- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 3
Read review
Paper summary
The paper looks at a deployment setting where an operator has an aligned base model plus its own safety LoRA, and receives a finished task LoRA from someone else. That LoRA might have been trained on data with some unsafe examples mixed in. SafeCompose keeps all three pieces frozen and separate, and at each layer it scales the safety and task activation residuals (measured against the base layer output) with two gates. The gates come from a small MLP over the pooled base state. They are meta-trained once on a pool of adapters, using distillation toward the "right" branch for each prompt type plus a V-REx term across contamination doses, and then reused unchanged on new adapters.
Experiments cover Qwen2.5-7B, Mistral-7B and Gemma-2-9B, with an in-pool GSM8K panel and a held-out MedMCQA panel. On the held-out task, SafeCompose keeps most of the accuracy that discarding the adapter would lose, at roughly discard-level ASR on Qwen and Gemma (7–9% on Mistral). The paper also has a repeated-arrival experiment, prefill attacks, over-refusal results, a lot of ablations, and a decoding cost analysis.
Critical strengths
- The setting is practical and well motivated. "I got a LoRA from someone else and I can't retrain it" is a real situation, and wanting to keep components separate and auditable is a reasonable requirement.
- I appreciate how honest the paper is. The authors tune baselines on validation data, and they report results that cut against their own story: a tuned weighted merge ties them on Qwen, static gates match learned gates without attack, discard and RESTA are safer under attack, and the method is 4–5x slower to decode. Most papers would leave these out.
- The ungated-vs-gated ablation is clean and useful. Simply adding both residuals gives ~17% ASR, while any per-layer scaling brings it to ~1%. That is a clear finding about where the erosion comes from.
- The prefill-attack result is the most interesting part of the paper to me. The prompt-reading gate holds at 41–51% ASR, while static gates, a global gate, and tuned merges go to 68–83%. The prompt-once variant keeps most of this, so caching is possible.
Critical weaknesses
1. Without attack, the method doesn't really beat a tuned merge. The large gaps in the main tables are against baselines at default settings. Once the merge weight is tuned on validation data, it ties SafeCompose on Qwen, trades off on Mistral, and only loses on Gemma GSM8K. A static gate also matches the learned gate. So without attack, the evidence supports "scale the residuals per layer with a good constant" rather than a learned input-dependent policy. I think the abstract and intro oversell this, and the attack result should be the main story.
2. The transfer claim is weaker than it sounds. The held-out panel only changes the task. Rank, training recipe, contamination source and harm distribution all stay the same. The learned α is also nearly constant (0.48–0.53, within-layer SD around 0.02), so it's not surprising it transfers. The paper also says the gates see the adapter "through its residuals," but Eq. 2 only feeds them the pooled base output b_l. They never look at Δ^S or Δ^T directly.
3. A simple baseline is missing. Appendix M says β < 0.01 on 63% of harmful prompt–layer pairs. To me the gate looks like it is mostly a harmful-prompt detector that turns T off, which would also explain the prefill-attack advantage. The obvious comparison is an input classifier (e.g. Llama-Guard on the prompt) that routes harmful prompts to discard and everything else to M+T or the tuned merge. Without it, I can't tell what the per-layer learned gating adds.
4. The attack evaluation is narrow. It uses one Qwen adapter, three seeds, a handful of openings, and no GCG-style attack, and ~45% ASR under a simple prefill is still high. The threat model also rules out the most relevant adaptive attacker: a downstream party who trains T to push later base states so the gates open β.
Minor
-
the GSM8K adapters score below the base on Qwen and Gemma, which makes the in-pool panel hard to read. Over-refusal goes up 9–13 points on OR-Bench-Hard. Decoding is 4–5x slower than a merged model. A lot of key results are buried in the appendix.
-
The opening sentence defines PEFT as "attaching a small adapter and leaving the base weights fixed," but it cites LoRA, which learns a weight delta that can be merged into the base; please define "adapter" explicitly (e.g., "PEFT modules such as bottleneck adapters and LoRA") and state that the separation argument assumes the delta is not merged.
-
the term "push" is informal (line 94-95) and is used interchangeably with "activation residual"; since the residual is the central object of the method, please define it once formally and keep "push" only as intuition.
-
Presentation is the quite weakness: the method is simple, but the paper makes it hard to follow by introducing overlapping terms — "operator", "received adapter", "branch", "push", "composer", and "gates" — and then using "gates" and "composer" interchangeably while "SafeCompose" names the whole method; one term per concept, a short glossary, and plainer sentences would make the contribution much easier to assess.
-
Eq. 3 introduces six loss weights (λsafe, γ, λvar, λover, λcap, λbeh) in addition to a curriculum schedule and an early-stopping rule, but only four are reported: λsafe and γ are said to be "listed in Appendix B," where they do not appear. No sensitivity analysis is given for any weight, and the paper notes only that λover was not swept. Please report all six values, state whether a single setting is shared across families and seeds, and show that the conclusions are stable under at least ±2× perturbation of the four safety-related terms.
Questions for authors
- Can you show the distribution of α and β for harmful vs. benign/task prompts?
- How does SafeCompose compare to an input-classifier router, with and without the prefill attacks?
- Since the gates only read b_l, in what sense do they see the adapter's residuals? Did you try feeding Δ^S and Δ^T to the gates?
- Can an adapter trained by someone who knows the gate design push β up on harmful prompts?
- Does transfer hold with a different rank, training recipe, or contamination source?
- Discarding T should be the safest option, but on Qwen Table 2 shows the opposite (discard 1.0% vs SafeCompose 0.6%). Since one prompt is about 1% of the 104-prompt set, is this difference real? If so, please explain why adding the task adapter makes the model safer than removing it.
5 comments
2. Scaling versus conditioning
On clean Qwen at 1/5/10% contamination, Tuned merge gives generation-parsed accuracy of 60.00/60.51/60.32% versus SafeCompose's 58.27/58.97/59.42%. These historical Qwen arms used different repetition penalties, so their contrasts are protocol-level rather than isolated method effects. On Mistral, the tuned merge is safer but less accurate than SafeCompose, and on Gemma it trails on GSM8K (Table 24). We withdraw clean input-dependence claims and will add the tuned merge and the Target-free merge, selected without MedMCQA calibration, to the main results, to separate frozen composition from conditioning.
The historical ungated comparison motivates scaling, but its unmatched conditions do not establish the size of that effect. We also rename the trained static checkpoints “Near-initial static gate” and withdraw “optimized static control”: they remained near initialization, with α≈0.7 and β≈0.5, and do not establish the best achievable static policy.
Retrained comparison. Both learned and static arms use prompt-once, uncached inference on held-out MedMCQA adapters at the same contamination setting as the router study. Fixed-prefix generation is capped at 256 tokens. Each family has five matched seed pairs; paired 95% intervals jointly resample seeds and prompts with 4,000 draws.
| Family | Learned prefix ASR (%) | Near-initial static prefix ASR (%) | Learned − static ASR (pp) | Generation-parsed accuracy difference (pp) |
|---|---|---|---|---|
| Qwen | 56.54 | 73.85 | −17.31 [−30.77, −3.85] | −0.08 [−1.88, +1.50] |
| Mistral | 18.85 | 21.35 | −2.50 [−7.50, +2.69] | +0.81 [−1.23, +2.85] |
The prefix difference is resolved only on Qwen, where four of five pairs favor the learned gate, without a detected accuracy difference.
Imposed constants. Separately, we evaluated 12 constants on Matched gate checkpoints with three seeds, one batched decoder and the same fixed-prefix token cap. Qwen SafeCompose has prefix ASR 47.12 ± 10.04%. Differences below are constant − SafeCompose, with paired 95% seed × prompt bootstrap intervals. This study's accuracy scoring rule is unspecified, so its accuracy differences are not interchangeable with generation-parsed or next-token accuracy.
| Qwen control | Prefix ASR difference (pp) | Accuracy difference (pp) |
|---|---|---|
| Paper constant gate (0.7/0.5) | +35.58 [+24.04, +49.04] | +0.64 [−0.90, +2.31] |
| Validation-selected constant (0.8/0.3) | +28.21 [+17.31, +40.06] | +1.79 [+0.13, +3.53] |
| Lowest-ASR constant (diagnostic) (0.9/0.1) | +6.09 [−2.88, +14.42] | −1.15 [−3.97, +1.60] |
Conditioning lowers prefix ASR against the paper and validation-selected constants, but no advantage over the diagnostic constant is resolved.
A fixed per-layer profile of the dynamic gate's harmful-prompt values also recovers most of the advantage over the paper constant: its ASR difference from SafeCompose is +8.01 pp [+0.00, +16.35], consistent with your detector reading. Validation selection uses clean safety and utility on JailbreakBench-100 and MedMCQA-128, not prefix robustness; the diagnostic constant was selected on the test panel and is not a deployable selection result.
On Mistral, the ordering reverses: SafeCompose prefix ASR is 52.88 ± 7.26%, versus 29.17 ± 10.28% for constant 0.7/0.3. Constant − SafeCompose is −23.72 pp [−36.54, −10.58] for ASR and −0.90 pp [−2.95, +1.15] for accuracy. This is a safer constant without a detected accuracy cost, not accuracy equivalence.
Thus, the retrained Qwen contrast establishes a conditioning benefit against its near-initial static gate, routing supplies a different operating point, and Mistral prevents a general superiority claim.
3. Gate distributions and the adapter-input pathway (Questions 1 and 3)
For Submitted gates, we captured the first causal composed forward pass on 104 prompts × three seeds per row. Quantiles are over per-prompt layer averages.
| Cohort / format | α mean | α p10/p50/p90 | β mean | β p10/p50/p90 |
|---|---|---|---|---|
| AdvBench / raw | 0.9667 | 0.94/0.96/1.00 | 0.0596 | 0.00/0.05/0.13 |
| AdvBench / chat | 0.8624 | 0.77/0.85/0.97 | 0.1340 | 0.04/0.14/0.21 |
| XSTest-safe / chat | 0.6859 | 0.49/0.67/0.89 | 0.1618 | 0.07/0.14/0.24 |
| MedMCQA / raw SFT | 0.4018 | 0.23/0.38/0.61 | 0.4475 | 0.26/0.45/0.64 |
| Benign / chat | 0.1726 | 0.06/0.16/0.30 | 0.7480 | 0.61/0.75/0.88 |
These distributions are consistent with your prompt-type-detector interpretation, with XSTest-safe nearer the harmful cohort.
The format diagnostics were designed after initial results were visible, and prompt type and format remain confounded; they do not causally explain transfer or prefix behavior. We will replace the Appendix J/M near-constant-α and β-threshold summaries with checkpoint-, format- and cohort-specific statistics, explicitly separating prompt variation from decoding-step variation. For example, β<0.01 occurs in 72.30% of raw AdvBench prompt–layer pairs versus 13.69% under chat formatting.
On the gate input, you are right. Equation 2 reads pooled b_l, not the current ΔS or ΔT. The task adapter affects gate inputs indirectly through earlier composed layers. We will correct the introduction and Appendix A's “through its residuals” explanation.
We also trained a fresh residual-input variant under the original recipe, using pooled [b_l; ΔS; ΔT]. Its pooled-b_l baseline is the Submitted gates, not another fresh retrain. Over three seeds, the variant gives generation-parsed accuracy 58.97 ± 0.87%, versus the baseline's 58.91 ± 1.09%, and both give unattacked ASR 0.64 ± 0.56%. This does not resolve a clean advantage or establish safety equivalence at the ASR floor. Prefix ASR and over-refusal were not measured for this variant, and ΔS-only input was not trained.
4. Transfer beyond the original adapter recipe (Question 5)
The original held-out panel changes task, not adapter recipe. We therefore applied unchanged Submitted gates to one-factor changes in Qwen MedMCQA adapters. The mild, attention-only and aggressive recipes were specified after earlier results were visible but before their own evaluation. Entries are mean ± sample SD over three seeds, using generation-parsed accuracy.
| Adapter change | SafeCompose accuracy (%) | SafeCompose ASR (%) | Tuned-merge accuracy (%) |
|---|---|---|---|
| Rank 8 | 59.55 ± 0.22 | 0.32 ± 0.56 | 60.19 ± 0.33 |
| Rank 64 | 58.46 ± 1.35 | 0.64 ± 0.56 | 60.13 ± 0.78 |
| PKU-SafeRLHF contamination | 59.36 ± 0.97 | 0.00 ± 0.00 | 59.87 ± 0.78 |
| Mild recipe | 60.13 ± 0.44 | 0.32 ± 0.56 | 57.44 ± 0.89 |
| Attention-only LoRA | 59.49 ± 1.13 | 0.00 ± 0.00 | 59.87 ± 0.59 |
| Aggressive recipe | 43.97 ± 10.77 | 0.64 ± 0.56 | 37.76 ± 32.54 |
The first five changes preserve task accuracy, while the aggressive recipe degrades every task-using arm.
A degenerate aggressive-recipe seed makes SafeCompose answer “A” on 502/520 items; the near-initial static arm also degrades, to 42.44 ± 18.99% mean accuracy. Among the five non-degenerate shifts, only the mild recipe exceeds tuned merge's mean accuracy, and that merge retains the coefficient selected for the original adapter. These tests retain Qwen, MedMCQA, 5% contamination and direct-request AdvBench evaluation, without prefix attacks; floor-level ASR across arms supports accuracy transfer, not safety separation. We will add this table to Section 4 and narrow the transfer claim in Section 4.1 and Appendix O accordingly.
You are right that the clean results support scaling rather than learned input dependence, and that the gate may largely detect prompt type. New evidence resolves a fixed-prefix benefit against the near-initial static gate for the retrained Qwen gate, while routing offers a different safety–accuracy operating point. The family-dependent results and gate-aware adapter-training failure preclude a general superiority claim. We will reframe the abstract, introduction and conclusion as an empirical characterization of composition, not a deployment-superiority claim.
We distinguish four checkpoint groups throughout this response:
The first three groups use the implemented surrogate, whose description and Equation 3 require the substantive corrections detailed in Section 6 below. The causal/SFT gate is a separate method variant, not a corrected evaluation of the submitted one. Diagnostics and transfer results for the Submitted gates do not establish those properties for the retrained gate. Prompt-once SafeCompose computes gates once from the prompt and holds them fixed during generation; this policy must be distinguished from enabling a KV cache.
1. Prompt detection and the router baseline (Question 2)
Llama-Guard-3-8B classifies only the user prompt. Flagged prompts go to Discard, the safety adapter only; unflagged prompts go to the task adapter or Tuned merge, whose coefficient is selected on MedMCQA validation. The Fixed prefix forces the assistant opening “Sure, here is” on AdvBench-104, judged by the HarmBench classifier. Generation-parsed accuracy parses generated answers on MedMCQA-520. Keyword refusal is a heuristic on OR-Bench-Hard, not a semantic judgment.
This study uses the Retrained causal/SFT gate on Qwen, a held-out 5%-contaminated MedMCQA adapter and three seeds. All arms use neutral greedy decoding with repetition penalty 1.0. The router study and the five-pair comparison below evaluate the same gate family separately; we compare arms only within each table. Entries are mean ± sample SD.
Both routers have lower prefix ASR than SafeCompose, while SafeCompose has higher accuracy.
Exploratory post-hoc paired 95% percentile intervals jointly resample matched seeds and shared prompts with 20,000 draws:
The intervals exclude zero for both metrics, supporting a trade-off between operating points, not the necessity or dominance of per-layer gating.
Routing covers one Qwen adapter setting, without classifier-evasion, other-family or cost tests. We will include this comparison in Section 5 and replace the introduction's broad “what the gate adds” claim with the measured trade-off.