Skip to content

REINS: Refusal-Enhanced Inhibitory Steering with Sparse Autoencoder Features

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/Geralt1020/REINS
Area: LLM Reasoning
Keywords: Sparse Autoencoders, Inference-time Steering, Refusal Enhancement, Inhibitory Steering, Contextual Disguise Safety

TL;DR

To overcome the failure of single-direction sparse autoencoder (SAE) steering on harmful prompts embedded in complex disguises, this paper establishes the hierarchical GUISE benchmark and develops REINS—a coordinated steering method that suppresses dynamic harmful continuation features while amplifying calibrated refusal features, alongside a prompt-side REINS-Gate that safeguards general capabilities.

Background & Motivation

Large language models deployed across open domains must dynamically adapt to evolving content safety policies and risk boundaries. Full parameter fine-tuning is computationally prohibitive, while permanent weight editing techniques frequently incur severe side effects and catastrophic forgetting of general tasks. Inference-time activation steering has thus emerged as a compelling, lightweight paradigm for behavioral adaptation without modifying base weights; in contrast to dense hidden-state directions that entangle semantic intent, stylistic markers, and task states, Sparse Autoencoders (SAEs) project residual activations into sparse, interpretable feature dictionaries, offering a modular, auditable interface for alignment control.

However, existing SAE-based safety steering approaches typically rely on isolated, single-direction interventions—either uniformly amplifying refusal features or enforcing unidirectional shifts within learned subspaces. When malicious requests are wrapped in complex contextual disguises, such as fictional roleplay, professional cyber red-teaming pretexts, academic hypotheticals, or intricate stylistic constraints, these single-direction interventions crumble. The wrapper framing masks surface-level toxicity keywords, leaving the internal latent circuits supporting harmful operational continuations fully active; injecting a refusal signal into such active circuits causes severe pathway competition, either failing to steer the generation away from the harmful task or driving the model into degraded, collapsed outputs (e.g., gibberish or empty strings).

The fundamental tension stems from an operational asymmetry: eliciting a reliable refusal requires two distinct, complementary operations within representation space—depriving the active harmful continuation trajectory of its mechanistic feature support, while concurrently providing a well-calibrated, coherent refusal trajectory. Core idea: coordinate dynamic Harm-Inhibit suppression with calibrated Refusal-Enhance injection in a shared SAE feature space, triggered on demand via a prompt-side REINS-Gate to achieve robust safety under contextual disguise without degrading general capabilities.

Method

Overall Architecture

REINS operates across the layerwise residual stream Sparse Autoencoders (BatchTopK SAEs) of a frozen language model to steer output generation safely. The end-to-end inference flow comprises a prompt-side gate and a dual-controller feature modification engine: first, the input prompt is evaluated by REINS-Gate using prompt-token SAE activations to determine whether intervention is necessary; if deemed high-risk, the unsteered model generates a brief early prefix to identify the active harmful route via gradient-based attribution for prompt-specific suppression (Harm-Inhibit), while an offline-calibrated set of sparse refusal features is added over early generation steps (Refusal-Enhance). A strict disjoint update rule prioritizes suppression whenever feature coordinates overlap, and the steered sparse features are decoded back to residual streams to produce coherent, safe refusals.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
    A["Input Prompt x"] --> B["Adaptive Prompt-Side Gate (REINS-Gate)<br/>Cosine similarity sgate(x) vs threshold τ"]
    B -->|sgate < τ Benign| C["Unsteered Generation Path<br/>Standard residual stream execution"]
    B -->|sgate ≥ τ High Risk| D["Dynamic Harmful Feature Suppression (Harm-Inhibit)<br/>Early prefix attribution Attrpx identifies & zeros active features"]
    B -->|sgate ≥ τ High Risk| E["Offline Calibrated Refusal Enhancement (Refusal-Enhance)<br/>Dual-negative purified features boosted for first MR steps"]
    D --> F["Disjoint Feature Update Scheduling<br/>Enforce suppression priority: ΩH and ΩR\\ΩH"]
    E --> F
    F --> G["SAE Reconstruction & Residual Stream Writeback<br/>Modified zl,t,j' mapped via SAE decoder into Transformer blocks"]
    G --> H["Safe Coherent Refusal Output"]

Key Designs

1. Dynamic Harmful Feature Suppression (Harm-Inhibit): Dismantling latent operational support masked by contextual wrappers

Because complex wrappers embed harmful objectives within legitimate professional or fictional task formulations, selecting features directly from prompt tokens tends to select benign framing tokens rather than the actual malicious intent. Harm-Inhibit addresses this by unmasking the model's emergent operational trajectory: it runs the frozen model forward for a maximum of 32 tokens, stopping at the earliest natural sentence or line boundary to yield an unsteered prefix span \(\tilde{p}_x\) and a refined attribution target \(p_x\). It then computes the summed gradient-times-activation attribution score for each layerwise SAE coordinate \((l, j) \in \mathcal{F}\):

\[A_{l,j}^H(x) = \text{Attr}_{p_x}(l, j) = \sum_{r=1}^{|p_x|} \left( z_{l,t_r,j} \cdot \frac{\partial \log p_\theta(p_{x,r} \mid x, p_{x,<r})}{\partial z_{l,t_r,j}} \right)\]

To prevent disrupting general linguistic scaffolding or foundational formatting mechanics, the final ranking score incorporates a normalized depth bias \(\rho_l\) and semantic context penalties:

\[q_{l,j}^H(x) = [A_{l,j}^H(x)]_+ + (0.25 + 1.10\rho_l) \min\{\omega_{l,j}^H, \nu_{l,j}^H\}\]

The layer depth term biases selection toward intermediate and upper layers (\(l \ge 14\)) where behavioral commitment occurs, while \(\omega_{l,j}^H\) protects refusal/safety contexts (weight 0.15) and penalizes formatting contexts (0.35). The top \(K_H\) features \(\mathcal{S}_H(x)\) are set to zero across all continuation positions \(T_H\), neutralizing the active harmful continuation pathway.

2. Offline Calibrated Refusal Enhancement (Refusal-Enhance): Injecting a purified, reusable refusal attractor

Simply zeroing harmful features eliminates unsafe details but frequently leaves the model stranded in high perplexity, resulting in evasive drift or degenerate collapse. Refusal-Enhance constructs a stable, prompt-independent refusal attractor by analyzing SAE activations during legitimate refusal generations. To eliminate template bias, activations are averaged across 16 instruction-answer pairs formed from four refusal instructions and four fixed refusal statements, yielding \(\mu_{l,j}^{\text{ref}}\). Crucially, to isolate pure refusal intent from generic assistant boilerplate or tokens shared with harmful task continuations, a dual-negative calibration margin is computed:

\[\Delta_{l,j}^R = [\mu_{l,j}^{\text{ref}} - \max\{\mu_{l,j}^{\text{neu}}, \mu_{l,j}^{\text{orig}}\}]_+\]

where \(\mu_{l,j}^{\text{neu}}\) denotes activations on neutral instruction-answer pairs and \(\mu_{l,j}^{\text{orig}}\) represents unsteered continuations on harmful calibration prompts. Augmented by depth and semantic context weighting, the ranking identifies a minimal budget of \(K_R\) coordinates \(\mathcal{S}_R\) (e.g., exactly 2 features for Qwen3.5-4B). During inference, scaled additions \(b_{l,j}^R = \alpha_R [\Delta_{l,j}^R]_+\) are applied strictly over the initial \(M_R = 8\) continuation tokens. In cases where coordinates collide, Harm-Inhibit takes strict precedence: updates are applied over the disjoint partition \(\Omega_R(x) = (\mathcal{S}_R \setminus \mathcal{S}_H(x)) \times T_R\).

3. Adaptive Prompt-Side Gate (REINS-Gate): Eliminating collateral disruption on benign multi-task inputs

Applying global steering across every user query inevitably degrades multi-task reasoning and standard instruction following. REINS-Gate introduces an input-side decision mechanism operating solely on prompt-token SAE representations before decoding commences. During calibration, the mean prompt-token feature vectors \(\mathbf{v}(x)\) of harmful prompts (positives) and benign/general prompts (negatives) are contrasted to derive a sparse discriminative projection vector \(\mathbf{w}_{\text{gate}}\). For any runtime query \(x\), its alignment score is computed as:

\[s_{\text{gate}}(x) = \cos(\mathbf{v}(x), \mathbf{w}_{\text{gate}})\]

Under a fixed false-positive budget on negative calibration prompts (\(\beta = 0.10\)), a threshold \(\tau\) is established. Queries satisfying \(s_{\text{gate}}(x) \ge \tau\) activate REINS; all other queries bypass steering entirely, eliminating collateral distortion on harmless interactions.

Benchmark & Evaluation Protocol

To rigorously diagnose safety mechanisms under contextual disguise, the authors introduce GUISE (Generalized Undercover Instruction Safety Evaluation). Built upon a three-level hierarchical taxonomy (5 categories: Hate, Crime, Violence, Pornography, Self-Harm; 15 subcategories; 45 fine-grained scenarios), GUISE generates 900 balanced prompt pairs via an iterative multi-agent pipeline using Gemini-3.1-Pro. Each sample contains a wrapped harmful prompt disguised in legitimate roles (creative writing, red-team auditing, scientific fieldwork) paired with a matched safe prompt sharing identical framing without the harmful objective. Outcomes are evaluated via automated LLM judges with secondary judge cross-verification and human adjudication across four safety partitions: Harmful Response Rate (HRR \(\downarrow\)), Safe Refusal Rate (SRR \(\uparrow\)), Other Safe Rate (OSR \(\uparrow\)), and Collapse Rate (CR \(\downarrow\)), alongside utility benchmarks (MMLU-Pro, GPQA) and Collateral Effect (CE \(\downarrow\)).

Experimental Results

Main Benchmark Comparisons

The primary evaluation on Qwen3.5-4B-Base across GUISE safety outcomes and multi-task capability benchmarks is presented below (all values in percentages %):

Method HRR ↓ SRR ↑ OSR ↑ CR ↓ MMLU-Pro ↑ GPQA ↑ CE ↓
Original (Unsteered) 90.3 3.7 5.7 0.3 50.0 40.0 0.0
Random-SAE (Negative Control) 90.4 3.3 5.3 1.0 50.0 40.0 1.0
Refusal-SAE (O'Brien et al., 2024) 65.0 2.3 13.7 19.0 48.0 40.0 6.2
SAE-SSV* (He et al., 2025) 78.4 5.0 14.3 2.3 48.0 38.7 5.6
CorrSteer-A (Cho et al., 2025) 43.7 23.7 12.3 20.3 46.3 42.0 16.6
REINS (Full Intervention) 26.3 63.7 7.3 2.7 45.4 35.1 26.4
REINS-Gate (Adaptive Gated) 26.7 62.7 8.3 2.3 49.9 40.0 0.1

Ablation Study

Component ablations on Qwen3.5-4B-Base on GUISE reveal the distinct operational roles of Harm-Inhibit and Refusal-Enhance:

Configuration HRR ↓ SRR ↑ OSR ↑ CR ↓ Findings & Description
Full REINS 26.3 63.7 7.3 2.7 Dual coordination yields lowest harmfulness and highest coherent refusal without collapse
w/o Refusal-Enhance (Harm-Inhibit only) 44.7 8.0 44.0 3.3 Suppresses harmful flow but lacks refusal direction; safe outputs drift into ambiguous OSR
w/o Harm-Inhibit (Refusal-Enhance only) 48.0 44.3 6.3 1.3 Boosts refusal signal but fails to override active disguised harmful continuation circuits
REINS-Gate 26.7 62.7 8.3 2.3 Preserves full defense while dropping collateral effect (CE) from 26.4% down to 0.1%

In external transfer evaluations across standard safety benchmarks, REINS demonstrates remarkable generalization: on AdvBench, HRR drops to 2.9% (while baseline CorrSteer-A exhibits 100% collapse); on JailbreakBench, HRR drops from 53.3% to 3.3% with 93.3% SRR; on HarmBench, HRR falls from 57.0% to 6.5% with 72.0% SRR.

Key Findings

  • Failure of Unilateral Intervention Under Disguise: Relying solely on harmful suppression shifts generations into evasive, non-actionable text (OSR rises to 44.0%), whereas solely enhancing refusal cannot overpower the entrenched task context of wrapped prompts (HRR remains at 48.0%). Combining both achieves a 26.3% HRR and 63.7% explicit refusal rate.
  • Genuine Refusal vs. Degenerate Collapse: Baseline methods like CorrSteer-A achieve apparent safety through severe output degradation (20.3% collapse on GUISE and 100% on AdvBench). In contrast, REINS maintains a minimal collapse rate of 2.7%, proving that behavioral gains stem from true refusal mode shifts rather than representation breakdown.
  • Locality Restoration via Gating: REINS-Gate triggers on 99.3% of harmful prompts while activating on only 3.5% of negative queries, completely restoring MMLU-Pro accuracy (49.9% vs. 50.0% unsteered) and slashing collateral distortion from 26.4% to 0.1%.

Highlights & Insights

  • Coordinated Dual-Action Steering: Demonstrates that safety steering in sparse latent spaces requires orthogonal cooperative forces: actively pruning the forward support of the unwanted behavior while seating the generation trajectory into a calibrated target attractor.
  • Early-Prefix Dynamic Attribution: Effectively circumvents semantic cloaking in disguised prompts by attributing gradients directly to early generated tokens, pinpointing the latent operational features of the true attack vector.
  • High Operational Efficiency: Requires calibrating only 2 to 8 sparse features offline without retraining. REINS requires only 7.79 seconds per prompt, reducing core runtime by 52.0% compared to CorrSteer-A and 87.3% compared to SAE-SSV*.

Limitations & Future Work

  • Model Scale Validation: Current evaluations are primarily situated on Qwen3.5-2B/4B and Gemma-3-1B backbones due to the substantial compute required for full-layer SAE pretraining; validation on 70B+ frontier models remains an important open step.
  • SAE Architecture and Hook Points: Experiments strictly utilize residual-stream BatchTopK SAEs at a single expansion factor, leaving the effects of attention-head hooks, MLP interior activations, and varying sparsity budgets unexplored.
  • Absence of Circuit-Level Causal Mapping: Feature selection is presently driven by activation statistics and gradient attribution rather than comprehensive mechanistic circuit tracing, suggesting future work in mapping causal circuit interactions during steering.
  • vs. Refusal-SAE (O'Brien et al., 2024): Refusal-SAE applies contrastive feature clamping between simple harmful and benign prompts, which lacks the intervention strength needed for complex disguises and induces up to 19.0% collapse; REINS achieves over 27x higher explicit refusal rates via dual-action coordination.
  • vs. CorrSteer-A (Cho et al., 2025): CorrSteer-A depends critically on spontaneous refusal occurrences within calibration data, causing total failure and 100% collapse on challenging benchmarks like AdvBench; REINS utilizes a reusable, independently calibrated refusal bank.
  • vs. Dense Activation Steering (Rimsky et al., 2024; Sheng et al., 2026): Dense hidden-state vectors inevitably modify entangled non-safety representations, causing substantial multi-task degradation; REINS operates on isolated sparse latent coordinates, enabling transparent auditing and negligible collateral impact via REINS-Gate.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ Establishes the dual-action paradigm for SAE safety steering, uncovering why single-direction interventions fail against wrapped adversarial prompts.
  • Empirical Rigor: ⭐⭐⭐⭐⭐ Comprehensive coverage across the new GUISE benchmark, multi-judge validations, three external transfer datasets, detailed ablations, and runtime profiling.
  • Writing Quality: ⭐⭐⭐⭐⭐ Exceptionally clear mathematical formulation, well-structured exposition, and thorough analytical depth in failure mode categorization.
  • Impact & Practicality: ⭐⭐⭐⭐⭐ Provides an inference-time, retraining-free safety solution with low latency overhead and minimal capability degradation, ideal for deployment safeguards.