acceptodds
Under review as a conference paper at ICLR 2027

The Tokens Strike Back: Safety-Aware Denoising with Directional Embedding-based Revision

Abstract

The Safety-Aware Denoiser (SAD) is a recent, training-free, safety-guidance framework for text diffusion models. During inference, SAD steers text samples toward safe regions of the token probability space. In this paper, we propose *SADDER (Safety-Aware Denoising with Directional Embedding-based Revision)*, which improves the safety and quality trade-offs of SAD by exploiting a safety-guidance signal in the embedding (or activation) space for each token position, then updating the token positions using the SAD update rule multiplied by the guidance signal. This multiplication acts as a gating mechanism: tokens with a high unsafe signal are more likely to be modified during inference following the SAD update rule. However, such an update does not automatically preserve the conditional mean, i.e., the denoiser of a fixed target distribution. We therefore theoretically derive new position-wise target distributions for which the modified denoiser is exact. Furthermore, we derive conditions under which the position-wise target distributions match a global target distribution. We test our *SADDER* method on popular Masked Diffusion Language Models (MDLMs) and Uniform-State Diffusion Models (USDMs) across unsafe continuation and jailbreak settings, and show that SADDER achieves a better balance between safety, generation quality, and inference cost than existing methods.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
diffusion language models, safe text generation, inference-time guidance, selective denoising, remasking

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_P5vxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies safety guidance for text diffusion models and proposes SADDER, an extension of SAD. SAD steers token predictions away from unsafe-reference distributions, but applying this correction uniformly across eligible positions can degrade generation quality. SADDER uses intermediate activations to identify potentially risky token positions and selectively applies SAD’s probability-space correction to those positions. The paper also characterizes conditions under which the modified predictions correspond to position-wise safe denoisers. Experiments on masked and uniform-state diffusion models evaluate safety, generation quality, and generator evaluation counts in unsafe continuation and jailbreak settings. The results show improved benign-generation quality relative to full SAD.

Critical strengths
  • The proposed method can be applied to pretrained models without additional training. This makes it a practical approach to inference-time safety guidance.
  • The paper provides theoretical analysis that characterizes when the modified predictions correspond to position-wise safe denoisers. It clarifies the method’s properties and assumptions.
Critical weaknesses
  • The contribution is incremental, and its generality remains limited. SADDER extends SAD by adding activation-based position selection, while its main correction mechanism relies on SAD. Although the position-selection strategy could potentially be applied to other safety-guidance methods, such combinations are not discussed. Consequently, the demonstrated contribution is limited to improving SAD, and its effectiveness as a more general safety-guidance principle remains unclear.
  • The evaluation of generation quality is insufficient to substantiate a central claim of the paper. The claimed improvement in the safety–quality trade-off relies primarily on GPT-2-large perplexity and BERTScore. Perplexity measures predictability under an external language model and does not necessarily capture human-perceived fluency, coherence, or usefulness. Similarly, BERTScore measures similarity to reference continuations rather than overall response quality. Although the authors acknowledge these limitations, additional validation, such as human evaluation, is needed because preserving generation quality is a central contribution.Could the authors report actual runtime measurements and compare the methods under matched NFE budgets, with candidate counts and denoising steps adjusted appropriately for each method? NFE does not capture auxiliary computation, such as the unsafe-reference likelihood calculations inherited from SAD, and some baselines achieve better safety or quality with additional generator evaluations. These comparisons would help clarify SADDER’s practical efficiency advantages.

How sensitive is SADDER to the size, composition, and source of the unsafe reference set? Could the authors provide ablations on these factors to clarify the method’s robustness and applicability beyond the selected reference sets?

Questions for authors
  • Could the authors report actual runtime measurements and compare the methods under matched NFE budgets, with candidate counts and denoising steps adjusted appropriately for each method? NFE does not capture auxiliary computation, such as the unsafe-reference likelihood calculations inherited from SAD, and some baselines achieve better safety or quality with additional generator evaluations. These comparisons would help clarify SADDER’s practical efficiency advantages.
  • How sensitive is SADDER to the size, composition, and source of the unsafe reference set? Could the authors provide ablations on these factors to clarify the method’s robustness and applicability beyond the selected reference sets?
5 comments
AuthorsOpenReview ↗

Response to Reviewer P5vx

Thank you for the review and the questions on cost and the reference set.

W1 (Contribution)

SADDER's contribution is to make safety corrections selectively: at every step, it decides from the model's own hidden states which positions to correct and leaves the others exactly as the model would write them. As shown in General response 2, on benign prompts, it leaves 10–59% of outputs identical to the unmodified model's (full SAD: at most 0.1%), and at equal benign fluency it is safer than full SAD on LLaDA and level or slightly safer on Dream.

The selection can be applied beyond masked models: the same gate runs on UDLMs (General response 1). We tried it inside best-of-N to pick the positions where candidates differ. To be precise, this selective best-of-N still samples N answers and has N judge calls, but the candidates share the same random draws everywhere except at the positions the gate flags. So the N answers differ mainly where the gate thinks the text is risky, and the search only tries alternatives there. The idea was to spend the search where it is risky and leave the rest of the text untouched. However, when testing for N=4, it did not help, being 1–9 points less safe than plain best-of-4, because search already leaves benign text mostly intact (benign fluency 1.59–1.67, unmodified 1.60–1.65). Our reading is that the gate pays off where the correction itself damages text, and Section 6 will present it as a position-selection step for such corrections.

W2 (Quality metrics)

We added an LLM fluency judge (Llama-3.1-8B, 0–2, checked against a 70B judge), a pairwise judge, Prometheus 2, and a blind human comparison of benign outputs by the author and an independent annotator (see General response 2). All of them prefer SADDER's benign text to full SAD's, and the independent annotator chose SADDER in 71% of the decided pairs. With the fluency judge, we compared both methods at equal benign quality, where SADDER is safer on LLaDA and level or slightly safer on Dream.

Q1 (Runtime and matched NFE)

Seconds per sample on one L40S, 32 tokens, 32 steps:

UnmodifiedFull SADSADDERBoN-8FK-8
LLaDA1.321.561.5410.5811.79
Dream1.431.691.6911.4811.58

SAD's reference computation adds about 18% of the base model's generation time. The gate adds nothing measurable, since it reads a hidden state of the same forward pass. For equal NFE we shortened the search trajectories to 32 generator evaluations and judged the harmful-prompt outputs:

LLaDA, RTP prompts

MethodUnsafe %Fluency (0–2)
Unmodified39.21.28
SADDER29.71.15
BoN-2, 16 steps29.91.09
BoN-4, 8 steps21.90.86

LLaDA, BeaverTails prompts

MethodUnsafe %Fluency (0–2)
Unmodified37.20.96
SADDER31.30.87
BoN-2, 16 steps29.20.85
BoN-4, 8 steps19.60.69

Dream, RTP prompts

MethodUnsafe %Fluency (0–2)
Unmodified45.21.32
SADDER39.01.19
BoN-2, 16 steps36.31.11
BoN-4, 8 steps28.20.88

Dream, BeaverTails prompts

MethodUnsafe %Fluency (0–2)
Unmodified39.40.97
SADDER34.70.94
BoN-2, 16 steps30.80.87
BoN-4, 8 steps23.10.74

At equal cost, Best-of-N is safer but SADDER keeps text quality. The difference is clearest on benign prompts, where short search trajectories drop fluency to 1.02–1.35 against 1.46–1.63 for SADDER. We will revise the introduction's cost claim accordingly.

W3, Q2 (Reference set)

Size: on LLaDA/RTP (unmodified 39.2% unsafe), every size helps and both SAD and SADDER stay within about a point of each other.

Unsafe references1003001,0002,000
SADDER35.331.834.229.7
Full SAD35.933.035.430.7

Source: we swapped references and gate direction between datasets (RTP prompts, unsafe %; a separate run with its own unmodified baseline).

Unsafe referencesGate direction built fromLLaDADuo
none (unmodified)–37.748.3
RTPRTP27.540.3
RTPBeaverTails30.341.2
BeaverTailsRTP or BeaverTails38.0–40.255.7–55.8

The gate direction transfers across datasets, but the unsafe references must match the prompts' domain, as in SAD, since its correction acts on the references' tokens.

Composition: with benign text or a uniform estimate in place of the unsafe references, the unsafe rate stays within 1 point of unmodified, so the gain comes from the unsafe content of the references.

We would like to thank you for taking the time to review our paper and for your valuable comments. As the discussion period is ending soon, we wanted to follow up and hope that our response has addressed the concerns raised in your review, including the runtime, matched-NFE, and reference set experiments you requested. Please let us know if anything remains unclear, as we would be happy to clarify. We sincerely appreciate your feedback and the opportunity to improve our work.

Thank you for the detailed response and additional experiments. These address some of my concerns. However, my main concern about the limited generality due to SADDER’s dependence on SAD remains. The additional experiments are interesting and strengthen the evidence for its value as an extension of SAD, but its contribution beyond SAD remains insufficiently established. I therefore maintain my score.

AI use disclosure: I apologize for omitting the AI-use documentation field. I used AI tools to polish the writing and check the accuracy of the content. I confirm that I read the paper myself and independently formed the assessment and final recommendation.

Reviewer_tT4LOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper proposes SADDER, a training-free safety guidance method for text diffusion models. It builds on SAD by using hidden-state differences between safe and harmful responses to identify potentially unsafe token positions. Instead of applying SAD guidance to all positions, SADDER only modifies positions selected by the activation-based score. The authors also provide a theoretical analysis of the modified denoiser and evaluate the method on masked and uniform-state diffusion models on unsafe continuation and jailbreak benchmarks.

Critical strengths

Strengths

  1. The proposed method uses a simple midpoint-centered activation score to identify potentially unsafe token positions. This allows SAD guidance to be applied selectively rather than to all positions.

  2. The paper provides a theoretical analysis of the modified denoiser, including the conditions under which it corresponds to the conditional mean of a position-wise safe distribution and when these distributions agree with a global target.

  3. The method separates the selection of token positions from the safety correction itself. This allows the reuse of the existing SAD guidance mechanism and applies the approach to both masked and uniform-state diffusion models.

Critical weaknesses

Weakness

  1. Limited Methodological Novelty: The proposed SADDER mainly combines two existing approaches: the Safety-Aware Denoiser (SAD) [1] and contrastive activation-based safety detection (CSD) [2]. The main modification is a simple midpoint-centered binary gate that determines where to apply the existing SAD correction. It is unclear whether this combination provides sufficient methodological novelty beyond prior work.

  2. Inconsistent Safety Improvements: The experimental results do not consistently demonstrate better safety than the original SAD method. In the unsafe continuation experiments, SADDER outperforms full SAD in only one of the four model-dataset settings. For example, on Dream/BeaverTails, SADDER has a substantially higher unsafe rate (34.1% vs. 22.7%). The jailbreak results are also mixed, with limited improvements against DIJA. These results weaken the claim that the proposed method offers a better safety-quality trade-off across different models and settings.

  3. Limited Practical Relevance of the Theoretical Analysis: The theoretical results rely on assumptions that are not satisfied by the actual implementation. In particular, Theorem 3.1 requires the guidance strength to match the posterior unsafe odds, whereas SADDER uses a binary activation-based gate and an approximate guidance weight. The unsafe denoiser assumed in the proof also differs from the practical estimator. Therefore, the theoretical results do not establish safety guarantees for the proposed algorithm.

References

[1] Yusuf et al. The Safety-Aware Denoiser for Text Diffusion Models. ICML 2026.

[2] Lee and Han. Adaptive Steering and Remasking for Safe Generation in Diffusion Language Models. 2026.

Questions for authors

Please see Critical Weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer tT4L

We thank the reviewer for their careful review. Our three general responses cover the novelty claim in General response 1, new evaluation methodology in General response 2, and how the gate relates to Thm 3.1 in General response 3.

W1 (Novelty beyond SAD and CSD)

Our contribution is selective safety guidance and a safety inference-time method that transfers to UDLMs. SAD's correction applies at every masked position, superseding the model's own predicted guess of the next token. CSD uses a similar contrastive direction to edit hidden states and to remask tokens; on uniform-state models, which have no mask token, we run only its hidden-state edit (CSD†). SADDER only reads the direction, so it runs unchanged on them.

The benefit of selectivity shows once quality is measured, which the new evaluation does. On benign prompts SADDER returns the unmodified output token for token in 10–59% of cases (full SAD: at most 0.1%), and at equal benign fluency it is safer than full SAD on LLaDA and level or slightly safer on Dream (General responses 1 and 2).

W2 (Safety compared with full SAD)

Table 1 compares each method at its own selected strength, so a stronger full SAD looks safer while it damages more text. We ran both methods over the same 11 strengths and compared them at equal benign fluency (General response 2, comparison with a weaker full SAD). At equal quality SADDER is safer on both LLaDA settings and level or slightly safer on Dream. Full SAD becomes safer than SADDER only at strengths where its benign fluency is 0.10–0.24 below SADDER’s in the same setting.

We tested why Dream/BeaverTails is our weakest setting. From partly generated texts kept early in sampling (100 per seed, two seeds), we restarted the sampler with a correction at every masked position for four steps. Judged on prompt and response (the unsafe rate of unmodified sampling on these 100 prompts is 37% when the response is judged alone, as in the tables), starting at step 0, the share of unsafe endings fell from 51.6% (in both seeds) to 44.0% and 44.4%; starting at step 4, it went from 49.1% to 46.8% and from 53.7% to 50.1%, within sampling noise. So on Dream/BeaverTails the outcome is largely set in the first few steps after the prompt, while the gate is still closed at every position in 91% of these texts. We will state this limitation in the paper.

On DIJA the gate does recognise the attack: it opens at 58% (LLaDA) and 85% (Dream) of masked positions, against 23% and 9% for the same requests asked plainly. The limit is SAD's correction itself, which does not lower attack success even when applied at every position.

Attack success under DIJA (HarmBench judge) and judged fluency of the responses (Llama-3.1-8B, 0–2):

LLaDA-8B-Instruct, DIJA

MethodAttack success %Fluency (0–2)
Unmodified78.71.03
Full SAD77.01.05
SADDER79.01.05
CSD†48.70.37
DiffuGuard42.70.21

Dream-7B-Instruct, DIJA

MethodAttack success %Fluency (0–2)
Unmodified82.00.93
Full SAD85.31.02
SADDER77.71.00
CSD†68.00.42
DiffuGuard33.30.13

CSD and DiffuGuard lower attack success, but their judged fluency falls to 0.13–0.42.

W3 (Practical relevance to theory)

We agree that Thm. 3.1 gives no guarantee for the sampler used; as Remark 1 states, it identifies the quantity the gate should approximate, and the centred score is a surrogate for that quantity without formal guarantees. The revision will state its assumptions, including the exact unsafe denoiser. What the theorem asks of a practical gate is to correct where the partly generated text is likely to turn harmful, and General response 3 measures exactly this. SADDER's score ranks harmful above safe partly masked texts at every masking level, above each of 200 random directions, and prompts it scores highest give unsafe answers two to four times as often as those it scores lowest. The theorem’s information sits in the gate’s decision while the guidance strength is a separate scale, set as in SAD.

We would like to thank you for taking the time to review our paper and for your valuable comments. As the discussion period is ending soon, we wanted to follow up and hope that our response has addressed the concerns raised in your review, including the comparison with full SAD at equal text quality. Please let us know if anything remains unclear, as we would be happy to clarify. We sincerely appreciate your feedback and the opportunity to improve our work.

As a friendly reminder before the discussion period closes, we wanted to point you to two experiments we have added since our last message. They bear on your first concern, about what SADDER contributes beyond combining SAD and CSD. They are described in our first follow-up and second follow-up to Reviewer P5vx and in the new Appendix E.7 and E.8 of the revised PDF (in blue).

First, we removed SAD's update and applied SADDER's gate, unchanged, to a different correction that edits hidden states. The unsafe rate falls from 39.2% to 22.9% on LLaDA and from 45.2% to 32.8% on Dream, while the fluency of outputs on benign prompts stays at the level of the unmodified model. Compared with the CSD† baseline, which relies on remasking, the unsafe rate is similar on LLaDA and lower on Dream, with much better text quality.

Second, with no steering/correction at all, we used SADDER's score to stop unpromising best-of-4 candidates early. This saves 37.5% of the generator evaluations and is safer than stopping candidates at random in three of the four settings we tested.

These results suggest that SADDER's selection rule and its underlying signal are useful beyond SAD's update. Please let us know if anything in our responses remains unclear, as we would be happy to clarify. Thank you again for your time.

Reviewer_rMmYOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper extends SAD, a training-free safety guidance method for text diffusion models. The idea is to use a harmful-vs-safe direction in the hidden states to find the risky token positions, and to apply the SAD correction only there. The authors also give a theorem on when this gated denoiser is an exact safe denoiser. Experiments are on masked and uniform-state diffusion models, for unsafe continuation and for diffusion-specific jailbreaks, plus an over-refusal test.

Critical strengths

Safety for diffusion LMs is an important topic and still not much explored. The method is simple, needs no training and costs the same as SAD, which is much cheaper than best-of-N or FK steering. I also appreciated that it is tested on both masked and uniform-state models, and that the authors are upfront about the cases where it does not help. The reduction in over-refusal is a nice result.

Critical weaknesses

(1) Novelty is limited. In practice the method combines the contrastive safety direction of Lee & Han with the SAD update, using the direction as a gate. Centering the score at the midpoint is a small change with respect to Lee & Han.

(2) I am not convinced by the safety results. What SADDER seems to improve over full SAD is mostly quality, not safety. On several settings full SAD is actually safer, and the quality gains in Fig. 1 look quite small. Harmful perplexity also gets noticeably worse than unmodified sampling. On the DIJA jailbreak SADDER does very little. In the end it is not clear to me when one should prefer SADDER over simply tuning SAD.

(3) The theory does not add much. Thm. 3.1 basically says the denoiser is exact if the gate equals the posterior unsafe odds, which feels close to a definition, and the authors admit in Remark 1 that the real gate has no guarantee. It would be more interesting to understand why a hyperplane in hidden space should approximate this quantity.

(4) The evaluation is somewhat weak. Safety is judged by a single classifier, and quality is measured with BERTScore and GPT-2 perplexity, which say little about fluency or helpfulness. Generations are also short. On top of that, there are several variants and many hyperparameters, and the best configuration changes from one setting to another, which makes the overall picture hard to follow.

Questions for authors
  1. In which cases would you recommend SADDER over full SAD with a weaker guidance strength? A safety vs quality comparison of the two across strengths, on all models, would help.

  2. How sensitive is the method to the chosen layer and to the data used to build the harmful/safe direction?

  3. Could you add a second safety judge or a small human study, and a better quality metric?

  4. Why does SADDER help so little on DIJA, while other methods do reduce the attack success rate?

5 comments
AuthorsOpenReview ↗

Response to Reviewer rMmY

Thanks for the detailed review.

W1 (Novelty)

SADDER turns a full-sequence safety correction into a selective one driven by the model's own hidden states on which positions to correct. The other positions are left as how the model would have written them. Lee & Han use a similar direction to edit the hidden states and remask, whereas SADDER only reads it and decides where to apply SAD's correction. See General response 1 for details.

W2, Q1 (When to prefer SADDER over a weaker SAD)

See General response 2 for our detailed response. At a high level, we suggest users choose SADDER when response quality on benign prompts matters. Use full SAD at a high strength when the lowest unsafe rate matters more than text quality.

Over the same 11 strengths, when full SAD is tuned to be as fluent as SADDER, it is less safe on both LLaDA settings and about level or slightly safer on Dream. When it is tuned to be as safe, its benign fluency drops.

On harmful prompts, a lower unsafe rate can come from text that is no longer legible. To test this, we split every harmful prompt output into three groups: unsafe, safe and fluent (fluent here is judged by fluency of at least 1 on the 0-2 scale introduced), and safe but not fluent. Both methods run at η=100,1,000,104,104\eta = 100, 1,000, 10^4, 10^4 (on LLaDA/BeaverTails the paper uses 3,000 for SADDER).

LLaDA, RTP prompts

MethodUnsafe %Safe and fluent %Safe, not fluent %
Unmodified39.247.613.2
Full SAD30.739.729.6
SADDER29.747.422.9

LLaDA, BeaverTails prompts

MethodUnsafe %Safe and fluent %Safe, not fluent %
Unmodified37.229.033.8
Full SAD27.733.838.4
SADDER31.325.942.8

Dream, RTP prompts

MethodUnsafe %Safe and fluent %Safe, not fluent %
Unmodified45.244.110.8
Full SAD34.642.223.3
SADDER39.043.217.8

Dream, BeaverTails prompts

MethodUnsafe %Safe and fluent %Safe, not fluent %
Unmodified39.432.628.1
Full SAD20.928.750.3
SADDER34.732.333.0

Both methods turn some unsafe answers into less fluent ones, since the correction acts mid-sentence. Full SAD lowers the safe and fluent share by up to 8 points (LLaDA/RTP 47.6 to 39.7) and leaves much more text safe but not fluent (50.3% against 33.0% for SADDER on Dream/BeaverTails). SADDER keeps it within a point of unmodified sampling in three settings (47.4, 43.2, 32.3). LLaDA/BeaverTails is the exception, where full SAD keeps safer and more fluent output than SADDER.

W3 (Why a hyperplane)

The direction is the difference between the mean hidden states of harmful and safe continuations of the same prompts, so the score is a linear read-out of where the model’s own states say the text is heading. We describe extra experiments we conducted in our General response 3 to test how useful SADDER's embedding directions are: on held-out references masked at the level from 95% to 35%, AUROC rises from 0.58–0.64 to 0.77–0.89 as text is revealed, above each of 200 random directions.

W4, Q3 (Evaluation)

We show the newly added evaluation metrics in our General response 2. In summary, we added Llama-Guard-4 and WildGuard as safety judges (they rank the methods as Llama-Guard-3 does), an LLM fluency judge, a pairwise judge, Prometheus 2, and a blind human comparison of benign outputs. We also generated longer outputs, keeping the strengths chosen at 32 tokens. Guidance covers the first half of the steps at 128 tokens and the first eighth at 256 tokens (300 prompts × 3 seeds there). Unsafe %, Llama-Guard-3:

Model, promptsTokensUnmodifiedFull SADSADDER
LLaDA, RTP12846.936.438.7
LLaDA, RTP25649.738.241.2
LLaDA, BT12842.131.332.8
Dream, RTP12850.740.645.7
Dream, BT12841.016.934.3

SADDER keeps most of full SAD's reduction on LLaDA and Duo (General response 2, part 2) and less on Dream. On the number of variants: every new result uses the gate alone, while the warm-up and remasking are ablations.

AuthorsOpenReview ↗

Continued Response to Reviewer rMmY

Q4 (DIJA)

The gate recognizes the attack (it opens at 58–85% of masked positions, against 9–23% on plain requests), but SAD’s correction does not lower attack success even when applied at every position. Our response to Reviewer tT4L (W2) shows the table.

Q2 (Sensitivity to layer and data)

We choose the layer where the gate has the largest difference between harmful and benign prompts. This rule does not need a safety judge, picks the paper's layer in all four settings, and picks the same layer in 19 of 20 random 90% subsets of the references. See details in our response to Reviewer P5vx (W3, Q2).

We would like to thank you for taking the time to review our paper and for your valuable comments. As the discussion period is ending soon, we wanted to follow up and hope that our response has addressed the concerns raised in your review, including the comparison against a weaker full SAD and the additional judges and human study you suggested. Please let us know if anything remains unclear, as we would be happy to clarify. We sincerely appreciate your feedback and the opportunity to improve our work.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General Response 1 - Novelty

We thank the reviewers for their careful review and questions on our method. We have uploaded a revised PDF in which all changes are in blue and the new experiments are in Appendix E. We'll address overlapping novelty concerns in this response.

Existing inference-time steering methods for diffusion language models tend to destroy the output for the sake of reducing ASR. SAD corrects every masked position, CSD edits hidden states and remasks tokens over a number of iterations, and DiffuGuard also remasks. SADDER introduces a way to steer generations with much smaller quality impacts. It does this by not modifying the original model's sampler on positions that were not flagged by the safety score. The new quality evaluation (General Response 2) shows what selectivity gives us.

On benign prompts SADDER returns the unmodified model's output token by token in 58% (RTP tested on LLaDA), 10% (BT tested on LLaDA), 59% (RTP tested on Dream) and 30% (BT tested on Dream) of cases, against at most 0.1% for full SAD. We also compare the method at equal benign fluency, meaning the fluency of their outputs on benign prompts as rated by a Llama-3.1-8B judge on a 0-2 scale. At equal benign fluency it is safer than full SAD on LLaDA and level or slightly safer on Dream, and judges and people prefer its benign text. A weaker full SAD, meaning SAD run at a lower strength η\eta, does not reproduce this on LLaDA. It changes every position (albeit, slightly less), so it either keeps more unsafe outputs or loses fluency. See General response 2 for more detail.

SADDER is for the practitioner who wants a training-free method that can improve safety without sacrificing the quality of the base model.

When would one want to use SADDER over SAD? These results identify where SADDER is the better choice: when the model must keep answering benign prompts as it normally would, SADDER gives the largest safety gain at that quality that we found, without training and at close to the base model’s cost. When the lowest unsafe rate matters more than text quality, full SAD at a high strength, or search based methods (Best-of-N, FK Steering) at several times the cost, remains the better option.

First inference-time steering safety study on UDLMs. To our knowledge, this is the first evaluation of inference-time safety guidance on text UDLMs, and SADDER’s gate carries over to them without change. This is especially useful as current strongest inference-time safety techniques we compare on (CSD, DiffuGuard) rely on remasking for their ASR performance, which is not an option for UDLMs as there's no mask token to re-open.

AuthorsOpenReview ↗

General Response 2 - Part 1/2 - Quality Evaluation

We thank the reviewers for highlighting the evaluation limitations on PPL and BERTScore, and we agree. We'll address overlapping evaluation concerns in this response.

For this evaluation we consider six automatic judges and a small human study. Three of the judges are safety classifiers on harmful prompts which give an unsafe rate, while three judges are on benign prompts. Two of these quality judges rate the fluency of the output and one of them picks the better of two outputs. To supplement the current ASR judge (Llama-Guard-3) for harmful prompts, we added Llama-Guard-4 and WildGuard to the evaluation. They both evaluate harmful prompts and give an unsafe percentage.

One of the fluency judges is Llama-3.1-8B. It is instructed to rate each response to a benign prompt from 0 (unreadable) to 2 (fluent). The instruction asks the model whether the provided text is grammatical or repetitive, so higher is better. To have confidence in this smaller model, we re-scored the outputs of an 11-strength sweep (23 runs, 300 benign outputs each) with Llama-3.1-70B. On individual outputs the two judges correlate at 0.67–0.69 while on each run’s average fluency they correlate at 0.97–0.995, so they rank the methods the same way. The second benign fluency judge is Prometheus 2. It was trained specifically to grade other models' text against a rubric. We give it a fluency rubric, so it checks the Llama judge's result with a different model on a finer scale. Prometheus 2's rating is from 1 to 5. The final quality judge we consider is a Qwen2.5-72B model. This judge compares responses to the same prompt from SADDER and full SAD side by side, and selects the more coherent response. This gives us another dimension of fluency that doesn't rely on an absolute score. All three quality judges are asked to ignore whether the content is harmful, for which the safety classifiers judge separately. The full judge protocol, with references, is in Appendix E.

We also added a small blind human study with two annotators who did not know which method generated which text. This study followed the same protocol as the Qwen model.

The following tables include the three safety judges evaluated on the paper's harmful prompts outputs, while the quality columns (which include the fluency score, the Prometheus 2 score, and the pairwise preference percentage from the Qwen2.5-72B model) score benign prompts on the runs at the same strengths. Please do note that the first tables re-judge the paper’s outputs while all other tables come from new runs at seeds 42–44 on the same prompts, so values for the same setting differ by up to 3 points.

LLaDA-8B, RTP prompts

MethodUnsafe %, Llama-Guard-3Unsafe %, Llama-Guard-4Unsafe %, WildGuardLlama-3.1-8B Fluency (0–2)Prometheus 2 (1–5)Pairwise: preferred by Qwen (%)
Unmodified38.631.366.31.623.26
Full SAD32.424.658.81.262.5932.0
SADDER29.422.251.11.613.2667.3

LLaDA-8B, BeaverTails prompts

MethodUnsafe %, Llama-Guard-3Unsafe %, Llama-Guard-4Unsafe %, WildGuardLlama-3.1-8B Fluency (0–2)Prometheus 2 (1–5)Pairwise: preferred by Qwen (%)
Unmodified38.635.058.41.623.26
Full SAD27.123.044.01.282.7637.3
SADDER29.928.553.41.463.1360.7

Dream-7B, RTP prompts

MethodUnsafe %, Llama-Guard-3Unsafe %, Llama-Guard-4Unsafe %, WildGuardLlama-3.1-8B Fluency (0–2)Prometheus 2 (1–5)Pairwise: preferred by Qwen (%)
Unmodified46.836.674.21.643.37
Full SAD33.427.764.31.362.8533.3
SADDER35.930.267.21.633.3966.3

Dream-7B, BeaverTails prompts

MethodUnsafe %, Llama-Guard-3Unsafe %, Llama-Guard-4Unsafe %, WildGuardLlama-3.1-8B Fluency (0–2)Prometheus 2 (1–5)Pairwise: preferred by Qwen (%)
Unmodified38.733.861.31.643.37
Full SAD22.718.149.21.442.9942.0
SADDER34.128.957.11.583.3557.7

As seen by the above tables, the three safety judges give different absolute rates but the same ordering of the methods. On quality, the judges show that SADDER's benign text scores close to the unmodified model's fluency while SAD at the strengths reported in the paper is weaker. Over all models/datasets, the Qwen2.5-72B model preferred SADDER's response. The human study also showed the same trend:

AnnotatorPairsSADDER betterFull SAD betterTie
Author2006336101
Independent annotator5029129

The independent annotator preferred SADDER in 71% of the pairs where they picked a winner (sign test p = 0.01).

AuthorsOpenReview ↗

General Response 2 - Part 2/2 - Duo and a Weaker Full SAD

This continues part 1, with the same judges. Applying the fluency metrics on Duo with RTP prompts, two output lengths at η=1.5×105\eta=1.5\times10^5 and gate at layer 9, we get similar results:

Duo, RTP prompts, 128 tokens

MethodUnsafe %, Llama-Guard-3Fluency (0–2)
Unmodified48.30.46
Full SAD39.40.41
SADDER40.70.47

Duo, RTP prompts, 512 tokens

MethodUnsafe %, Llama-Guard-3Fluency (0–2)
Unmodified45.00.67
Full SAD29.00.57
SADDER29.70.71

What about a weaker full SAD against SADDER? A weaker full SAD might match SADDER's quality and ASR reduction. We ran both methods over the same 11 strengths and compared a full SAD run at the strength where its benign text is as fluent as SADDER's, and its weakest strength that is as safe as SADDER. The safety judge is Llama-Guard-3, and we measure the benign fluency from the 8B judge. SADDER’s row is its safest strength whose benign fluency is no lower than full SAD’s, so on LLaDA/BeaverTails and Dream/RTP it differs from the paper’s (1,000 and 10510^5, against 3,000 and 10410^4).

As the results below show, reducing full SAD’s strength does not uniformly recover SADDER’s observed safety–fluency operating points. The differences are most pronounced on LLaDA and smaller on Dream.

LLaDA-8B, RTP prompts

MethodStrength η\etaUnsafe %Fluency (0–2)
Unmodified–39.21.62
Full SAD, as fluent as SADDER0.337.21.57
SADDER10029.71.61
Full SAD, closest to SADDER's safety10030.71.34

LLaDA-8B, BeaverTails prompts

MethodStrength η\etaUnsafe %Fluency (0–2)
Unmodified–37.21.62
Full SAD, as fluent as SADDER1045.01.43
SADDER1,00031.31.46
Full SAD, as safe as SADDER30029.31.33

Dream-7B, RTP prompts

MethodStrength η\etaUnsafe %Fluency (0–2)
Unmodified–45.21.64
Full SAD, as fluent as SADDER140.31.60
SADDER10510^537.41.62
Full SAD, as safe as SADDER1,00036.61.38

Dream-7B, BeaverTails prompts

MethodStrength η\etaUnsafe %Fluency (0–2)
Unmodified–39.41.64
Full SAD, as fluent as SADDER335.31.57
SADDER10410^434.71.58
Full SAD, as safe as SADDER1033.81.48

As one can see, when full SAD is made as fluent as SADDER, it removes fewer unsafe outputs. When it is made as safe, it loses fluency points.

Note: on LLaDA/BeaverTails every full SAD strength up to η = 100 is at least as unsafe as unmodified sampling (37.3–48.0%, at fluency 1.36–1.60), so no weaker full SAD matches SADDER there. The question’s own words are absent from the unsafe references, so a weak correction raises their probability and the answer more often starts by repeating the harmful question. Appendix Figure 4 shows all 11 strengths.

AuthorsOpenReview ↗

General Response 3 - The Gate and Thm 3.1

We agree with tT4L and rMmY that Theorem 3.1 gives no guarantee for our binary-gated sampler. Since we have no access to the oracle giving us the precise unsafe odds, we turn our attention to study whether the embedding activation direction we use in our method provides an informative risk-discrimination signal, like the ideal unsafe odds would do. In other words, we would like to evaluate what the theorem asks out of a practical gate, correcting where the partly generated text is likely to turn harmful and leaving the other positions alone. We tested this in two ways, on reference texts whose label is known, and on the model's own answers.

Can the score tell harmful from benign text while most of it is still masked?

The gate needs to decide which positions to correct before an answer is written, so its score needs to differentiate between harmful and benign text when most tokens are still masked. To test this, we took held-out harmful and benign responses to the same harmful prompts from the gate's reference data (RTP or BeaverTails, matching the setting). We masked 95% down to 35% of the continuation tokens and scored the masked positions. The last column of the following table compares the prompts. It asks whether the score can tell a harmful prompt from a benign one before the answer is written. AUROC is the chance that a harmful example scores above a benign one (with 0.5 chance). To see whether the risk comes from the learned direction or just any direction at this layer, we repeated the test with 200 random directions:

LLaDA-8B

PromptsDirection95% masked75% masked55% masked35% maskedAnswer fully masked (prompts)
RTPSADDER (layer 31)0.580.770.830.890.92
RTPbest of 200 random0.550.590.600.630.75
BTSADDER (layer 31)0.620.750.790.800.66
BTbest of 200 random0.600.670.620.620.64

Dream-7B

PromptsDirection95% masked75% masked55% masked35% maskedAnswer fully masked (prompts)
RTPSADDER (layer 27)0.590.750.820.860.89
RTPbest of 200 random0.560.600.610.630.74
BTSADDER (layer 27)0.640.760.780.770.68
BTbest of 200 random0.590.640.600.600.67

The score separates harmful/benign better as text is revealed, and SADDER's direction is above the 200 random directions in every cell (the median of these random directions is 0.51-0.54). This margin is wide on masked references and narrow on BeaverTails prompts with the continuation fully masked and the prompt/question is fully visible. Figure 7 of the revised appendix repeats the reference check at every layer.

A limitation of this experiment is that we assume all positions of harmful responses are labeled as unsafe. Since the reference data has no position-wise labels, it is challenging to design an experiment to test whether SADDER's harmful positions match the ground truth harmful positions. However, the above experiment tests whether harmfulness is readable throughout the representation, measuring overall risk predictions.

Does the score track the probability that the model's own answer turns out unsafe?

Thm. 3.1 scales the correction by the posterior probability that the finished text is unsafe, given the current partly generated text. At the first step, where the continuation is fully masked, this probability can be estimated directly. For each of 100 harmful prompts per setting, we sampled 24 answers from the unmodified model and took the share judged unsafe by Llama-Guard-3. We then split the prompts at the 20th, 40th, 60th, and 80th percentiles of SADDER's score at that step, giving 5 groups of 20 prompts. Each cell in the table is the average share of unsafe answers at that group.

Setting0–20th percentile20–40th40–60th60–80th80–100thAll prompts
LLaDA, RTP21.925.424.650.457.135.9
LLaDA, BT24.245.055.040.076.248.1
Dream, RTP29.439.238.139.264.642.1
Dream, BT21.537.753.156.287.951.3

Prompts above the 80th percentile of the score give unsafe answers 57–88% of the time, against 22–29% for prompts below the 20th. The middle groups are not always in order, since each holds only 20 prompts. The score ranks the posterior probability that the theorem asks the correction to follow, though it does not give its exact value.

Together, the two tests show that the score ranks texts by how likely they are to turn harmful, from the first step and more clearly once part of the answer is visible, though not perfectly. This is what Thm. 3.1 asks of the gate: it marks where the posterior unsafe odds are high, and the guidance strength takes the place of their exact value.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.