acceptodds
Withdrawn from ICLR 2027

FCRS: Learning Fine-grained Safety Signals with Cross-layer Consistency for Refusal Steering

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Steering models to refuse malicious prompts is vital for a variety of LLM safety applications. However, achieving robust refusal activation steering remains challenging because operating on the full activation vector limits the capture of fine-grained safety signals and treating each layer hinders cross-layer consistency. To address these challenges, a novel refusal activation steering method is proposed by integrating fine-grained activation unit selection with cross-layer consistency modeling, including a fine-grained activation unit selector learning (FAUSL) module and a cross-layer consistent safety-relevant signal detector learning (CCSDL) module. The proposed method enjoys several merits. First, the proposed FAUSL module can select the most safety-relevant AUs effectively via an AU importance scoring network. Second, the CCSDL module can model long-range dependencies across layers to extract consistent safety-relevant signals. Extensive experimental results on eleven challenging benchmarks show that our proposed method significantly outperforms state-of-the-art refusal activation steering methods.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Fine-Grained, Cross-layer Consistency

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_MB16OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

FCRS proposes a detector-gated refusal-steering framework for frozen instruction-tuned language models. Its fine-grained activation unit selector learning (FAUSL) module encodes the magnitude and direction of malicious–benign activation differences, learns AU relevance scores, and fixes the top 128 AUs per selected layer. Its cross-layer consistent safety-relevant signal detector learning (CCSDL) module aggregates selected AU activations through layer-specific prototype attention, applies cross-layer self-attention, and predicts a score that gates refusal steering. The modules are trained sequentially, using a heuristic-derived AU selection target followed by safety-label regression with a descriptor-diversity regularizer. The evaluation covers three language models, seven jailbreak benchmarks, and four benign-utility benchmarks. The paper reports improved defense performance while matching the corresponding unmodified models’ utility scores at the displayed precision. A component ablation on Qwen2.5-7B-Instruct examines the selector and detector designs.

Critical strengths
  1. An application-relevant design that explicitly separates detection from intervention. The frozen-backbone, selectively gated design targets the safety–over-refusal trade-off rather than maximizing refusal indiscriminately. Offline AU selection and online hierarchical attention provide a coherent way to combine selected activation features across layers without modifying the base model’s parameters (Sections 3–4.1; Figure 2, page 4). This is consequential because a useful safety intervention must preserve normal task performance as well as block malicious requests. I view the integration of feature selection with conditional, multi-layer detection as a potentially useful contribution.
  2. Promising empirical results across multiple backbones and attack families, with explicit utility evaluation. The evaluation is not restricted to one model or one attack style. For example, Table 1 reports an increase from 63% to 86% DSR on Llama-3.1-8B-Instruct under Cipher, and from 88% to 97% on Qwen2.5-7B-Instruct under PAIR, relative to AlphaSteer. Table 2 reports FCRS matching the corresponding base model on all four displayed utility metrics for each of the three backbones. Including XSTest alongside instruction-following and mathematical-reasoning evaluations makes the empirical scope more informative than attack-only evaluation. These results provide a promising signal that merits further consideration, subject to the validation concerns below.
Critical weaknesses
  1. The end-to-end defense and evaluation protocol are insufficiently specified, and key reported aggregates require reconciliation. Section 3.3 ends with a safety score that triggers refusal steering, but the manuscript does not provide a reproducible specification of the intervention itself: how the refusal direction is constructed, where it is injected, its strength, and which prompt or generation positions it affects. Sections 3–4.1 also leave the actual selected layer indices and the training/validation data sources, sizes, and malicious–benign pairing procedure unclear. Section 4.2 repeatedly states that evaluation follows AlphaSteer, but does not explicitly identify the inherited configuration or establish separation between detector training, threshold calibration, and benchmark testing. It is also unclear whether attacks were optimized against FCRS or reused from attacks on an undefended model. These details materially affect both reproducibility and the scope of the robustness claim; the detector alone does not determine the end-to-end defense success rate. The reported aggregates also need an audit. In Table 1 (page 8), the seven displayed Llama/AlphaSteer entries have an unweighted mean of approximately 92.71%, rather than the reported 91.93%. The Gemma/Jailbreak Antidote entries similarly average approximately 29.86%, rather than 43.94%. A different aggregation rule could explain some discrepancies, but the weights and denominators are not specified. This matters because the aggregate results are used to quantify improvements over baselines. These are questions about the support for the central empirical claim, not merely presentation details.
  2. The experiments do not adequately isolate the value of learned AU selection or establish the claimed cross-layer consistency mechanism. Equation (8) trains FAUSL against a Top-K target constructed from activation magnitude and directional consistency. The exact rule producing that target is not specified, and there is no comparison against directly using the target-generating heuristic. Consequently, the advantage of learning the selector, rather than applying its supervision rule, remains unclear. With CCSDL held fixed, Table 3 shows 98.14% DSR for random AU selection and 99.00% for FAUSL, a 0.86-percentage-point difference. A gain near the performance ceiling may be meaningful, but its reliability and attribution cannot be assessed without uncertainty estimates or repeated selections/training runs. For CCSDL, Equation (9) is regression of the final safety score to the sample label, rather than an explicit cross-layer agreement objective; Equation (10) encourages descriptor diversity. This does not make the architecture invalid, and diversity need not conflict with consistent predictions. However, attention-based fusion alone does not establish the paper’s claimed consistency mechanism. The comparison between variants [C] and [D] replaces a layer-independent MLP detector with the full attention-based architecture, so it supports the usefulness of that replacement without isolating consistency from model capacity and generic multi-layer fusion. A capacity-matched fusion control and an operational definition or measurement of consistency are needed to substantiate the proposed explanation for the gains. These gaps limit the scientific conclusions that can be drawn from the reported performance.
Questions for authors
  1. Clarification — complete algorithm. What is the complete inference procedure, including the selected layer indices, refusal-direction construction, intervention strength, affected token positions, and whether detection requires an additional prompt pass before intervention? Please also specify the exact rule used to construct the FAUSL target YlY^l in Equation (8). A compact algorithm with the relevant model-specific settings would resolve substantial ambiguity about what was evaluated.
  2. Clarification — data separation and threat model. What datasets and sample counts are used for the two training stages and for ordinary-benign/hard-benign threshold calibration, and how are malicious–benign pairs formed? How are these sets separated from the reported test benchmarks? For the jailbreak evaluations, please specify sample counts, the success-judging procedure, and whether attacks were generated against the complete FCRS defense or transferred from an undefended model. Which baseline results were rerun, and which were inherited from prior work?
  3. Targeted validation of existing claims — selection and cross-layer modeling. With the data, steering intervention, and calibration protocol held fixed, how does FAUSL compare with directly selecting the AUs used to construct its supervision target? How does CCSDL compare with a capacity-matched detector that combines the same layer descriptors through concatenation or pooling rather than cross-layer attention? Please provide repeated-run variability for at least the random-selection versus FAUSL comparison, and clarify what measurable property distinguishes “cross-layer consistency” from ordinary multi-layer feature fusion.
  4. Clarification — metric aggregation and utility preservation. What is the exact aggregation rule for Avg DSR, and can the displayed per-attack results and aggregate values be reconciled using the underlying counts? Table 2 matches the unmodified-model metrics in all 12 model–task entries at the reported precision. What are the actual benign gate-activation rates and held-out detector true-positive/false-positive rates at the deployed threshold? Please clarify whether the utility results reflect identical generations or merely identical aggregate scores, and confirm that they were measured with the full gated system enabled.
Reviewer_vzVrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FCRS, a refusal activation steering framework that combines fine-grained activation unit (AU) selection with cross-layer safety detection. It contains two modules: (1) FAUSL, which learns to select safety-relevant AUs using a MLP, and (2) CCSDL, which uses attention mechanisms to capture cross-layer dependencies and output a safety score for conditional refusal steering. The paper evaluates FCRS on three LLMs across seven jailbreak benchmarks and four utility benchmarks. The results show improvements in defense success rate over existing SOTA methods, while maintaining comparable utility. Ablation studies are conducted to examine the effectiveness of the two modules.

Critical strengths
  1. The paper introduces cross-layer attention to capture dependencies among safety-relevant representations, rather than processing layer-wise activations independently as in AlphaSteer. This is a reasonable approach to improving the consistency of safety detection across layers.

  2. The experimental results are promising. FCRS achieves higher average defense success rates than the evaluated baselines across three LLMs while maintaining comparable utility. The improvements on certain jailbreak benchmarks suggest potential benefits for robustness against diverse attacks.

Critical weaknesses
  1. Training data are insufficiently specified. The paper does not identify the training datasets, explain malicious–benign pair construction, or clarify dataset overlap with evaluation benchmarks, limiting reproducibility and assessment of generalization.

  2. The steering intervention is unclear. Although threshold calibration is described, the construction of the refusal direction, intervention locations, and steering strength remain unspecified.

  3. FAUSL provides limited incremental gains. In Table 3, FAUSL improves average DSR by only 0.86% over random AU selection with CCSDL (99.00% vs. 98.14%). The ablation is also limited to Qwen2.5-7B-Instruct, without reporting variability across runs. The paper also lacks direct comparisons with the target-generating selection rule or the selection method in Fine-Grained Activation Steering [1].

  4. FAUSL's supervision targets are unclear. Section 3.4 does not specify the exact rule used to combine activation magnitude and directional consistency to construct the supervision targets YlY^l.

  5. Generalization across model sizes is unclear. Although the evaluated models follow prior work, FCRS uses a two-stage training procedure to learn AU selection and cross-layer safety detection. These learned components may be sensitive to differences in model size and activation patterns. Evaluation beyond the tested 7B–9B range would help establish whether the training procedure generalizes across scales.

[1] Feng, Zijian, Tianjiao Li, Zixiao Zhu, Hanzhang Zhou, Junlang Qian, Li Zhang, Chua Deryl, Lee Mak, Gee Ng, and Kezhi Mao. "Fine-grained activation steering: Steering less, achieving more." In International Conference on Learning Representations, vol. 2026, pp. 39421-39443. 2026.

Questions for authors
  1. What datasets and splits are used to train and validate FAUSL and CCSDL, and how are malicious–benign pairs constructed? Do any examples or attack types overlap with the evaluation benchmarks?

  2. Once the safety score triggers steering, how is the refusal direction constructed, which layers and token positions are modified, and how is the intervention strength determined?

  3. Is FAUSL's performance gain over random AU selection consistent across random selections and training seeds? How does FAUSL compare with direct selection using its target-generating rule or the selection method in Fine-Grained Activation Steering [1]?

Reviewer_tnf5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes FCRS, a detector-gated refusal steering method: a frozen LLM’s activations are scored by a learned detector, and a refusal vector is added only when the score exceeds a validation-calibrated threshold. A FAUSL module first selects K=128 dimensions per layer: a two-MLP scorer, applied to direction- and magnitude-encoded malicious–benign activation differences, is regressed onto a heuristic top-K target, and its top-K indices are then frozen. A CCSDL detector is then trained with binary malicious/benign labels on these selected activations. In CCSDL, per-layer learnable prototypes cross-attend to embedded activation values, cross-layer self-attention mixes the layer summaries, and an MLP head outputs a scalar safety score. On Llama-3.1-8B, Qwen2.5-7B and Gemma-2-9B, the method reports higher average defense success rates than AlphaSteer across seven jailbreak attacks. Utility scores are identical to those of the unsteered models.

Critical strengths

[S1] The design is simple and practical. A small attention-based probe over a frozen LLM, combined with a binary gate, is cheap and modular. It leaves benign behavior untouched whenever the detector does not fire, which avoids the utility losses.

[S2] The results suggest that detection quality, rather than the sophistication of the steering operator, is the main bottleneck. A plain refusal vector behind a good detector outperforms AlphaSteer’s null-space-constrained steering in DSR.

[S3] The random-selection ablation is informative, though underemphasized: 128 random dimensions per layer with CCSDL reach 98.14 DSR, versus 99.00 for the full method. This indicates that safety-relevant information in the residual stream is highly redundant, a useful observation for future probe and steering design.

[S4] The evaluation covers three model families and seven attack types under the established AlphaSteer protocol, which makes the comparison straightforward.

Critical weaknesses

[W1] The central “fine-grained selection” claim is not supported.

The near-parity of random selection (98.14) and FAUSL (99.00) undercuts the paper’s main motivation. FAUSL itself is trained to regress onto a hand-crafted top-K mask (magnitude plus directional consistency, formula unspecified), computed from the same data it is applied to. A well-trained scorer therefore simply just reproduces the heuristic, which contradicts the “data-driven vs. human priors” framing. Missing baselines include the heuristic used directly, Feng et al.’s direction-only criterion, and a sweep over K.

[W2] The cross-layer consistency claim is weakly evidenced.

Cross-layer inconsistency in AlphaSteer is shown only by a schematic (Fig. 1c), with no quantitative measure for either method. The per-layer ablation (53.57) is implausibly far below AlphaSteer’s per-layer linear detectors (97.29), possibly suggesting an under-tuned baseline. How its per-layer outputs are combined is not described. Simple multi-layer baselines are absent, e.g., linear or MLP probes on concatenated full or randomly subsampled activations.

[W3] Generalization, the key question for a supervised gate, is not addressed.

The training data for both stages is not described, so it is unclear whether jailbreak-style prompts resembling the test attacks were seen in training. There is no held-out-attack or adaptive-attack evaluation, and no per-benchmark detector firing rates.

[W4] Important method details are missing or ambiguous.

  • Which layers are used.
  • How the refusal vector is constructed, where it is applied, and its strength.
  • How the 4×128 descriptor matrix C is reduced to a scalar.
  • Whether the embedded activation values carry dimension identity. Without it, attention over the selected values is permutation-invariant.

[W5] Limited novelty and informative results

In general, this paper does not propose a very novel method, and does not provide enough informative results and analysis through it's empirical experiments, considering the existing literature and established methods in llm safety alignment.

Questions for authors

No questions at this time. Concerns are stated in the weaknesses section.

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

Related papers

Discussion (0)

Sign in to comment.