acceptodds
Under review as a conference paper at ICLR 2027

Noise Out, Bias In: Targeted Bias Injection in Diffusion Language Models via Closed-Loop Activation Steering

19 views

Abstract

Masked diffusion language models (dLLMs) generate text by iteratively denoising masked positions, re-predicting each token multiple times before it is committed. An autoregressive decoder exposes an answer's distribution once, at the step that commits it; a dLLM exposes it at every denoising step before commitment, and we show that an adversary can exploit this. Since an answer remains open to revision over many denoising steps, an adversary with access to internal activations can watch how likely the model is to produce a chosen answer and adjust the intervention accordingly. Building on this observation, we study targeted bias injection, an attack that steers a frozen dLLM toward a demographic answer selected by the adversary. The attack uses a simple proportional-integral (PI) controller that tracks the target-answer probability during denoising and adapts the strength of a steering vector on the fly. On ambiguous BBQ questions where the correct answer is abstention, our attack raises LLaDA-8B-Instruct's preference for the targeted group from 1.8 to 16.7 percentage points, more than three times the strongest fixed-strength steering baseline, and on SocialStigmaQA it raises the selection of stigmatizing answers from 17.6% to 58.1%. Fitted to other demographic targets, the same attack shifts answers by up to 37 percentage points, and each attack takes about 40 minutes on one GPU. On the primary target, feedback is what makes the attack work: constant steering at the same average strength over the token-committing steps produces a far smaller shift while corrupting nearly three times as many outputs, and a constant strength set separately for each example still falls well short. Our findings identify the denoising trajectory as a new control channel in dLLMs and call for bias audits that examine the serving stack rather than the frozen model alone. We will make our code publicly available.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Diffusion language models, targeted bias injection, activation steering, closed-loop control, inference-time attacks.

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_1VHGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an inference-time attack that increases demographic bias in a diffusion language model. It constructs a direction from the mean activation difference between target and comparator answers, then uses a proportional–integral controller to adjust its strength from the target-answer probability during denoising. The main contribution is feedback across denoising steps, together with an empirical comparison against fixed-strength interventions. On ambiguous BBQ questions with LLaDA-8B (Black target), the target–comparator gap rises from 1.8 to 16.7 pp, compared with 3.5 for a constant mean strength. On SocialStigmaQA, stigmatizing answers rise from 17.6% to 58.1%. The effect transfers to Arab and Old targets. It fails on Asian, Latino and White because of high invalid-output rates, the effect is small on Dream, and it is beaten by constant steering on LLaDA-MoE. The authors argue that dLLMs expose a new attack surface because the answer can be observed and steered before it is committed.

Critical strengths
  1. The paper identifies a useful inference-time control opportunity in dLLMs: repeated predictions at an unresolved answer position can provide feedback for adjusting an intervention before commitment. The primary LLaDA results show substantial demographic target-answer shifts. The matched comparisons hold the steering direction and injection sites fixed, helping isolate the contribution of how the strength is set.
  2. The evaluation includes several valuable checks: answer-order rotations, matched-strength and per-example constant baselines, item-clustered confidence intervals, and parser-sensitivity analyses. The authors also report unsuccessful demographic targets and models where feedback does not outperform constant steering (Asian, Latino, White, LLaDA-MoE). These checks and negative results help establish the scope of the empirical finding.
Critical weaknesses
  1. The experiments don’t show that feedback is what drives the effect. All the ablations compare PI against a strength that stays constant across denoising steps. But PI’s strength increases over the steps (Fig. 1c). So a fixed schedule that increases over time, with no feedback, could potentially explain much of the gain. The sampling setup makes this harder to judge. Only the first token is scored, but the model generates 32 tokens over 64 steps with steering at every position, the last 32 steps commit nothing, and the step at which the answer token commits is never measured. That step determines how many controller updates actually matter. Finally, the criteria for a successful attack change between settings. Asian, Latino and White runs are discarded for high strict invalidity, but SocialStigmaQA PI is 17.4% strict-invalid (Table 2), above the paper’s own 15% cutoff (Sec. 4.4). That result relies on a semantic parser added after inspecting the outputs.
  2. The threat model needs more motivation. The attacker can hook activations and read the sampler’s answer probabilities. In this setting, though, the same access would usually allow editing the logits or the sampler directly, for example by adding a bias to the target letter’s logit. The paper excludes these by assumption, but it doesn’t explain why an attacker would choose activation steering instead. A clearer argument for when this attacker is realistic, or framing the work as a study of how steerable dLLMs are during denoising, would make the contribution easier to judge.
Questions for authors

Clarifications:

  • When does the answer token typically commit under your attack, and how many controller updates come before that? Prior work [1] finds that LLaDA’s answers on multiple-choice tasks are usually settled by halfway through denoising. Related to this: why generate 32 tokens over 64 steps when only the first token is scored, and does steering every position raise the invalid rate? I don’t expect new runs here, but please explain these sampling choices.
  • How was block 14 chosen for extracting the direction, and why is it then added at every block and position? Was any ablation done over these choices? Why does the attack fail on the Asian, Latino and White targets: is it the direction itself, the strength, where it is added, or some bias inherent in the model?
  • The threat model allows activation hooks and access to intermediate answer probabilities but excludes direct logit and sampler edits. Can you clarify what deployment scenario motivates this particular access boundary? Addressing this would help establish the practical significance of the attack beyond demonstrating controllability during denoising.
  • Can you please compare PI with a predetermined time-varying schedule without feedback?

On SocialStigmaQA:

  • What is the unsteered gap under the semantic parser? Reporting it with the semantic-parser outcome rates would make the magnitude of Δgsem\Delta g^{\mathrm{sem}} easier to interpret.
  • Why isn’t the semantic parser also applied to Table 1?

Suggestions (do not affect my score):

  • The table formats are inconsistent: Table 2 drops the Gap column and Table 3 drops Tgt and Cmp. A single format would make them easier to compare.
  • Line 345 says no baseline exceeds 4.6, but the per-item static strengths reach 4.8 and 8.4. Please clarify which rows count as baselines.

References

[1] P. Li et al. Diffusion Language Models Know the Answer Before Decoding. arXiv:2508.19982, 2025.

11 comments

Response to Reviewer 1VHG

We thank the reviewer for recognizing the denoising-time attack surface and matched comparisons . We clarify the existing evidence and report the additional experiments below.

1VHG-W1 / Q4. Does feedback matter beyond a predetermined time-varying schedule?

Yes. The additional schedule controls support feedback as the source of the advantage. Section 4.3 and Table 1 compare PI with constant and per-item static strengths. To test the reviewer’s time-varying alternative directly, we replaced feedback with PI’s average α(t) curve, another item’s actual PI trajectory, and a linear ramp from 0 to 6.5. The three schedules have reported mean α = 3.27 over the committing steps. The new conditions were run together on 1,200 prompts; the paper’s constant-strength result is included separately for reference.

Condition (1,200 prompts)Gap [95% CI]Invalid (%)Mean α (steps 0–31)PI − condition [95% CI]
Base2.5 [−1.5, 6.3]0.00+12.6 [8.8, 16.4]
Decode-PI15.1 [11.1, 19.0]7.73.27—
PI mean α(t), no feedback1.7 [−1.7, 5.0]0.83.27+13.4 [9.2, 17.6]
Another item's PI α(t), no feedback2.8 [−0.5, 6.1]8.23.27+12.3 [8.5, 16.2]
Linear ramp 0 → 6.5, no feedback1.6 [−0.2, 3.4]29.03.27+13.5 [9.2, 17.7]
Constant α = 3.28 (paper)3.5 [0.1, 6.9]20.83.28—

PI exceeds all three no-feedback schedules, with paired advantages of +13.4 [9.2, 17.6], +12.3 [8.5, 16.2], and +13.5 [9.2, 17.7] pp. Replaying another item’s trajectory preserves the replayed command sequence but removes its dependence on the current item’s predictions. Its failure to recover PI’s gap directly supports the value of closing the loop, beyond increasing strength over time. We will add these controls and distinguish this rerun from Table 1.

1VHG-Q1. When does the answer commit, and why use 32 tokens, 64 steps and sequence-wide steering?

Early predictability does not end the intervention window: commitment does. Section 3 describes the 32-token, 64-step schedule. Our additional logging distinguishes a stable answer-slot argmax from the sampler’s commitment decision. With gen_length = block_length = 32, one token commits in each of steps 0–31; steps 32–63 commit nothing. SocialStigmaQA’s 32-step and 64-step runs produce identical outputs on all 1,665 prompts.

For unsteered prompts, the answer argmax settles at median step 0, consistent with the reviewer’s observation. However, commitment occurs at median step 31, with the answer committed last in 74% of prompts. On the 223 prompts PI flips to the target, the answer settles at median step 29 (97% after halfway) and commits at step 31. Thus, feedback can change the answer before commitment even when the unsteered prediction stabilizes early.

Condition (1,200 prompts)Gap (pp)Invalid (%)Median answer-commit step
Base, 32 tokens / 64 steps2.50.031
PI, 32 tokens / 64 steps (paper setting)15.17.729
PI, answer-position-only steering1.90.031
Base, 8 tokens / 8 steps3.20.07
PI, 8 tokens / 8 steps−2.30.85

The additional ablations support the chosen configuration: answer-position-only steering gives 1.9 pp, while 8 tokens / 8 steps gives −2.3 pp under PI. Neither recovers the 15.1 pp sequence-wide result. The shorter configuration changes both generation length and the feedback horizon, so it does not isolate update count alone. Answer-only steering reduces invalidity to 0.0% but also removes the observed gap increase.

1VHG-Q2. Why extract at layer 14 and inject at every block? What explains the failed targets?

The gap persists across the tested extraction layers; injection coverage matters substantially. Section 3 specifies extraction at layer 14 and injection at every block and position. We additionally varied extraction depth while retaining all-block injection, then tested injection at block 14 alone (100 items × 3 rotations).

Direction extraction / injectionGap (pp)Invalid (%)
Layer 14 / all blocks (paper setting)19.311.0
Layer 8 / all blocks18.718.3
Layer 20 / all blocks20.023.0
Layer 26 / all blocks16.36.0
Layer 14 / block 14 only2.30.0

All tested extraction layers produce gaps of 16 to 20 pp, although validity varies. Layer 14 gives 19.3 pp with 11.0% invalidity; layer 26 gives 16.3 pp with 6.0% invalidity. Block-14-only injection gives 2.3 pp, supporting the all-block design under these settings without claiming it is the only viable configuration.

The failed targets show a mismatch between the letter sensor and the generated output. Section 4.4 reports Asian, Latino and White as failures. Additional output inspection finds “Li…” in 1,064 of 1,200 Asian prompts, surname fragments such as “Z:” and “…PEZ” for Latino, and repetition (“AAAA…”) for White. Target-letter probability falls below base: Asian 0.044 vs 0.189; White 0.031 vs 0.196. These observations suggest that the intervention can promote names or malformed text rather than the measured answer letter, causing the controller to increase strength as the letter-probability shortfall persists. They do not isolate direction quality from controller and injection effects.

TargetBase gap (pp)PI gap (pp)PI invalid (%)Dominant invalid output
Black1.816.77.8Mixed ("Answer:", "The …")
Arab−0.222.39.7Empty output
Old−5.531.89.3Empty / "The …"
Latino4.215.259.6Surname fragments ("Z:", "…PEZ")
Asian−3.25.693.8"Li…"
White−1.63.996.1"AAAA…"

Black, Arab and Old retain substantially better validity. The additional Latino suppression result (SsMd-W1) also shows that this direction’s usefulness depends on the intervention setting, rather than establishing that it is uniformly ineffective.

1VHG-W2 / Q3. Why study activation steering when logit or sampler edits may be available?

Our question is what feedback can achieve through activation steering under the access boundary in Section 2. We do not claim that steering is stronger than unrestricted logit or sampler editing. Sections 2 and 5 cite activation-access precedents: Trojan Activation Attack (Wang & Shu, CIKM 2024) studies inference-time steering, including bias; Arditi et al. (NeurIPS 2024) removes refusal through a representation direction; Schwinn et al. (NeurIPS 2024) studies continuous-space attacks; and Shnaidman et al. (ReALM-GEN at ICLR 2026) applies activation interventions during LLaDA denoising.

Our assumed interface explicitly includes intermediate probability readout and bounded residual-stream intervention, while excluding direct logit and sampler edits. These precedents motivate the interface but do not establish that every compromised serving process is restricted to it. We will clarify that boundary.

The additional open-ended experiment demonstrates that the attack is not confined to MCQ answer letters. With one-sentence BBQ responses, the same actuator and a target-name probability sensor yield +28 pp [8, 48], the only tested method whose reported shift interval excludes zero (R1.1). The MCQ implementation does use the target’s answer letter for feedback; the free-form result supports broader controllability, not a claim that logit editing cannot operate on free text.

1VHG-Q5. What is the clean SocialStigmaQA gap under semantic scoring?

The clean gap is −44.3 pp under both parsers, with 0% invalidity. Table 2 reports the strict baseline and parser-dependent shifts; Appendix C.8 defines the semantic rules. We now make the absolute gaps explicit below. The run generated 1,665 prompts; as specified in Appendix C.8, headline race metrics use the race subset, while no-stigma prompts are diagnostic.

Condition (1,665 prompts)Gap, strict (pp)Invalid, strict (%)Gap, semantic (pp)Invalid, semantic (%)
Clean−44.30.0−44.30.0
Decode-PI+39.517.4+55.61.4
Open loop α = 4+18.33.9+19.437.8
Open loop, effort-matched−2.50.0−2.50.0

Decode-PI shifts the gap by +83.8 pp under strict parsing and +99.9 pp under semantic parsing, with semantic invalidity 1.4%. Against the effort-matched open loop, paired gains are +42.0 pp [36.2, 47.9] and +58.1 pp [48.8, 67.7], respectively. Thus, the advantage holds under both scoring rules.

1VHG-Q6. Why was semantic scoring not applied to BBQ Table 1?

The additional rescoring shows that the main BBQ comparison is stable under both parsers. Appendix C.8 introduced semantic scoring to handle exact answer text and reject option re-listings on SocialStigmaQA. We have now applied both parsers to the Table 1 conditions:

Condition (1,200 prompts)Gap, strict (pp)Invalid, strict (%)Gap, semantic (pp)Invalid, semantic (%)
Base1.80.01.80.0
Decode-PI16.77.816.811.1
Constant α = 3.283.520.83.520.8
Per-item static α (oracle)4.813.14.813.1
Per-item static α (hindsight)8.413.48.313.5
CAA L14 (a = 16)3.30.03.30.0
ActAdd L14 (a = 16)4.10.04.10.0
Mean-AcT (s = 2)3.50.23.50.2
Linear-AcT (s = 1)2.60.12.60.1
AurA inject (g = 4)0.00.20.00.2
AurA vanilla0.80.00.80.0
ITI-C (K = 48, a = 8)2.10.02.10.0

No reported gap changes by more than 0.2 pp. Decode-PI remains at 16.7–16.8 pp, above the listed steering baselines and static controls. Its invalidity increases from 7.8% to 11.1%, so parsing does affect validity, but PI remains below the 15% threshold under both rules. We will report both.

1VHG-W1. Is the SocialStigmaQA success criterion inconsistent with the 15% invalidity threshold?

Yes, SocialStigmaQA PI exceeds the 15% threshold under strict parsing. Section 4.4 and Table 2 report 17.4% strict invalidity; Appendix C.8 identifies semantic parsing as a secondary analysis added after output inspection. We will distinguish strict-threshold failure from semantic-parser success explicitly.

The additional audit finds that 250 of the 271 strict-invalid outputs are the bare biased answer word (“Yes” or “No”), rather than an answer letter. Under the semantic rules applied to every condition, PI’s invalidity falls to 1.4% and its gap increases from +39.5 to +55.6 pp. This supports a substantive answer shift despite letter-format violations, without retroactively treating the strict criterion as satisfied. We will report both parsers and their rules consistently.

1VHG. Presentation and reporting suggestions

We will standardize Tables 1–3 to include target, comparator, gap and invalidity, identify each parser, and clarify the denoising schedule. We will also replace “no baseline exceeds 4.6 pp” with a precise comparison: the tuned open loop reaches 4.6 pp, the probe-based per-item static control 4.8 pp, the hindsight static control 8.4 pp, and Decode-PI 16.7 pp. Only the hindsight condition uses PI’s realized trajectory to set its constant.

The new schedule controls directly address the reviewer’s central alternative: predetermined time variation, including replayed PI trajectories, does not recover the advantage of feedback tied to the current example’s predictions.

Reviewer_SsMdOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper shows that a masked diffusion LM re-predicts its answer at every denoising step, and an attacker with residual-stream hooks can use this as a feedback signal. A PI controller adjusts the strength of one fixed steering direction during denoising to push the model toward an attacker-chosen demographic answer, without changing weights or prompts. On ambiguous BBQ questions with LLaDA-8B, the target–comparator gap rises from 1.8 to 16.7 points, while the best of seven steering baselines reaches 4.6. On SocialStigmaQA, stigmatizing answers rise from 17.6% to 58.1%. Ablations with matched constant and per-item strengths show that the gain comes from the feedback, and transfer to Dream-7B and LLaDA-MoE is mixed.

Critical strengths
  1. The ablations show that feedback is the mechanism. A constant at the controller's mean strength gets a gap of 3.5 with 20.8% invalid outputs, and a per-item hindsight oracle gets 8.4. PI gets 16.7 with 7.8% invalid, and its lead over the oracle excludes zero.
  2. The paper reports its failures, and its control theory predicts them. A tuned constant wins on LLaDA-MoE, the gain on Dream is small, and three targets fail on invalidity. This matches the paper's account that feedback helps only when per-example strength thresholds vary.
Critical weaknesses

Honestly this paper does feel like LLM written, but considering the idea is good and the authors seems to have added some human effort to it, Which I value. But it still has some elements here and there which I feel is something what LLMs leave out like if we read the reproducibility statement, it reads like an internal to-do list ("replacing absolute paths…")...

But holistically "AI Slop" problems are less.

At some places it feels like there is hodgepodge of writing, which affects presentations.

In terms of experiments, these are 2 of my concerns:

  1. The adversary controls forward hooks on the residual stream of a deployed model. Such an adversary can also edit logits or the output directly. The paper rules this out by definition ("direct edits to logits or responses ... lie outside this interface"). The paper does not argue why a realistic attacker would restrict itself to steering, e.g. because steering is harder to detect. Without that argument, "3x more than fixed-strength steering" is a result about closed-loop control during denoising, and its practical importance as a threat is unclear. The defense (zero setpoint) is tested only where the unsteered gap is already about 0.

  2. The PI gap is 39.5 when the target is option A and 5.3 averaged over B and C (Section 4.5), so most of the headline effect comes from one answer position. The Limitations section says that several per-target strengths were selected on evaluation data. The study is multiple-choice only, with a single-letter sensor. On SocialStigmaQA, calibration and evaluation share all 37 templates, and PI also moves answers on prompts without race descriptions, so that result shows transfer of answer control, not stigma-specific steering (the authors say this).

Questions for authors
  1. Why would an adversary with residual-stream hooks prefer steering over logit edits? Is there an audit or detection setting in which steering is stealthier? If not, how should readers read the threat model?
  2. Please report the Table 1 comparison (PI, matched open loop, hindsight oracle) separately for target positions A, B and C. Does PI beat the hindsight oracle at B and C?
  3. Which per-target strengths were tuned on evaluation items? What are the results when all hyperparameters are fixed on held-out items?
  4. The attack lowers abstention and raises invalid outputs. At what false-positive rate would an endpoint audit on a fixed probe set detect it?
  5. Contribution 3 says feedback wins when no single strength serves every example "and that a tuned constant does better". Under which condition does the constant do better?

Overall, I will increase my score happily If I get clarity over these questions.

4 comments

Response to Reviewer SsMd

We thank the reviewer for the careful reading and recognition of the central idea. Below, we clarify the existing evidence and report additional analyses addressing the concerns.

SsMd-Q1. Why steer activations rather than edit logits?

The two interventions act at different points. As described in Sections 2–3, activation steering modifies the model’s internal representations through a fixed direction shared across prompts, while logit editing directly changes output-token probabilities. Our direction is fitted from answer texts rather than option letters; the controller uses intermediate target probabilities to adjust its strength during denoising. This studies how the activation-steering interface also used for legitimate model control can induce targeted bias.

We compared both interventions on the same first 100 BBQ items × 3 rotations (300 prompts):

Condition (300 prompts)Target (%)Comparator (%)Abstain (%)Invalid (%)Gap (pp)
Base13.713.073.30.00.7
Decode-PI (ours)33.715.741.39.318.0
Logit edit, b=134.06.359.70.027.7
Logit edit, b=261.04.734.30.056.3
Logit edit, b=495.73.01.30.092.7

Direct logit editing achieves larger shifts in this comparison. Our contribution is feedback-driven manipulation through the residual stream during denoising, rather than outperforming direct output manipulation. The main Table 1 result remains 1.8 → 16.7 pp on 1,200 prompts.

The additional open-ended experiment shows that the activation intervention transfers beyond MCQ answers: with the same actuator and a sensor monitoring the target’s name rather than its answer letter, Decode-PI shifts the gap by +28 pp [8, 48] on free-form BBQ responses (Q2). Thus, the steering direction is reusable beyond letter-based outputs, although the sensor must identify the target.

SsMd-Q2. Does PI beat the hindsight oracle at target positions B and C?

The open-ended experiment directly addresses the concern that the effect depends on answer letters or position A. We evaluated 50 open-ended BBQ items with one-sentence responses, scored by whether they name the target, comparator, both or neither. The sensor reads the target’s name probability; the residual-stream intervention is unchanged.

Condition (50 items)Mean αTarget (%)Comparator (%)Degenerate (%)Gap (pp)Δgap vs. base [95% CI] (pp)
Base020360−16—
Decode-PI1.9936248+12+28 [8, 48]
Open loop, effort-matched α = 2.472.4734304+4+20 [−6, 46]
Open loop α = 33.002828140+16 [−8, 40]
Mean-AcT (s = 2)—323200+16 [−10, 42]
ITI-C (K = 48, a = 8)—30322−2+14 [−8, 36]
CAA / ActAdd / Linear-AcT—20–2232–340–6−12+4
AurA (inject / vanilla)—8–2424–400−160

Decode-PI shifts the gap by +28 pp [8, 48], the only tested method whose reported interval excludes zero. This demonstrates a target-directed shift without MCQ answer options, supporting the attack beyond the answer-position effects identified in Section 4.5.

The complementary per-letter analysis also gives PI higher point estimates than the hindsight oracle at A, B and C, with significant advantages at A and C. It measures

P(pick L∣target at L)−P(pick L∣comparator at L)P(\text{pick }L \mid \text{target at }L) - P(\text{pick }L \mid \text{comparator at }L)

Its mean across A, B and C equals the pooled gap; this is a per-letter contrast rather than a target-position-conditioned gap.

LetterDecode-PIHindsight static oraclePI − oracle [95% CI]
A17.04.0+13.0 [4.8, 21.0]
B12.09.8+2.2 [−3.0, 7.5]
C21.011.5+9.5 [3.0, 16.0]

The significant advantage at C (+9.5) provides further evidence beyond letter A; the B advantage is not statistically resolved. We will add both analyses to the revision.

SsMd-Q3. Were strengths tuned on evaluation items?

Yes, some were, as disclosed in Section 6 and Appendices A.5 and B.2–B.3. The race-target open-loop strengths and the Old-target controller bound were selected using evaluation items. For the primary Black-target experiment, controller settings were selected on 100 calibration items included in the 400-item headline pool; Appendix A.5 separately evaluates the remaining 300 disjoint items.

On those held-out items, Decode-PI reaches 16.2 pp: +14.1 pp over base [10.0, 18.4] and +12.8 pp over the effort-matched open loop. The calibration gap is 18.0 pp, placing the held-out result within 2 pp. Three fresh draws give +16.4, +14.7 and +17.0 pp over base. These results support the primary effect outside calibration, without resolving evaluation-set tuning for every secondary setting. We will identify the tuning split for each result explicitly.

SsMd-Q4. Can an endpoint audit detect the intervention, and at what FPR?

Yes. Section 6 identifies reduced abstention as an endpoint-visible signature. We now quantify an audit based on the drop in abstention on a fixed probe set:

Number of probesTPR at 1% FPRTPR at 5% FPR
100.210.42
250.730.85
500.970.99
1001.001.00

At 1% FPR, 10 probes miss the attack 79% of the time, while 50 probes achieve 0.97 TPR. These results concern probes targeting the evaluated demographic axis, not an axis-agnostic audit. Full output comparison against a clean reference flags 42% of prompts but requires access to that reference. We will report the probe construction and threshold-calibration protocol with these results.

SsMd-Q5. When can constant steering match feedback?

When a coherent constant approximates the intervention needed during denoising. Section 4.4 already reports that constant steering wins on LLaDA-MoE; Section 6 discusses this limitation. The additional trajectory analysis finds within-sequence command variance of 0.06 on Dream-7B and 0.19 on LLaDA-MoE, with 0% invalid on LLaDA-MoE. There, PI − CAA = −2.2 pp [−3.6, −0.9].

On LLaDA-8B, additional comparisons show a +12 to +13 pp advantage over every tested no-feedback schedule, including PI’s mean α(t) curve and replays of other items’ PI trajectories. These controls support the value of adapting to the current example’s predictions, beyond varying strength over time. We will revise Contribution 3 to state this empirical condition without claiming that strength variation universally guarantees a feedback advantage.

SsMd-W1. Defense was evaluated only where the base gap was near zero

Yes, the original defense evaluation had this limitation, as acknowledged in Sections 4.5 and 6. We additionally tested suppression on Latino, with a significant base gap: 400 items × 3 rotations (1,200 prompts), setpoint 0, command range [-6,0].

ConditionGap (pp), 95% CIInvalid (%)Mean αChange vs. base (pp), 95% CI
Base4.2 [1.1, 7.4]0.00—
PI suppression1.3 [−1.5, 4.1]0.3−0.59−2.9 [−5.1, −0.8]
Open loop, α = −0.59 (matched mean)3.7 [0.6, 6.7]0.0−0.59−0.6 [−1.8, 0.6]

PI significantly reduces the gap, leaving an interval that includes zero, with 0.3% invalid. The matched constant does not significantly reduce it; PI − matched open loop = −2.3 pp [−4.2, −0.5]. This extends the evidence for suppression to a setting with significant initial bias, without establishing that all residual bias is eliminated.

SsMd-W2. SocialStigmaQA shares templates between calibration and evaluation

Yes, all 37 templates are shared, as disclosed in Section 4.4 and Appendix C.8. The split does, however, hold out two other factors entirely. Calibration uses 200 yes-biased and 200 no-biased prompts drawn only from non-race stigmas in the original style. Evaluation covers 7 race stigmas × 37 templates in both the original and positive styles. The split manifest is included in the supplementary material.

Because the positive style never appears in fitting, it provides a held-out test. PI shifts the gap from −56.1 to +44.7 pp on the positive style, compared with −32.6 to +66.5 pp on the original style.

Condition (gap, pp, semantic parser)Original stylePositive style (not in fitting)
Clean−32.6−56.1
Decode-PI+66.5+44.7
Effort-matched open loop−2.6−2.4
Open loop, α = 4+20.1+18.8

The matched open loops use PI's direction and injection sites, so they isolate the effect of feedback-controlled strength. Pooled over both styles, PI exceeds the effort-matched open loop by +42.0 pp [36.2, 47.9] under strict parsing and +58.1 pp [48.8, 67.7] under semantic parsing.

Clustering intervals over the 37 templates accounts for dependence within templates, but it does not establish transfer to unseen templates. We will state this explicitly in the revision and keep the paper's interpretation: the result shows transfer of answer control across stigmas and prompt styles, not stigma-specific steering.

Dear Reviewer SsMd,

Thank you again for your thoughtful review and for recognizing the central contribution of feedback-driven steering. In our response, we addressed the following:

  1. Added a direct logit-editing comparison and demonstrated significant gains on open-ended generation.
  2. Provided per-letter comparisons, held-out evaluations, and endpoint detection results to address position bias, tuning, and detectability.
  3. Demonstrated bias suppression on a target with significant initial bias and clarified the SocialStigmaQA evaluation, including transfer to a held-out prompt style.

As the discussion period ends soon, we would be grateful to hear whether these additional results address your concerns. We are happy to answer any further questions.

Reviewer_bTb4OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper considers a threat model where an adversary can supply a steering vector into dLLM denoising, and has the goal of manipulating the demographic distribution returned by the dLLM. The paper proposes a pretty straightforward method for doing this manipulation, and shows that it works on some dLLMs and two relevant benchmarks.

Critical strengths

I have not seen any papers that investigate this particular threat before. There is some novelty here.

Given the threat model, the proposed approach outperforms relevant baselines.

Critical weaknesses

To be honest, the paper's biggest flaw is just that the threat model is very weak - I think it is very unlikely that an adversary is actually going to have this particular access. In cases where an adversary is allowed to supply a malicious steering vector (maybe steering vectors are part of the API?), the paper doesn't really consider the "backdoor" part of this, where the adversary would ideally have the steering vector achieve another good purpose rather than being specifically built for the purpose of bias. On top of the steering vector bit, there's the feedback part of this that makes the threat model even stronger.

I'm also pretty skeptical of the MCQ evaluation used in the paper. I'm nervous that in more harmful settings that might involve open ended generation, the approach would be too brittle to work well.

Questions for authors

I'd like the authors to defend the threat model! Have there been any APIs/CVEs that permit an adversary to achieve this threat model (and not something much stronger like direct control over sampled tokens)?

3 comments

Response to Reviewer bTb4

We thank the reviewer for recognizing the novelty of this threat in diffusion language models and the improvement over relevant baselines. We address the concerns below.

R1.1. Does the attack generalize beyond multiple-choice questions?

Yes, our additional open-ended experiment supports this. The submitted paper evaluates multiple-choice questions (Section 4.1) and identifies free-form generation as a limitation (Section 6). We selected BBQ (Parrish et al., 2022) and SocialStigmaQA (Nagireddy et al., 2024) following their use of question answering to assess stereotyped responses under insufficient context and stigma amplification in social situations, respectively. Their questions and annotations provide a controlled setting for measuring biased answer shifts. We agree that multiple-choice results alone do not establish open-ended effectiveness.

Following the reviewer’s suggestion, we evaluated open-ended BBQ, retaining the same contexts and questions but requesting a one-sentence free-form response. Responses are classified by whether they name the target, comparator, both, or neither. The gap is the target rate minus the comparator rate, in percentage points.

Condition (50 items)Mean αTarget (%)Comparator (%)Degenerate (%)Gap (pp)Δgap vs. base (pp), 95% CI
Base020360−16—
Decode-PI1.9936248+12+28 [8, 48]
Open loop, effort-matched (α = 2.47)2.4734304+4+20 [−6, 46]
Open loop (α = 3)3.002828140+16 [−8, 40]
Mean-AcT (s = 2)—323200+16 [−10, 42]
ITI-C (K = 48, a = 8)—30322−2+14 [−8, 36]
CAA / ActAdd / Linear-AcT—20–2232–340–6−12+4
AurA (inject / vanilla)—8–2424–400−160

Decode-PI shifts the gap by 28 percentage points (95% CI [8, 48]), the only tested method whose reported confidence interval excludes zero. The sensor monitors the probability of the target’s name, rather than an answer letter; the residual-stream intervention remains unchanged. These results support free-form transfer on this 50-item evaluation. We will include them and their scope in the revision.

R1.2. How realistic is the threat model? Are there relevant APIs or CVEs?

Section 2 considers an adversary inside the serving stack, not an arbitrary caller of a black-box API. The assumed interface permits residual-stream hooks and intermediate probability readout, while excluding weight updates and direct edits to logits, the sampler, or sampled tokens. A compromised serving component or an untrusted steering module motivates this setting; the restricted access boundary is an explicit modeling assumption.

As discussed in Sections 2 and 5, Trojan Activation Attack (Wang & Shu, CIKM 2024) provides the closest precedent: inference-time activation-vector injection into open-source LLMs, including bias manipulation on BOLD. Arditi et al. (NeurIPS 2024) demonstrate refusal removal through activation interventions without weight updates; Soft Prompt Threats (Schwinn et al., NeurIPS 2024) studies continuous-space attacks with model access; and Shnaidman et al. (ReALM-GEN at ICLR 2026) applies activation steering across LLaDA-8B-Instruct’s reverse-diffusion steps. These works motivate studying activation access, rather than establish our exact deployment boundary.

We do not claim a public API exploit or CVE. As the reviewer recognizes, studying this threat in diffusion language models is novel. Our work investigates this attack surface, and we anticipate future work examining more realistic deployment scenarios and access constraints.

The feedback readout is already included in Section 2’s assumed interface: it reads probabilities computed during denoising without changing token sampling. It is not an additional capability introduced by the controller, nor do we claim that all activation-hook interfaces automatically expose logits. Our question is what an attacker can induce through residual-stream steering alone. We will clarify this distinction from unrestricted control of the serving process.

R1.3. Why is this a backdoor rather than a malicious steering vector?

Our contribution is an inference-time, closed-loop steering attack. Section 3 specifies this mechanism; the backdoor literature motivates the access model, not a claim that our experiments establish a benign-purpose, secretly triggered backdoor.

Conventional backdoor attacks often implant trigger-dependent behavior during training or fine-tuning while preserving normal behavior on clean inputs. Our attack leaves the weights unchanged. Its controller adapts to the measured target probability and accumulated error (Section 3, Equation (4)); this state dependence does not itself establish a secret trigger or an independently benign function.

A future extension could embed the attack in a module that performs an authorized task while covertly shifting demographic outcomes in selected contexts. We will distinguish this scenario from our demonstrated attack and identify benign-task preservation and trigger-specific activation as future evaluations.

We will add the open-ended results and clarify both the attacker-access assumptions and the distinction from a conventional backdoor.

Dear Reviewer bTb4,

Thank you again for your careful review and for recognizing the novelty of our threat model and improvements over existing baselines. In our response, we addressed the following:

  1. Demonstrated that the attack extends beyond multiple-choice questions, achieving a significant +28 pp shift on open-ended BBQ.
  2. Clarified the activation-access threat model, its precedents, and the distinction from unrestricted control of logits or sampled tokens.
  3. Distinguished our inference-time steering attack from conventional backdoors and clarified the scope of our contribution.

As the discussion period ends soon, we would be grateful to hear whether these results and clarifications address your concerns. We are happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.