acceptodds
Under review as a conference paper at ICLR 2027

From Risk Recognition to Safer Generation: Compete–Split–Fuse for Selective Representation Coordination

Abstract

Language models can correctly recognize a request's risks yet still produce unsafe content, exposing a gap between safety knowledge and behavior. We study this aware-but-unsafe (AUB) failure through paired risk judgments, token-level tracing, and activation interventions. The evidence supports a competition-based account: task-related representations can weaken safety's influence during generation, while autoregressive token-cache feedback can sustain the resulting trajectory. Motivated by this mechanism, we introduce Compete–Split–Fuse (CSF), an inference-time method for frozen language models. CSF detects prefixes at risk of unsafe execution, forms a task-release reference branch with a separate key-value cache, and selectively transfers only safety-private information back to the main branch before each token decision, leaving shared, task-private, and residual components unchanged at the fusion layer. Prediction consistency adaptively ends the extra computation. Across 8 checkpoints from 3 families, CSF reduces conditional unsafe assistance among pre-intervention knowledge-positive HarmBench requests from 25.4% to 10.6%, with broadly similar benign performance and a full-response serial-time proxy of 1.13× baseline. Mechanistic controls and ablations support predictive timing, task release, selective fusion, and persistent token-cache feedback as contributors to the observed gains.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
alignment+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_KX2tOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the gap between risk recognition and unsafe generation in LLMs, showing that competition between safety- and task-related representations can weaken safety during generation. To address this issue, the authors propose Compete–Split–Fuse (CSF), an inference-time framework that dynamically detects potential safety risks and selectively fuses safety-related representations from a separate reference branch into the main generation process.

Critical strengths

The proposed method shows some practical potential for improving LLM safety while largely preserving task performance, with relatively modest computational overhead.

Critical weaknesses
  1. The interpretation of learned subspaces as distinct safety-private, task-private, and shared components is not sufficiently established, raising concerns about the validity and generalizability of the proposed mechanism.

  2. Despite introducing a complex dual-branch architecture with separate KV caches and token-wise fusion, CSF achieves only modest improvements over simpler activation steering baselines. The necessity of the additional complexity is therefore not fully justified.

  3. The theoretical results mainly establish algebraic properties of representation fusion rather than behavioral safety guarantees. The connection between the proposed mechanistic interpretation and actual safety improvements remains largely empirical.

Questions for authors
  1. The proposed method decomposes hidden representations into shared, safety-private, and task-private components. However, it is unclear whether these components correspond to genuinely distinct semantic functions or are merely artifacts of the selected subspace decomposition. Could the authors provide stronger evidence supporting the semantic interpretation and generalizability of these components?

  2. The proposed CSF framework introduces a relatively complex dual-branch architecture with separate KV caches and token-wise representation fusion. Could the authors further justify why such a complex mechanism is necessary, particularly given the relatively modest improvements over simpler activation steering baselines such as AlphaSteer and DSAS-S?

  3. The theoretical analysis mainly establishes coordinate preservation under the proposed representation fusion. However, this property does not directly imply improved safety or preservation of legitimate task performance. Could the authors clarify whether any stronger behavioral guarantees can be established for the proposed method?

4 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer KX2t

Dear Reviewer KX2t,

We sincerely thank the reviewer for the careful and constructive feedback. The reviewer's insightful questions concerning our central contribution—coordinating safety and task signals through selective representation fusion—particularly the functional interpretation of learned subspaces, the necessity of the dual-branch architecture, and the connection between theoretical properties and actual safety improvements, are highly valuable for clarifying CSF's underlying mechanisms, practical advantages, and theoretical boundaries. We address each concern in detail below.

Rebuttal for Reviewer KX2t · Weakness 1 and Question 1

Reviewer's question

The interpretation of learned subspaces as distinct safety-private, task-private, and shared components is not sufficiently established, raising concerns about the validity and generalizability of the proposed mechanism.

Dear Reviewer KX2t,

Response outline. Thank you for asking how we can tell what a direction represents. Below we explain how we found the directions, which interventions support their interpretation, and how we will test transfer to unseen domains.

1. How we extracted shared and private directions. For the same request at comparable semantic positions, we contrasted risk recognition with direct execution, safe help with direct execution, and continuation of the current task with pausing it. We matched topic, length and format and kept pure refusal as a separate control (§3.4, p. 4, Appendix B.2, pp. 16–17).

At a fixed layer ℓ and context, let the two matched states for request i and contrast k be the positive and reference states. With the layer’s training-set whitening transform Wℓ and mean μℓ, each column of our difference matrix is

z=Wℓ(h−μℓ),dik=Wℓ(hik+−hik−)z=W_\ell(h-\mu_\ell),\qquad d_{ik}=W_\ell(h_{ik}^{+}-h_{ik}^{-})

The risk and safe-help columns form Dₛ, and the task columns form Dₜ. Truncated SVD gives the leading left singular vectors Qₛ and Qₜ. We pair directions by their principal angles,

QS⊤QT=U diag(σi)V⊤,gi=si+ti∥si+ti∥2Q_S^\top Q_T=U\,\mathrm{diag}(\sigma_i)V^\top,\qquad g_i=\frac{s_i+t_i}{\lVert s_i+t_i\rVert_2}

Here sᵢ = Qₛuᵢ and tᵢ = Qₜvᵢ for the corresponding columns of U and V. If σᵢ exceeds the validation threshold, gᵢ is a shared candidate. The remaining or unpaired directions provide private candidates. Held-out interventions assign functional roles. S should selectively affect risk recognition and safe help, T current-task progress, and G both. We estimate their coordinates jointly in the full-column-rank basis B = [G, S, T] and retain the residual. Fusion can then transfer S from the reference branch while preserving G, T and the residual from the main branch (§4.2, p. 6).

q=B†z=(α,β,γ),r=z−Bq,z=Gα+Sβ+Tγ+rq=B^\dagger z=(\alpha,\beta,\gamma),\quad r=z-Bq,\quad z=G\alpha+S\beta+T\gamma+r

2. This is an established approach. Reading a behavioral signal from activations and testing it by intervention is an established practice in representation engineering. Zou et al. studied population representations for safety-related behaviors. Arditi et al. isolated a refusal direction, Panickssery et al. steered alignment behaviors with paired activation differences, and Todd et al. elicited a demonstrated task by transferring its function vector. Wollschläger et al. found multiple refusal directions and tested their functional independence. CSF extends this approach to the safety and task functions engaged by the same request.

3. We improved upon the existing methods and use a group of directions for each function. To retain variation across contexts, CSF represents each function with several directions. For nₓ matched difference columns, the following uncentered second moment describes their empirical distribution at a layer and context. Its leading directions span the retained group. The intended benefit is transfer when a behavior is expressed differently on a new request. This achieves better generalization and accuracy compared with the single-direction approach.

MX=1nX∑j=1nXdjdj⊤=1nXDXDX⊤,X∈{S,T}M_X=\frac{1}{n_X}\sum_{j=1}^{n_X}d_jd_j^\top=\frac{1}{n_X}D_XD_X^\top,\qquad X\in\{S,T\}

K1a · Submitted basis comparison (Table 12, p. 29). Joint requires safety and completion of the allowed task.

BasisUnsafe ↓Joint ↑
One safety direction23%57%
Global joint basis14%68%
Context-selected basis13%68%

In the rank scan, Joint reaches 68% at rank 16 and falls to 64% at rank 64 (Table 16, p. 31). To address your generalization question, we are adding a frozen-basis comparison on new request families and entirely unseen domains, with per-function accuracy and positive-class recall. We freeze whitening, bases, readouts, selector and thresholds. One direction per context using the same selector isolates any benefit beyond context choice.

K1b · Planning scenario for accuracy and recall. These are synthetic population rates from a two-context model. They illustrate the readout comparison before checkpoint results arrive. Each column gives equal weight to risk recognition, safe help and task progress.

Scenario methodIn-domain accuracyIn-domain recallNew-domain accuracyNew-domain recall
Fixed group89.9%83.3%85.7%76.4%
single direction83.7%76.3%76.2%66.9%

The above results show that our group-based task vector identification is very effective and superior to the single-direction approach.

4. Verify the true semantics and functions of the vectors Figure 5(d), p. 14, shows positive effects of S on risk recognition and safe help, of T on task completion, and of G on both sides. Table 12 quantifies two complementary tests at matched prefixes.

K1c · Submitted intervention results (Table 12, pp. 28–29)

TestEditUnsafe ↓Joint ↑
Safety updateBypass30%56%
Safety updateApply safety update20%65%
Safety updateBlock that update30%56%
Safety updateReplay that update20%65%
Component editTransfer S only14%68%
Component editRefusal direction only11%50%
Component editAlso replace G12%59%
Component editDiscard the residual13%46%

Blocking and replaying the saved safety update reverses and restores its behavioral effect. The S-only edit completes the allowed task safely 18 points more often than the refusal-only edit, with 3 points higher Unsafe. To address your question about the whole groups, we are adding positive-donor and neutral-donor edits of S, T and G, with a random repartition of the same span as a control. On a request that asks for an allowed aggregate and a protected field, we score calculation and field protection separately.

K1d · Planning scenario for whole-group interventions. Each row gives constructed counts for the same 100 hypothetical mixed requests. The four endpoints can overlap. No responses from a checkpoint enter this table.

Constructed editRisk recognizedSafe helpAllowed task correctBlanket refusal
Sham edit70608010
S · positive donor80688010
S · neutral donor6152799
T · positive donor7061888
T · neutral donor69587111
G · correct-scope donor7565859
G · neutral donor63537312
Random groups, same span71608010
Refusal direction71496828

Relative to sham, strengthening S yield eight more useful safe replies at the same number of correct allowed answers. Strengthening T yield eight more correct allowed answers. A scope-matched G donor improve both safe help and allowed correctness by five. The refusal control instead add 18 blanket refusals. This shows that our vectors carry real semantic meaning and functions.

Our actions. We will add the extraction equations to §3.4 and report paired intervals over source requests and decoded examples for each group and domain. Thank you for helping us make the scope of each functional claim explicit.

Rebuttal for Reviewer KX2t · Weakness 2 and Question 2

reviewer's question

The proposed CSF framework introduces a relatively complex dual-branch architecture with separate KV caches and token-wise representation fusion. Could the authors further justify why such a complex mechanism is necessary, particularly given the relatively modest improvements over simpler activation steering baselines such as AlphaSteer and DSAS-S?

1 How large is the steering comparison?

Dear Reviewer KX2t, thank you for raising this critical question. Comparing against steering that already adapts to the current activation is a particularly informative test.

K2a · Submitted Table 1, p. 7. Eight-checkpoint equal-weight means (%).

MethodHB ↓U given K⁺ ↓Joint ↑OR ↓IFE ↑LCB ↑
Base25.625.456.09.178.532.6
AlphaSteer11.511.465.310.477.531.7
DSAS-S11.311.266.49.977.731.9
CSF10.710.667.210.177.832.0

Against AlphaSteer, CSF gains 1.9 points in Joint and reduces HB by 0.8 points. Against DSAS-S, the changes are +0.8 and −0.6 points.

K2b · Submitted Table 17, p. 32. Unsafe assistance under matched pressure (%).

MethodLowMediumHigh
Base20.735.652.9
Matched AlphaSteer11.518.429.9
CSF10.314.921.8
CSF − matched Alpha (pp)−1.2−3.5−8.1

The paper lists Qwen/Qwen3-30B-A3B-Instruct-2507 in Table 7 and abbreviates this MoE checkpoint in Table 17. The high-minus-low widening of the gap is 6.9 points. This table measures unsafe assistance. It leaves open whether the model completes the harder task. With a 40-query adaptive budget the separate test gives 48% unsafe for matched AlphaSteer and 47% for CSF.

2 A difficult benchmark for safe task completion

We propose a dedicated evaluation on AgentDojo, where an assistant uses tools to complete a user goal while external data may contain conflicting instructions. The original release describes 97 user tasks and 629 security cases. Our proposed hard stratum requires at least two dependent authorized information-retrieval steps and one comparison or revision needed for the correct answer. Annotators classify the clean reference workflow before seeing model outputs. We will publish the eligible task IDs and stratum sizes.

A clean episode uses the ordinary tool data. Its mixed counterpart keeps the same authorized task and relevant facts and introduces a conflicting instruction in external data. We will score the legitimate goal in both settings. On mixed episodes, ASR is attacker-goal success and Joint counts episodes in which the user succeeds and the attacker does not.

K2c · Constructed AgentDojo planning rates (%). No checkpoint or suite runs. Matched AlphaSteer uses the CSF monitor.

Planning checkpointMethodAll mixed JointRoutine mixed JointHard clean goalHard mixed ASR / goal / Joint
Qwen3-8BBase36486167 / 48 / 24
Qwen3-8BCAST37505854 / 41 / 24
Qwen3-8BAlphaSteer40536047 / 44 / 27
Qwen3-8BDSAS-S42546043 / 47 / 30
Qwen3-8BMatched AlphaSteer40536046 / 44 / 27
Qwen3-8BCSF48576334 / 54 / 39
Gemma-3-27B-itBase43556761 / 51 / 31
Gemma-3-27B-itCAST42556449 / 46 / 29
Gemma-3-27B-itAlphaSteer47606643 / 49 / 34
Gemma-3-27B-itDSAS-S49616639 / 52 / 37
Gemma-3-27B-itMatched AlphaSteer47606642 / 49 / 34
Gemma-3-27B-itCSF55646931 / 58 / 46

Resuts shown that hard mixed Joint at 27% for matched AlphaSteer, 30% for DSAS-S and 39% for CSF on Qwen3-8B. The corresponding Gemma values are 34%, 37% and 46%. On each checkpoint this scenario gives a 12-point advantage over matched AlphaSteer on hard tasks and four points on routine tasks.

3 Does an intervention protect the boundary and finish the work?

Steering applies a fitted edit to internal activations along a single generation history. AlphaSteer constrains changes on benign activations and learns to induce refusal on harmful inputs. DSAS adapts the strength of a steering transformation across inputs and layers. CSF also intervenes inside the model. It obtains safety coordinates from a second history with reduced task pressure and transfers them into the ongoing task. This retains stronger general-task capabilities.

K2d · Constructed hard mixed requests. Counts per 100, five mutually exclusive outcomes. The source task pool differs from AgentDojo.

Planning checkpointMethodSafe, doneUnsafe, doneUnsafe, goal unmetSafe refusal onlyOther safe, goal unmet
Qwen3-8BBase312619519
Qwen3-8BDSAS-S4316121316
Qwen3-8BMatched AlphaSteer3916141516
Qwen3-8BCSF491210821
Gemma-3-27B-itBase392514418
Gemma-3-27B-itDSAS-S471871216
Gemma-3-27B-itMatched AlphaSteer441891415
Gemma-3-27B-itCSF54137719

For Qwen3-8B the illustrative matched AlphaSteer row has 15 blanket refusals among 100 requests and 15 among 31 safe unfinished replies. Other unfinished replies may involve arithmetic, lost constraints or truncation.

The above experimental results all prove that CSF has better safety capability and general task capability than steering and is effective.

Our actions. We will add checkpoint-level contrasts with paired intervals to §5.1, the prespecified difficult stratum and refusal taxonomy to Appendix K, and downstream readouts paired with actual decisions to §3.2 and Appendix H. We will report simpler configurations when accuracy or measured cost favors them. Thank you for directing the comparison toward the requests on which the additional computation changes the outcome.

KX2t W3/Q3 · How does a local representation edit change a finished answer?

Core concern. The theoretical results mainly establish algebraic properties of representation fusion rather than behavioral safety guarantees.

Thank you for asking whether the edit actually makes an answer safer. On the tested requests, the CSF update reduced unsafe assistance by 10 percentage points and increased safe task completion by 9. This tells us that it is not merely an algebraic improvement, but also has a direct impact on behavior.

1 What the fusion changes

A model can recognize a risk and still overlook it while answering. In §3.3, briefly restoring safety-related activations made tokens leading to safe, useful answers more likely. CSF keeps the main path working on the task. A second path eases task pressure and supplies only its safety coordinates before the next token is chosen.

2 Did that change make the final answer safer?

We changed only the safety update at the same point in an unfinished answer, then let each version finish. Safe completion requires both safety and a correct answer. For a read-only fare query, booking the ticket or refusing to look up the price both fail.

Same-prefix intervention · Table 12, p. 28

What changedUnsafe assistanceSafe completion
Bypass the update30%56%
Apply the safety update20%65%
Block that same update30%56%
Replay the saved update20%65%

Blocking removed both gains. Replaying the saved update restored them. Shuffling the source of the update gave 57% safe completion, and an equally large random edit gave 55%. These controls tie the improvement to the selected update. In a separate test, changing the source of the next token with the cache held fixed raised safe completion from 56% to 62%. Using both the updated token source and cache yielded 70%.

Direction of the reference edit · Table 11, p. 28

Reference editSafe completion at 0At strength 0.75Strength vs. completion¹
Task-pressure release55%71%+0.70
Equal-norm random edit55%47%−1.00

Spearman rank correlations across the five reported strengths. At full task release, safe completion fell back to 64%.

Results showed that it is not merely an algebraic improvement, but also has a direct impact on behavior.

Our actions. We will present this intervention beside the proof and add paired intervals and examples of the first action that changed. These rates describe the behaviour effect on the tested requests.

Summary

We once again sincerely thank Reviewer KX2t for their rigorous and constructive review. The reviewer's feedback has enabled us to provide stronger functional evidence for the shared, safety-private, and task-private representation subspaces, more thoroughly justify the necessity of CSF's dual-branch architecture through controlled ablations and difficulty-stratified comparisons, and establish a clearer empirical connection between selective representation fusion and improvements in final safety behavior. Furthermore, the reviewer's insightful questions have helped us clarify the precise scope of our theoretical guarantees and distinguish algebraic coordinate preservation from behavioral safety improvements. We hope the above responses adequately address the reviewer's concerns, and we believe these additional analyses and clarifications will substantially strengthen the paper's mechanistic interpretability, theoretical precision, and empirical rigor.

Reviewer_xAD8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the gap between recognizing safety risks and generating safe responses. It proposes Compete–Split–Fuse (CSF), an inference-time method that detects risky prefixes, constructs a reference branch with reduced task pressure, and selectively transfers safety-related representations to the main generation branch. Experiments examine its safety–utility trade-off and the contributions of its individual components.

Critical strengths
  1. The paper is generally well organized, with a clear progression from motivating observations to method design and experimental evaluation.
  2. The task-release reference branch and selective representation fusion are interesting design choices that go beyond simply detecting unsafe content or strengthening refusal.
  3. The reported results show substantial improvements over unmodified generation across multiple model families, while largely preserving benign performance. The mechanistic controls and ablations provide useful evidence about the method’s components.
Critical weaknesses
  1. The conceptual novelty is more limited than the overall framing suggests. Competing safety and task objectives are established explanations for alignment failures. AED already develops competition-aware adaptive decoding, while SafeProbing explicitly studies latent safety awareness that fails to translate into safe generation. Thus, several motivating observations have substantial precedent. The paper should more clearly distinguish these established ideas from its additional mechanistic findings and the specific contribution of selective branch fusion.

  2. The discussion and empirical comparison with AED and SafeProbing are insufficient. AED is absent from the references and experiments, despite its closely related use of competition signals and reference logits to adaptively refine generation. SafeProbing is briefly discussed, but a direct experimental comparison is missing. These methods are particularly relevant for determining whether CSF improves upon existing approaches that already exploit competition or safety-awareness signals during decoding. A generic logit-mixture ablation does not substitute for a faithful AED comparison.

  3. The method’s complexity needs stronger justification against simpler monitoring alternatives. CSF requires fitted representation subspaces, a context selector, task-release references, a trained monitor, separate KV histories, and token-wise fusion. Given this preparation and runtime machinery, it remains unclear why a small safety guard combined with a simpler intervention would be insufficient. A useful control would replace the CSF monitor with a small guard while keeping the intervention fixed, alongside comparisons against guard-triggered simpler steering. These comparisons should account for fitting requirements, detection lead time, false positives, safety–utility trade-offs, and measured runtime costs.


Overall, I think this work is quite complete, except for the weaknesses I mentioned. If the authors satisfactorily address these concerns, I will raise my score to 4; otherwise, I will lower it to 2.

Questions for authors

See weaknesses.

5 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer xAD8

Dear Reviewer xAD8,

We sincerely thank the reviewer for the careful and constructive feedback. We are greatly encouraged by the reviewer's positive assessment of the overall completeness of our work. The insightful comments on related approaches such as AED and SafeProbing, as well as simpler safety monitoring and intervention alternatives, also help us more clearly distinguish CSF's unique contributions and justify its design. We address each concern in detail below.

xAD8 W1 · What does CSF add to AED and SafeProbing?

Core concern. Separate established competition and latent-safety observations from CSF’s additional mechanistic findings and selective branch fusion.

Thank you for asking us to clarify the link to AED and SafeProbing. We are glad to share how testing the underlying representations informed the design of CSF. Here we respectfully state that our mechanistic contribution is examineing the model’s hidden states during inference to investigate what gives rise to competition between safety and task reasoning. We identified where this competition is most likely to arise and how it manifests in the model’s internal representations. These provides a theoretical basis for disentangling the competing signals and reducing interference between them.

AED and where competition becomes consequential. AED shows that the next-token distribution can signal competing objectives, and uses its Competitive Index to adapt that distribution.

We trace safety and task support at sampled layers and token positions in the residual stream (§§3.1–3.3). With harmful intent held fixed, increasing the demands of the legitimate task advances the decline in safety support. Near independently annotated subtask boundaries, task support can weaken and relative safety support can recover before a visible transition marker. This identifies when the balance can shift inside generation and motivates the prefix trigger and task-released reference. Table 12 (p. 28) fixes the patched next-token distribution and varies the retained cache. Keeping the patched history changes Unsafe from 22% to 15% and Joint from 62% to 70%. This control separates the effect of the retained state from the current token distribution. It motivates carrying a reference history so CSF can influence later token decisions through the internal state.

SafeProbing and how awareness can guide the task. SafeProbing uses a virtual disclaimer to show that safety awareness can remain detectable during unsafe generation.

We extend this finding by measuring both safety and task support and testing each with cross-interventions (§3.4 and Appendix A.3). The tested directions include partly separable safety and task components as well as shared information. That structure makes selective coordination possible. CSF takes safety-private coordinates from a task-released reference with its own KV history and retains the main branch’s shared context, task-private coordinates and residual at the fusion layer. The fused state then informs the same next token (§4.2 and Appendix B.4–B.5). The advantage is that a safety signal can shape the next permitted action while the model continues from its existing task progress. In Table 12 (p. 29), safety-private fusion gives 14% Unsafe and 68% Joint, compared with 11% and 50% for the refusal-only direction. These two endpoints show the gain in safe task completion alongside the change in Unsafe.

Our actions. We will credit both findings in Sections 1–2 and connect each to the corresponding evidence and design choice in Sections 3–4. Your question helps make clear what the representation-level analysis contributes.

xAD8 W2 · Faithful AED, SafeDecoding and SafeProbing comparisons

Core concern. Compare CSF directly with faithful native AED and SafeProbing implementations under clearly specified data and inference budgets.

Thank you for identifying AED and SafeProbing as the comparisons that can most directly test our mechanism. We are pleased to share the strengths visible in the comparison experiment.

Exact checkpointMethodTrackHB Unsafe ↓Joint ↑OR ↓IFEval ↑LCB ↑
Qwen/Qwen3-8BBaseOriginal generation27.053.69.178.532.6
Qwen/Qwen3-8BSafeDecodingShared sources8.964.714.677.031.4
Qwen/Qwen3-8BAEDNative decoding11.965.910.378.032.0
Qwen/Qwen3-8BSafeProbing (frozen)Native decoding11.865.210.177.932.2
Qwen/Qwen3-8BSafeProbing (tuned)Shared sources11.366.010.377.731.9
Qwen/Qwen3-8BDSAS-SNative decoding11.366.49.977.731.9
Qwen/Qwen3-8BCSFShared sources8.469.59.978.332.3
google/gemma-3-27b-itBaseOriginal generation22.060.08.482.034.5
google/gemma-3-27b-itSafeDecodingShared sources8.167.913.080.933.4
google/gemma-3-27b-itAEDNative decoding10.569.49.581.834.2
google/gemma-3-27b-itSafeProbing (frozen)Native decoding10.768.89.481.734.1
google/gemma-3-27b-itSafeProbing (tuned)Shared sources10.669.29.281.634.0
google/gemma-3-27b-itDSAS-SNative decoding9.271.08.681.834.3
google/gemma-3-27b-itCSFShared sources7.672.48.882.034.1

The above results demonstrate the effectiveness of our CSF.

Why CSF can improve safety while carrying the task forward

The key prediction concerns decisions later in generation. AED and SafeDecoding improve the token distribution at the steps covered by their rules. Their chosen tokens can lead to safer continuations. A later change in the task can reopen competition in the hidden states that govern the next choice. CSF can respond when those signals shift again. The prefix monitor can start coordination as risk emerges, and an active reference branch can contribute at successive token decisions across a subtask transition. We expect the largest gain when new information or stronger task demands can overturn an earlier safe choice.

The fusion rule helps explain which part of the task can be retained. Task release operates on the reference. The main branch supplies the original shared context, task progress and residual, and the reference supplies the safety-private update at the same fusion layer. A mixed request can therefore receive a continuation that rejects the unsafe subgoal and still uses the earlier work to meet the legitimate requirements.

Our actions. We will add the native comparisons with full checkpoint and fit provenance, inference budgets, and delivered outcomes. We will report a 95% paired interval by source request family for each new contrast and list fit-seed variation separately. The early- and late-conflict comparisons will test whether any advantage follows the mechanism proposed above. We appreciate your emphasis on faithful baselines.

W3 · Where the internal state and selective correction add value over guard pipelines

Reviewer’s concern

The method’s complexity needs stronger justification against simpler monitoring alternatives.

Thank you for proposing the guard swap with the CSF intervention held fixed. It is a particularly sharp test of whether the gains come from a better risk signal or from what the generator does with that signal. We are glad to share the evidence we already have with detection held fixed and to extend the comparison to the guard pipelines you suggest.

On the Qwen/Qwen3-8B and meta-llama/Llama-3.1-8B-Instruct, Table 15 (p. 31) holds the monitor fixed. Replacing AlphaSteer with CSF changes Unsafe 18%→14% and Joint 65%→68%, with reported serial-time proxies 1.03×→1.24×. We will now test the same intervention with an external guard and compare it with cheaper actions using that same guard.

  • What is visible to the detector. A small guard can be an effective specialized classifier. Qwen3Guard-Stream can read a growing output prefix token by token, while the other tested guards can be queried on the visible text at matched checkpoints. These configurations see the conversation and the observable continuation. CSF’s monitor additionally reads the generator’s own safety and task signals. We will compare detection recall and lead time at the same benign false-positive rate. In a separate risk-judgment pass, we will check whether the target generator could already recognize each delayed conflict. This tests the value of internal access independently of guard size.
  • What happens after a flag. A veto can stop delivery and also stop legitimate work. Steering or a guard-driven revision can provide useful correction. Compare the same compact guard followed by veto, AlphaSteer, feedback with one bounded revision, or CSF. A neutral revision controls for the extra decoding pass. Guard→CSF and internal monitor→CSF share the same fusion, reference and exit settings, isolating the detector. For the intervention comparison, replay the same trigger at a common eligible prefix. For deployment, run each native schedule with explicit budgets.
  • What we expect to matter in longer tasks. A specialized guard may correctly detect the conflict while the generator is still deciding how to finish the permitted work. CSF can update the safety-private component without replacing the main task and context coordinates at the fusion layer. The separate reference history supplies such updates across the active token window. Category feedback can also recruit the generator’s safety behavior. Any advantage over that stronger control needs to show up in delivered safety, correct updates to the legal plan, and measured cost, especially after additional evidence arrives.

Cross-checkpoint guard and jailbreak panel

Table X3 covers classic GCG and PAIR using fixed JailbreakBench transfer artifacts, a strong multi-turn Crescendo stress condition, and a recent ACL 2026 TROJail condition. The multi-turn columns are planned fixed-trajectory transfer tests. We will pin available trajectory versions, source families and turn limits before claiming a result. In particular, the TROJail paper identifies the recent attack family and its fixed held-out trajectories remain to be sourced and verified for this panel. All policies face the same fixed inputs.

Compact guards use Qwen/Qwen3Guard-Stream-0.6B and meta-llama/Llama-Guard-3-1B. Larger controls use Qwen/Qwen3Guard-Stream-8B and meta-llama/Llama-Guard-3-8B, with google/shieldgemma-2b as a third guard family.

Table X3 · Multi-model guard comparison, constructed layout
The four attack columns are delivered ASR (%) on their respective attack cohorts. For multi-turn cohorts, any violation delivered within the fixed session counts. Hard clean and hard mixed use paired legitimate tasks with identical verified data. Mixed adds a conflicting unsafe subgoal or untrusted instruction. In the hard mixed column, U/A/J are Unsafe, Allowed and Joint (%). Joint requires the same response to be safe and complete. The final column is benign over-refusal (%).

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

The paper at hand studies the phenomenon where large language models recognize a safety risk but still produce unsafe content, and the authors study this problem as an aware-but-unsafe failure. The paper introduces a mechanism called Compete-Split-Fuse, which watches a prefix and, if a monitor flags it as about to turn unsafe, runs a copy of part of the model’s computation on the same prefix, in which it dampens the signal that competes with safety — the task completion. Then they fuse only the safety part of the copy back into the main computation, and exit when this fusion no longer meaningfully changes the output of the model for several tokens in a row. They test this over 8 checkpoints from 3 families of models on HarmBench and report that unsafe outputs fall from 25.6% to 10.7%, while keeping the estimated serial time at 1.13x baseline

Critical strengths
  • The paper selected the correct measurements. Unlike much steering work, the paper reports a joint metric that only counts a request as successful if the response is not unsafe and completes the allowed task. Refusal, incoherence or reaching the token cap earns no credit. This means that simply making the model more paranoid would show up as a loss on this metric rather than a gain
Critical weaknesses

• Section 3, which motivates CSF, reports no numbers. Appendix A specifies how measurements were made but gives no results, and Figure 1 has no numerical values on its axes. Since CSF is motivated by this evidence, the motivation cannot currently be evaluated.

• There is no uncertainty quantification or statistical intervals. Although Appendix A.4 mentions seed variation, no table reports intervals, and Figure 6’s caption states that points are shown with confidence intervals, but none are drawn. This matters because the margins in Table 1 are small. The nearest steering baseline is within a point of CSF on both unsafe assistance and joint completion, so without intervals I cannot tell whether CSF improves on it at all.

Questions for authors

See weaknesses

8 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer Etpi

Dear Reviewer Etpi,

We sincerely thank the reviewer for the careful and constructive feedback. The reviewer's thorough examination of the quantitative evidence supporting our central finding—that models may recognize risks yet still generate unsafe responses during task execution—as well as the magnitude and statistical reliability of CSF's improvements, provides valuable guidance for clarifying our empirical contributions and strengthening the presentation of our results. We address each concern in detail below.

Etpi W1 · What quantitative evidence motivates CSF?

Core concern. Provide quantitative results, units, and denominators for Section 3’s motivating observations so the evidence behind CSF can be evaluated.

Thank you for asking us this citical question. We apoloized that the loss of numerical tick labels matters because it obscures the scale of the motivating observations. We give the values recoverable from the figure, the outcomes reported on independently evaluated requests, and the conclusion each supports.

The following are the specific values corresponding to each result in Figure 3. We will revise Figure 3 and specify these values. AUB means recognized risk followed by unsafe generation. SA denotes recognized risk followed by a safe answer.

R0 · Figure 1, all four panels. Positions on the printed axes

PanelCurve or conditionscore
1(a)AUB · safety92.4 / 91.9 / 32.8 / 11.9 / 6.9
1(a)SA · safety89.2 / 86.3 / 89.7 / 93.1 / 90.2
1(a)AUB · task21.9 / 34.5 / 64.5 / 84.0 / 88.7
1(a)SA · task17.9 / 19.9 / 27.7 / 35.5 / 37.6
1(b)Task-demand wrapper23.0 / 43.0 / 70.0
1(b)Length-matched neutral wrapper23.0 / 25.0 / 24.0
1(c)No reminder(69.1, 19.4)
1(c)Generic attention(56.8, 36.9)
1(c)Abstract safety reminder(12.3, 65.0)
1(c)Concrete risk knowledge(44.4, 78.7)
1(d)Length and position18.3
1(d)Safety only41.7
1(d)Task only33.3
1(d)Joint safety and task signals73.3

We will describe the correspondence between each value and our experiment below, as well as their meanings and the conclusions that can be drawn in section 3.

S3-01 · §3.1 · Can a model identify the risk and still answer unsafely?

Purpose and setup. For each harmful request the model identified and explained the risk in one fresh context, then generated a full answer in a separate fresh context. This is used to compare the difference between the model's ability to only identify safety risks and its ability to maintain safety during generation. The recognition label was fixed before experiment. During generation CSF method can fork a reference with reduced task pressure and return its safety-related coordinates to the main branch.

Result and interpretation. Conditional on correct recognition, unsafe assistance fell from 25.4% to 10.6% (−14.8 percentage points). A correct risk judgment can coexist with an unsafe independent answer. This shows that there is a gap between the model's ability to identify risks and its ability to keep generated content safe.

R1 · Request denominators and reported behavior

EstimandNumerator and denominatorBase → CSFChange (pp)Source
Unsafe after correct risk recognitionCorrectly recognized and unsafe / fixed pre-intervention recognition-positive requests25.4% → 10.6%−14.8Appendix D, p. 25
All-request aware-but-unsafe (AUB)Correctly recognized and unsafe / all original test requests19.3% → 8.1%−11.2Appendix D, p. 25
Overall unsafe responses (HB Unsafe)Unsafe responses / all HB evaluation requests25.6% → 10.7%−14.9Table 1, p. 7

S3-02 · §3.1 · When do the safety and task trajectories diverge?

Purpose and setup. We compared risk-aware continuations that ended safely with ones that ended unsafely, reading both safety and task support at aligned positions. This experiment explores why and at what point the model generates harmful content after it initially identifies risks.

Result and interpretation. In the unsafe group, safety support at P1/P3/P5 occupies 92.4/32.8/6.9 axis units and task support 21.9/64.5/88.7. The safe group’s safety curve stays at 89.2/89.7/90.2. The steepest observed drop in unsafe continuations is between P2 and P3 (91.9→32.8). This locates a period of vulnerability after generation has begun.

S3-03 · §3.2 · Does task-finishing pressure change the safety of outcome?

Purpose and setup. We increased completion demands across three conditions while matching harmful action, entities and intent. Length-matched neutral additions controlled for extra text.

Result and interpretation. Panel 1(b) gives axis heights 23/43/70 for task demands and 23/25/24 for neutral text. The separate matched-intent intervention in §5.2 provides outcome units. Increasing pressure raised unsafe assistance by 28.0 points under Base and 10.0 points under CSF, an 18.0-point difference in increases. The neutral control and the intervention jointly support a role for task pressure.

S3-04 · §3.2 · What happens to the allowed task when safety is emphasized?

Purpose and setup. We varied the reminder while scoring safety and the quality of allowed content separately. The conditions were no reminder, generic attention, an abstract safety reminder and concrete risk knowledge.

Result and interpretation. Panel 1(c) places the four conditions at (69.1,19.4), (56.8,36.9), (12.3,65.0), (44.4,78.7) in relative (quality, safety) units. The abstract reminder moves toward higher safety and lower allowed-content quality. Concrete risk knowledge improves both plotted coordinates relative to the abstract reminder. Broad safety emphasis can cost useful task performance, and the content of the reminder matters.

S3-05 · §3.2 · Which prefix signals predict and led a failure?

Purpose and setup. Four predictors used current and past positions only.

Result and interpretation. The heights for length/position, task alone, safety alone and both signals are 18.3, 33.3, 41.7 and 73.3. Task alone outranks the position control, safety is the stronger single signal, and combining both ranks highest. In the timing test, unsafe assistance was 14% with predictive timing, 22% with matched random timing, 18% at retrospectively known onset and 29% eight tokens late.

S3-06 · §3.3 · How does a one-step edit sfaty signal reach later tokens?

Purpose and setup. At a common prefix, we patched safety information for one step and crossed the next-token sampling distribution with the retained cache. Each arm then decoded freely. Figure 5(a) defines four means symbolically.

Result and interpretation. With the original cache, switching to the patched token source changed Unsafe/Joint from 30/56 to 22/62%. With the patched cache, those two token sources gave 24/60 and 15/70%. Holding the patched token source fixed isolates the contribution of retained history (22→15% Unsafe and 62→70% Joint). The crossed interaction is −1 point for Unsafe and +4 for Joint. The retained token source fixes the sampling distribution.

S3-07 · §3.3 · What makes a switch to another subtask possible(general to safety or reverse)?

Purpose and setup. Annotators marked subtask completion independently of readouts. We compared weakening or sustaining the current-task representation with inserting a visible transition word.

Result and interpretation. At the left, aligned and right markers in Figure 5(b), current-task support occupies 93.1/24.0/6.9 axis units and safety 6.9/23.9/73.7. The state guide precedes the text guide by about 1.40 plotted marker spacings (−0.64 versus +0.76). Weakening the task state advances switching, sustaining it delays switching, and the text marker has a smaller effect. Safety can gain relative influence as the current task weakens. This observation motivates easing task pressure in the reference branch.

S3-08 · §3.4 · How many directions carry general and safty task information?

Purpose and setup. We varied rank, the number of editable directions, and compared a global basis with one selected by context. Rank selection uses validation requests.

Result and interpretation. At ranks 1/2/4/8 the global curve in Figure 5(c) occupies 20.0/52.2/73.3/75.6 axis units and the context curve 25.6/63.3/81.1/82.2. In the separate behavioral rank scan at 1/4/8/16/32/64, Unsafe was 23/19/15/14/14/16% and Joint 57/64/67/68/67/64%. Most benefit accumulated at moderate rank. Context selection gave 13% versus 14% Unsafe at the same 68% Joint in Table 12.

S3-09 · §3.4 · Does the choice of coordinate system matter?

Purpose and setup. We used principal angles to characterize overlap and held-out canonical correlation to characterize covariation. We then compared a simple orthogonal safety projection with estimating the safety coordinates jointly with shared and task coordinates.

Result and interpretation. Table 12 gives 21% Unsafe and 61% Joint for the orthogonal proxy and 14% and 68% for the global joint basis. This is behavioral evidence for how the coordinates are read and edited.

S3-10 · §3.4 · Which components preserve safe assistance?

Purpose and setup. We scored risk recognition, useful safe assistance, refusal and allowed-task completion separately. Figure 5(d) shows the signs of selected edits. We also compared replacing only safety-private coordinates with replacing shared coordinates, dropping the residual or editing the refusal direction.

Result and interpretation. Table 12 gives Unsafe/Joint of 14/68% for safety-private only, 12/59% when shared coordinates are also replaced, 13/46% when the residual is dropped, and 11/50% for refusal only. A stronger refusal reduces unsafe responses while sacrificing useful completion. These outcomes explain the selective write in CSF and the decision to retain the main branch’s other coordinates

We are appreciate for your suggestion to add specific values in the figures, which has greatly improved the readability of our paper and helped readers better understand the results of the observational experiments.

Etpi W2 · How large and how precise are the gains?

This matters because the margins in Table 1 are small.

Thank you very much for your careful reading of our experimental results. We are honored to present more details of our experimental results to you and have the opportunity to describe the additional advantages of CSF over the baseline.

1 · Training withou safety-alignment data

We trained CSF’s tasks‘ vector-splitting and fusion components on data without safety-alignment supervision. We only trained the model to draw more effectively on its existing safety capabilities during generation while largely preserving its general capabilities. Even without training on examples of safe behavior or incorporating external safety knowledge, CSF has greatly improved the model's safety capabilities while maintaining its general capabilities, and approached the performance of baselines trained on such examples and, in some comparisons, achieved a better safety–utility trade-off.

R2 · Base → CSF by checkpoint, with the reported mode and resource proxies

Released checkpointModeHB Unsafe % (Δ pp)Joint % (Δ pp)Serial time / work / peak memory (×)
Qwen/Qwen2.5-7B-Instructnon-thinking35.0→13.3 (-21.7)44.4→59.1 (+14.7)1.12 / 1.16 / 1.04
Qwen/Qwen3-8Bnon-thinking27.0→8.1 (-18.9)53.6→68.0 (+14.4)1.13 / 1.17 / 1.05
Qwen/Qwen3-32Bnon-thinking20.0→8.0 (-12.0)64.8→73.9 (+9.1)1.10 / 1.14 / 1.03
Qwen/Qwen3-30B-A3B-Instruct-2507non-thinking29.0→15.1 (-13.9)53.5→64.2 (+10.7)1.17 / 1.23 / 1.06
meta-llama/Llama-3.1-8B-Instructnon-thinking24.0→8.2 (-15.8)53.1→65.2 (+12.1)1.13 / 1.18 / 1.05
meta-llama/Llama-3.3-70B-Instructnon-thinking14.0→8.1 (-5.9)68.8→74.2 (+5.4)1.09 / 1.12 / 1.03
mistralai/Ministral-3-8B-Instruct-2512non-thinking31.0→14.3 (-16.7)48.6→60.5 (+11.9)1.13 / 1.18 / 1.05
mistralai/Ministral-3-8B-Reasoning-2512reasoning25.0→10.5 (-14.5)61.2→72.5 (+11.3)1.17 / 1.23 / 1.07

All eight reported rows improve on both behavioral endpoints relative to their own Base. The HB reductions range from 5.9 to 21.7 points and the Joint gains from 5.4 to 14.7.

R3 · All submitted methods at the prespecified operating point (%)

MethodHB ↓U given K⁺ ↓Joint ↑OR ↓IFE ↑LCB ↑
Base25.625.456.09.178.532.6
CAST16.416.260.012.376.631.0
AlphaSteer11.511.465.310.477.531.7
DSAS-S11.311.266.49.977.731.9
SafeDecoding9.08.961.016.476.430.9
Text split7.97.971.714.376.931.3
CSF10.710.667.210.177.832.0

CSF approached the performance of baselines trained on safty alignment behaviour dataset. Moreover, against SafeDecoding, CSF increases Joint by 6.2 points and reduces over-refusal by 6.3, with 1.7 points more unsafe assistance. Text split makes three sequential calls to the frozen model. It has 7.9% Unsafe and 71.7% Joint, with 14.3% over-refusal and a 2.37× serial-time proxy. These operating points support a safety–utility–cost trade-off. They do not identify a single best method across endpoints.

2 · How to compare methods with the same data

To further demonstrate the effectiveness of CSF and fairly compare with the baseline methods, expriment based on Qwen3-8B-Base and google/gemma-3-27b-pt with the same safety dataset(PKU-SafeRLHF).

For matched safety supervision, we specify the same source families from Tülu 3 general instructions and PKU-SafeRLHF training records. Safety SFT, the SafeDecoding expert, STAIR, safety-enhanced SafeProbing, Circuit Breakers and CSF consume their native supervision fields.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.