acceptodds
Withdrawn from ICLR 2027

RAS: Measuring LLM Safety Through Refusal Alignment

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Repeated safety evaluation of open-weight large language models requires generating and judging responses for many related checkpoints. We introduce (Refusal Alignment Score), a reference-relative measure of harmful-request refusal computed by the five-stage \SafeVec procedure. \SafeVec extracts layer-wise refusal directions from an aligned reference model, measures their alignment in related targets, and calibrates a 0–100 score using behavioral attack success rates (ASR) from calibration models. Subsequent target scoring uses input forward passes alone. Across 18 held-out variants from the Llama, Gemma, and Qwen families, correlates with (Pearson , Spearman ) and achieves 83.0% pairwise ordering agreement. Across five Llama prompt configurations, reference directions have mean cosine similarity . With calibration ASR already available, setup-inclusive evaluation of five targets is – faster than the specified behavioral protocol. These results support refusal alignment as an efficient signal for screening related checkpoints and prioritizing behavioral audits.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Safety Evaluation, LLM Safety

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_7GBwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces RAS (Refusal Alignment Score), a lightweight metric for screening safety in related open-weight LLM checkpoints. The core idea is that within a model family, degradation in harmful-request refusal correlates with reduced alignment between the target model's activations and refusal directions extracted from an aligned reference. The five-stage SafeVec procedure extracts layer-wise refusal directions from the reference, selects a scoring window, computes alignment on harmful and jailbreak prompts, and calibrates a 0-100 score using behavioral ASR from calibration models.

Critical strengths

S1. The paper addresses a relatively novel topic. Given a reference model (e.g., Qwen) and a large set of its variant checkpoints, the question of how to perform rapid safety audits across these checkpoints is practically motivated and underexplored.

S2. The proposed Refusal Alignment Score enables low-cost safety auditing. By relying on input forward passes rather than response generation and judging, RAS offers a lightweight screening signal that can amortize evaluation cost across many checkpoints.

Critical weaknesses

My main concern with this paper is its practical utility.

W1. The setting is quite restrictive. RAS simultaneously requires (a) white-box access to residual-stream activations, (b) a reference model in the same family, and (c) multiple related checkpoints. Many realistic evaluation scenarios do not satisfy all three conditions at once, which limits the applicability of the method.

W2. The baseline cost being amortized is already modest. Behavioral evaluation on a benchmark like HEx-PHI with a single judge is not expensive in the first place.

W3. Missing comparisons to simple and intuitive baselines. The paper demonstrates correlation between RAS and behavioral safety, but does not compare against cheaper ways of approximating behavioral evaluation itself, such as:

(a) Subsampling the evaluation set: running the full behavioral protocol on only 1%-10% of the safety evaluation prompts;

(b) Capping generation length: performing behavioral evaluation with a very small output token budget (e.g., 10 tokens), since safety-relevant behavior (refusal or compliance) typically manifests at the start of the response.

Both strategies can substantially reduce evaluation cost, may correlate well with final safety scores, and are arguably simpler to deploy than RAS. I encourage the authors to include these baselines so that the contribution of RAS over naive cost-reduction strategies is clearer.

Questions for authors

see weakness

Reviewer_gX7SOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes RAS, a 0–100 score for screening refusal degradation in open-weight derivatives of an aligned reference model. A target is scored by the cosine alignment of its last-token activations with the reference’s mean-difference refusal directions on 11 harmful and 11 jailbreak prompts. The drop from the reference is then mapped to 0–100 using the ASR of 5–7 calibration models per family (Section 3, pp.2–4). On 18 held-out Llama, Gemma and Qwen variants, RAS correlates with 100(1−ASR) (pooled Pearson 0.934, Spearman 0.851; Table 2, p.7), and the paper reports 3.2–80.5× speedups when calibration ASR is already available (Table 3, p.8).

Critical strengths
  1. The procedure is fully specified and cheap per target. Table 14 (p.24) can be reproduced exactly from Tables 4–6, and scoring a target needs only forward passes on 22 prompts (Table 3, p.8). This makes the claims checkable, and the approach would be practical if its validity were established.
Critical weaknesses
  1. The validation does not support the screening claim. Six of the 18 targets are abliterated or heretic variants by name (Table 7, p.17), built by removing the very direction RAS measures. The high Pearson values therefore largely reflect separation between intact and ablated clusters (Gemma r = 0.994 but ρ = 0.771, CI [0.00, 1.00]; Table 2, p.7). For variants not named as ablated, which is the motivating fine-tuning case, RAS can miss the failure: Aion-RP keeps a much higher RAS than B (Fig. 4, p.7), and Lexi-Uncensored-V2 keeps unsafe alignment of 0.064 at ASR 0.939 (Table 4, p.14). Coverage is also narrow: the 33 probe prompts are hand-written with no taxonomy or per-category analysis (Section 4.1, p.5), the behavioral comparator contains no attacks (App. A, p.14), references with ASR up to 0.236 are all anchored at 97.70 (Prop. 1, p.4), and only 4–8B dense models are tested. Fixing this requires a new evaluation study, not a revision.
  2. RAS shows no value beyond the ASR it approximates. It is validated only against one protocol’s ASR, and its calibration inherits that protocol’s judge (Section 5, p.9), so its only demonstrated benefit is cost. Yet no comparable-cost baseline is tested, such as first-token refusal log-odds (Arditi et al., 2024) or short generations on the same prompts. Once calibration ASR is counted, a five-target audit is slower than behavioral evaluation in every family (Table 3 and Eq. 13, p.8). Uncited work already relates fine-tuning safety drift to geometry relative to a reference direction (Du et al., 2025; Asif and Amiri, 2026; Poppi and Lukas, 2026). That leaves the calibration as the main novelty, and its benefit over raw scores is not shown (App. I.2, p.24; Table 8, p.18).
Questions for authors
  1. [Clarification] How was each target produced? Please report the Table 2 statistics separately for ablation-based and fine-tuning-based targets, give per-target RAS and ASR, and confirm that QT6 and QC3 are distinct checkpoints.
  2. [Clarification] Were τ, η, λ, c and β fixed before any held-out target was scored, and was the judge validated on outputs of uncensored models?
  3. [Requested experiment] How do comparable-cost baselines (first-token refusal log-odds, short generations) correlate with B on the same targets?
  4. [Requested experiment] Does RAS track ASR on controlled fine-tuning checkpoints and under automated attacks such as PAIR or GCG?
Reviewer_GzHQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes the Refusal Alignment Score (RAS), a representation-based metric to efficiently evaluate harmful-request refusal across related open-weight LLM checkpoints. Computed via the five-stage SafeVec procedure, RAS extracts layer-wise refusal directions from a reference model and applies family-specific calibration to yield a 0–100 score, eliminating the need for text generation or behavioral judging. Evaluated on 18 checkpoints across Llama, Gemma, and Qwen, RAS achieves a 0.934 pooled Pearson correlation and 83.0% pairwise ordering agreement with traditional behavioral judges. The metric also delivers 3.2–80.5× computational speedups when evaluating multiple target models.

Critical strengths

From my observation below are the critical strengths:

  1. Robust Methodology & Reusable Protocol: The paper presents a systematic SafeVec procedure consisting of five stages that separates reference construction and calibration from target scoring. By fixing these parameters before evaluation and using 18 held out target models, the authors prevent calibration label leakage. Furthermore, evaluating new checkpoints using only input forward passes, without generating responses or performing behavioral judging, substantially reduces computational costs and enables more efficient safety audits.

  2. Comprehensive Empirical Validation: Experiments demonstrate a strong association between representation scores and behavioral refusal across three model families, with a Pearson correlation of 0.934 and 83.0% pairwise ordering agreement. Additional robustness analyses, including prompt transfer experiments and progressive refusal ablation, provide supporting evidence for the metric's stability and sensitivity, suggesting its potential utility in practical safety evaluations.

Critical weaknesses
  1. Limited Evidence for Detecting Subtle Safety Degradation The experimental evaluation raises several concerns regarding the generalizability of RAS:

a. Limited diversity of safety degradation: The evaluation primarily includes aligned and heavily uncensored models, so the strong correlations may reflect broad group separation rather than sensitivity to subtle safety changes. Several checkpoints were also explicitly modified to remove refusal representations, making their reduced alignment scores expected.

b. Limited checkpoint coverage: The evaluation includes only 18 held out checkpoints. The authors should investigate models with moderate safety degradation introduced through standard fine tuning and report their correlations and pairwise ranking agreement. This is particularly important given Gemma's statistically insignificant Spearman correlation ((p = 0.0724)).

c. Limited behavioral benchmark diversity: The authors should validate RAS against additional safety benchmarks, such as HarmBench, JailbreakBench, or StrongREJECT, to assess whether the observed correlations generalize across different harmful prompt distributions.

  1. Lack of Competitive Baselines and Unclear Incremental Contribution The proposed method builds on existing refusal representation techniques, but its incremental advantage remains insufficiently established:

a. Missing representation baselines: The authors should compare SafeVec against raw refusal direction cosine similarity, direction projection, and linear activation probes, reporting Pearson correlation, Spearman correlation, and pairwise ordering agreement on the same 18 checkpoints.

b. Unclear calibration contribution: Since the calibration mapping preserves within family rankings, the authors should clarify its additional value for absolute score interpretation and comparisons across model families.

c. Missing efficiency baselines: The authors should compare SafeVec against behavioral evaluations using subsets of 10, 20, 50, and 100 harmful prompts, reporting ranking accuracy and computational costs to establish whether the proposed method offers a meaningful efficiency advantage.

Questions for authors
  1. Sensitivity to Subtle Safety Degradation: Can the authors report correlations and pairwise ranking agreement for checkpoints with moderate safety degradation to demonstrate that RAS detects subtle changes rather than broad differences between model groups?

  2. Generalization Across Safety Benchmarks: Does RAS maintain consistent correlations and checkpoint rankings on other safety benchmarks, such as HarmBench, JailbreakBench, or StrongREJECT, without recalibration?

  3. Comparison with Representation Baselines: How does SafeVec compare against simpler methods, such as raw refusal direction cosine similarity, direction projection, and linear activation probes, on the same 18 checkpoints?

  4. Contribution of Behavioral Calibration: Since calibration preserves within family rankings, what additional benefit does it provide for absolute score interpretation and cross family comparability?

  5. Computational Efficiency Comparison: How does SafeVec compare with behavioral evaluations using 10, 20, 50, and 100 harmful prompts in terms of ranking accuracy and computational cost?

  6. Code Availability and Reproducibility: Could the authors provide a publicly accessible code repository containing the SafeVec implementation, evaluation scripts, prompts, and calibration configurations to facilitate independent reproduction of the reported results?

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

This paper introduces the Refusal Alignment Score (RAS), computed through the SafeVec procedure, for efficiently screening related open-weight model checkpoints. The method extracts layer-wise harmful-versus-benign activation directions from an aligned reference model, measures target alignment with these directions on harmful and jailbreak-framed prompts, and applies a family-specific calibration informed by behavioral attack success rates (ASR). Once this configuration is established, target scoring requires input forward passes without response generation or judging. Experiments on 18 held-out checkpoints across three separately calibrated model families show substantial association with behavioral refusal measurements and computational savings under the evaluated protocol.

Critical strengths
  • Repeated behavioral evaluation of related checkpoints can be expensive. An inexpensive signal that helps prioritize more detailed audits would be useful.

  • The reported pooled Pearson correlation of 0.934 and Spearman correlation of 0.851 suggest that reference-relative refusal alignment captures meaningful behavioral differences within the evaluated model population.

  • The paper acknowledges that RAS is reference-relative, requires compatible representations, and does not constitute a safety probability. It also reports model-specific discrepancies, uncertainty estimates, and calibration sensitivity.

Critical weaknesses
  • The presentation obscures the intuition, and the calibration complexity is insufficiently justified. The underlying hypothesis is relatively simple: stronger alignment with an aligned reference’s refusal-related direction should predict stronger refusal. However, this intuition is obscured by numerous intermediate quantities, including anchor scales, severity ratios, family multipliers, and sigmoid parameters. The paper should explain what each component contributes to the measurement objective and compare the complete procedure against simpler alternatives. Since Proposition 1 establishes that calibration preserves all within-family rankings, the additional machinery needs stronger justification through improved cross-family comparison or meaningful score interpretation.

  • The method depends on representation compatibility, and representation changes can be mistaken for safety degradation. Section 3.1 assumes compatible hidden-state coordinates and corresponding decoder layers across reference, calibration, and target models. This substantially restricts applicability to related checkpoints; matching hidden dimensions alone does not establish semantic comparability. Even within this setting, adaptation may change how refusal is represented while preserving or improving behavioral safety. Table 5 already provides a relevant example: the Gemma reasoning-distilled variant has lower reference-direction alignment but lower ASR than the reference, 0.076 versus 0.230. A monotonic calibration cannot correct this reversed ordering. The paper does not sufficiently establish when reference misalignment reflects safety degradation and when it reflects a different representation of safe behavior.

  • The evaluation does not establish the trade-off between computational savings and behavioral estimation accuracy. Runtime and behavioral agreement are reported separately at the selected operating configuration. The paper does not show how prediction or ranking quality changes as the number of probe prompts, evaluated layers, or computational budget varies. Section 4.4 examines variability in activation scores, which does not directly establish accuracy relative to behavioral ASR. A particularly important baseline is behavioral evaluation on a smaller, carefully selected prompt subset under the same computational budget. Without this comparison, it remains unclear whether RAS offers a better accuracy–cost trade-off than inexpensive behavioral sampling. Calibration-label acquisition and manual layer selection also need to be considered when assessing practical savings.

  • The model population provides limited evidence of generalization across safety failure mechanisms. The evaluation includes numerous abliterated and uncensored variants. These are useful test cases, but refusal-direction removal directly alters the type of representation measured by RAS. The controlled reference-derived experiment similarly weakens refusal-related directions and then evaluates a score based on such directions. These results establish sensitivity to relevant interventions, but broader validation is needed for safety changes arising from ordinary fine-tuning and other adaptation mechanisms. Particularly informative cases would include substantial representation changes with preserved safety, and degraded safety despite retained reference-direction alignment.

  • The validated scope does not cover conversation-level harmfulness. RAS measures prompt-conditioned activations before response generation. The paper provides no defined aggregation procedure or validation for harmful assistance accumulated across turns, refusal that deteriorates during follow-up interactions, or harmful outcomes assembled from individually benign-looking exchanges. Supplying conversation histories as inputs would still require validation against conversation-level outcomes. The authors acknowledge multi-turn interaction as a limitation, but this remains an important constraint on the metric’s practical coverage. Benign helpfulness and over-refusal also require separate evaluation.

Questions for authors
  1. What improvements does the full calibration procedure provide over raw cosine alignment, reference-relative alignment drops, and simpler calibration methods, particularly for cross-family comparisons?

  2. How robust is RAS to adaptation that changes refusal representations while preserving safety? Can the authors characterize or predict failures such as the Gemma reasoning-distilled example?

  3. Can the authors provide behavioral agreement versus computation curves and compare RAS with smaller behavioral evaluations under matched time budgets? If ASR prediction is an intended use, can a mapping fitted exclusively on calibration models be evaluated using held-out prediction error?

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

Related papers

Discussion (0)

Sign in to comment.