which it weakens or changes.
Critical weaknesses
  • The incremental contribution beyond established redundancy effects is limited. (1) Redundancy masking leave-one-out effects is already understood, and the sole-blocker condition follows directly from OR composition. (2) The main new contribution is empirical evidence of how prevalent this phenomenon is in the selected guard stacks. (3) The additional practical value of coalition attribution remains unclear, since leave-one-in evaluation resolves much of the ambiguity and iterative backward elimination already finds the exact optimum in the reported experiments.

  • The practical relevance of the strongest findings remains uncertain. (1) The primary stack is deliberately untuned, and matched-FPR calibration substantially reduces its over-refusal. (2) Harmful-side saturation is evaluated on public benchmarks that may overlap with guard training data, leaving its prevalence on unseen or adaptive attacks unclear. These limitations make it difficult to assess how severe the evaluation blind spot remains in calibrated, operationally plausible stacks.

Questions for authors
  1. Beyond illustrating the distinction between removal effects and coalition-averaged contributions, what additional empirical insights does the attribution analysis provide?
  2. What concrete component-selection or maintenance decision is improved by coalition attribution beyond singleton evaluation and iterative backward elimination?
  3. Do the existing artifacts include an evaluation that simultaneously uses matched-FPR calibration, one guard per vendor, a fixed self-refusal component, and a held-out harmful distribution? If so, how much harmful-side removal support remains?
Reviewer_pu6zOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper studies what leave-one-out ablation tables measure in redundant refusal stacks. The authors compose 8 released guard models plus the assistant's own refusals, and enumerate all 512 subsets, comparing each guard's leave-one-out score with its Shapley value. They find that leave-one-out score is a poor predictor of Shapley value, while leave-one-in score reproduces the Shapley ordering up to a few adjacent swaps.

Contributions:

  • The paper shows that with default thresholds, every harmful prompt is blocked multiple times, so no single removal changes any harmful decision and the ablation table only reflects over-refusal.
  • The paper shows that in their setting, the guards with zero ablation score contain most of the Shapley credit, demonstrating that leave-one-out score is a poor predictor of Shapley value.
  • The paper finds that the assembled stack has a huge FPR (57.6–98.6%), and that calibrating guards to a common benign FPR improves the situation.
  • The paper proposes cheaper alternatives to evaluating the contribution of individual guards than full enumeration, like leave-one-in scores, greedy backward elimination, and matched calibration.
Critical strengths
  • The paper thoroughly compares leave-one-out scores to fully computed Shapley scores, and provides cheap, directly usable alternatives (leave-one-in scores, greedy backward elimination) to the failure mode it describes.
Critical weaknesses
  • The central claim, that leave-one-out scores are uninformative when every harmful prompt is blocked by multiple guards, is obvious. The paper’s primary contribution is therefore mainly a demonstration of this in one particular setting.
  • Nobody would realistically deploy the stack used in the paper (8 guards in a union, with an FPR of up to 98.6%), and there's no evidence that anyone actually prunes/evaluates guard stacks by reading full-stack ablation tables. The most useful practical takeaway, recalibrating guards to a fixed FPR, doesn't depend on the paper's main analysis.
  • The result that leave-one-out score is a poor predictor for Shapley value is mostly a problem if Shapley value is what we care about, and the paper doesn't argue for that. Shapley averages a guard's contribution over all 512 subsets of guards, most of which are totally unrealistic. The important question is rather "which subset of the guards should I deploy?", which greedy backward elimination answers already.
Questions for authors
  1. Are there cases where anyone prunes/evaluates guard stacks using full-stack ablation in the way the paper warns against?
  2. How do the results hold up in more difficult settings where harmful refusals aren't as overdetermined?
  3. Why is Shapley value the right measure of a guard's importance?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.