- 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
-
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.
-
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.
-
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
-
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?
-
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?
-
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
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 method | In-domain accuracy | In-domain recall | New-domain accuracy | New-domain recall |
|---|---|---|---|---|
| Fixed group | 89.9% | 83.3% | 85.7% | 76.4% |
| single direction | 83.7% | 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)
| Test | Edit | Unsafe ↓ | Joint ↑ |
|---|---|---|---|
| Safety update | Bypass | 30% | 56% |
| Safety update | Apply safety update | 20% | 65% |
| Safety update | Block that update | 30% | 56% |
| Safety update | Replay that update | 20% | 65% |
| Component edit | Transfer S only | 14% | 68% |
| Component edit | Refusal direction only | 11% | 50% |
| Component edit | Also replace G | 12% | 59% |
| Component edit | Discard the residual | 13% | 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 edit | Risk recognized | Safe help | Allowed task correct | Blanket refusal |
|---|---|---|---|---|
| Sham edit | 70 | 60 | 80 | 10 |
| S · positive donor | 80 | 68 | 80 | 10 |
| S · neutral donor | 61 | 52 | 79 | 9 |
| T · positive donor | 70 | 61 | 88 | 8 |
| T · neutral donor | 69 | 58 | 71 | 11 |
| G · correct-scope donor | 75 | 65 | 85 | 9 |
| G · neutral donor | 63 | 53 | 73 | 12 |
| Random groups, same span | 71 | 60 | 80 | 10 |
| Refusal direction | 71 | 49 | 68 | 28 |
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 (%).
| Method | HB ↓ | U given K⁺ ↓ | Joint ↑ | OR ↓ | IFE ↑ | LCB ↑ |
|---|---|---|---|---|---|---|
| Base | 25.6 | 25.4 | 56.0 | 9.1 | 78.5 | 32.6 |
| AlphaSteer | 11.5 | 11.4 | 65.3 | 10.4 | 77.5 | 31.7 |
| DSAS-S | 11.3 | 11.2 | 66.4 | 9.9 | 77.7 | 31.9 |
| CSF | 10.7 | 10.6 | 67.2 | 10.1 | 77.8 | 32.0 |
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 (%).
| Method | Low | Medium | High |
|---|---|---|---|
| Base | 20.7 | 35.6 | 52.9 |
| Matched AlphaSteer | 11.5 | 18.4 | 29.9 |
| CSF | 10.3 | 14.9 | 21.8 |
| 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 checkpoint | Method | All mixed Joint | Routine mixed Joint | Hard clean goal | Hard mixed ASR / goal / Joint |
|---|---|---|---|---|---|
| Qwen3-8B | Base | 36 | 48 | 61 | 67 / 48 / 24 |
| Qwen3-8B | CAST | 37 | 50 | 58 | 54 / 41 / 24 |
| Qwen3-8B | AlphaSteer | 40 | 53 | 60 | 47 / 44 / 27 |
| Qwen3-8B | DSAS-S | 42 | 54 | 60 | 43 / 47 / 30 |
| Qwen3-8B | Matched AlphaSteer | 40 | 53 | 60 | 46 / 44 / 27 |
| Qwen3-8B | CSF | 48 | 57 | 63 | 34 / 54 / 39 |
| Gemma-3-27B-it | Base | 43 | 55 | 67 | 61 / 51 / 31 |
| Gemma-3-27B-it | CAST | 42 | 55 | 64 | 49 / 46 / 29 |
| Gemma-3-27B-it | AlphaSteer | 47 | 60 | 66 | 43 / 49 / 34 |
| Gemma-3-27B-it | DSAS-S | 49 | 61 | 66 | 39 / 52 / 37 |
| Gemma-3-27B-it | Matched AlphaSteer | 47 | 60 | 66 | 42 / 49 / 34 |
| Gemma-3-27B-it | CSF | 55 | 64 | 69 | 31 / 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 checkpoint | Method | Safe, done | Unsafe, done | Unsafe, goal unmet | Safe refusal only | Other safe, goal unmet |
|---|---|---|---|---|---|---|
| Qwen3-8B | Base | 31 | 26 | 19 | 5 | 19 |
| Qwen3-8B | DSAS-S | 43 | 16 | 12 | 13 | 16 |
| Qwen3-8B | Matched AlphaSteer | 39 | 16 | 14 | 15 | 16 |
| Qwen3-8B | CSF | 49 | 12 | 10 | 8 | 21 |
| Gemma-3-27B-it | Base | 39 | 25 | 14 | 4 | 18 |
| Gemma-3-27B-it | DSAS-S | 47 | 18 | 7 | 12 | 16 |
| Gemma-3-27B-it | Matched AlphaSteer | 44 | 18 | 9 | 14 | 15 |
| Gemma-3-27B-it | CSF | 54 | 13 | 7 | 7 | 19 |
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 changed | Unsafe assistance | Safe completion |
|---|---|---|
| Bypass the update | 30% | 56% |
| Apply the safety update | 20% | 65% |
| Block that same update | 30% | 56% |
| Replay the saved update | 20% | 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 edit | Safe completion at 0 | At strength 0.75 | Strength vs. completion¹ |
|---|---|---|---|
| Task-pressure release | 55% | 71% | +0.70 |
| Equal-norm random edit | 55% | 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.
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
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−)
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=Udiag(σi)V⊤,gi=∥si+ti∥2si+ti
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+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=nX1∑j=1nXdjdj⊤=nX1DXDX⊤,X∈{S,T}
K1a · Submitted basis comparison (Table 12, p. 29). Joint requires safety and completion of the allowed task.
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.