acceptodds
Under review as a conference paper at ICLR 2027

Beyond Model Weights: Sampling-Dependent Fairness in Diffusion Language Models

Abstract

Diffusion language models (DLMs) are emerging as a practical alternative to autoregressive language models (ARLMs), yet their fairness remains largely uncharacterized. DLMs introduce sampling controls with no direct autoregressive counterpart, such as block length and remasking strategy, that directly set their speed–quality operating point. This raises a basic question: does a DLM's measured social bias remain stable across these settings? Across LLaDA, LLaDA-1.5, and Dream on Open-BBQ, we find that all three models exhibit social bias, but both the magnitude of measured bias and the relative ordering of models by bias are highly sensitive to the sampling configuration. This sampling dependence also changes what successful debiasing means. While mitigation principles established for ARLMs transfer effectively to DLMs and reduce bias without degrading downstream performance, sampling dependence persists even after mitigation. Our results show that fairness in DLMs cannot be characterized independently of the sampling configuration, even after debiasing.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
diffusion language models, fairness, social bias, evaluation, debiasing

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_W54BOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how different sampling settings affect social bias in diffusion language models. Changing block length, remasking strategy, and the number of tokens generated at each step changes both the bias scores and which model appears less biased. The paper also tests whether ARLM debiasing methods work on DLMs. These methods reduce bias, but the results still vary across sampling settings, suggesting that DLM-specific generation settings should also be considered when evaluating and mitigating bias.

Critical strengths
  • The paper systematically compares different sampling configurations across three DLMs. Showing that these settings change both bias scores and model rankings is useful for fairness evaluation across different models.

  • I like that the paper includes mitigation as well as evaluation. The transferability of ARLM debiasing methods to DLMs is an interesting finding.

  • The mitigation experiments show that reducing bias and reducing variation across sampling settings are not necessarily the same. Some fairness terms help with both, which is useful for understanding what successful debiasing should achieve.

Critical weaknesses
  • Finding social bias in DLMs is not particularly surprising, given what is already known about bias in language models. The more interesting finding is how sampling changes that bias, why this happens, and how it differs from decoding effects in ARLMs.

  • I understand why the authors chose the short-answer task. However, the prompt still provides answer options, and evaluation uses the selected option. It is unclear how much this differs from multiple-choice evaluation, so the “open-ended” claim should be toned down.

  • Some mitigation methods substantially reduce variation across sampling settings. The paper should explain whether the remaining differences are large enough to matter, rather than focusing only on the fact that some dependence remains.

Questions for authors
  • Since the training data and model weights are fixed within each sweep, why do different remasking strategies produce different bias scores? Do you think this comes from generation order, parallel token prediction, or the amount of refinement?

  • ARLM bias also changes with decoding settings. What is different or specific to DLMs here?

  • How much of the bias difference across the three models might come from differences in their training data itself, rather than their diffusion generation process?

  • The prompt provides answer options. How is this different from multiple-choice evaluation? Did you measure bias in the reasoning as well as the final answer?

  • Do the findings extend to free-form generation without predefined answer options?

  • Based on your results, which sampling settings would you recommend to reduce bias, and what would be the cost in speed or answer quality?

4 comments
AuthorsOpenReview ↗

Response to Reviewer W54B (1/2)

We thank you for appreciating that the paper covers mitigation as well as evaluation, and that reducing bias and reducing variation across sampling configurations are not necessarily the same. We respond to each comment below. To address each point fully, we reply in two parts. In the revised manuscript, the changes that directly relate to our responses are shown in red.

W3.1, Q3.2

We agree that how sampling changes bias matters more than its existence. Still, research on DLMs has focused on generation quality and sampling speed, and their social biases had not been evaluated systematically in published work (§1).

Why sampling changes bias. Two additional analyses address this, and we have added both to the revision.

  1. Under random remasking, once the reasoning names one of the two subjects, the model answers "none" less often, which raises the score (see also our response to Reviewer YtFv, Q1.1).
  2. The variation of bias scores remains significant when restricted to high-validity configurations and to shared valid items, so it is not a by-product of failed extraction (see our response to Reviewer RTri, W2.1).

Difference from ARLMs. The difference appears under greedy decoding: an ARLM then produces a single output and a single bias score, whereas LLaDA and LLaDA-1.5, which we decode greedily, still change their bias score across sampling configurations (Fig. 2).

Remasking strategy, block length, and TPS do not exist in ARLMs (§2.1), so their effect cannot be evaluated directly in ARLMs. Conversely, we have started an additional evaluation that applies temperature, top-p, and top-k to LLaDA. With these results, we will add to the final version an analysis of how the controls shared with ARLMs and the DLM-specific controls each affect social bias.

W3.2, Q3.4

We would like to clarify what "open-ended" refers to.

What "Open-ended" refers to in Open-BBQ. In Open-BBQ, "open-ended" describes the response format, not the answer space. Every item involves two subjects, and all three formats are scored over the same three options: the two subjects and "unknown" (App. A.1). In the original short-answer format, the model writes a free-form answer, and a GPT-4o judge is given the three options and maps the answer to one of them. The final answer is therefore selected from a closed set in the original short-answer format as well.

Our setting. As stated in §3.1, we use the short-answer format with a boxed final answer for three reasons.

  1. The original multiple-choice format does not have the model write its chain of thought, so the effect of the sampling configuration on the reasoning cannot be observed.
  2. The original short-answer format relies on an LLM judge. In our pilot study, the guardrails of the judge model added noise to the fairness measurement of the evaluated model.
  3. The boxed answer remains readable at a large TPS. At TPS = 8, the box failure rate of LLaDA-1.5 stays near 10% (Fig. 2), whereas in our pilot study the judge model failed to extract an answer from nearly 40% of the samples at the same TPS (App. A.1).

That said, showing the options and requiring a boxed answer may themselves affect how social bias appears in the model's output. We adopted the short-answer format with a boxed final answer nevertheless, because the three reasons above make it necessary for a valid experiment.

Bias in the reasoning. We score bias on the final answer only, on which the BBQ bias score [1] is defined. We did not score the reasoning itself: in our pilot study, many reasoning traces could be read as supporting any of the three options, so they could not be labelled reliably (App. A.1). We did, however, analyse how the reasoning relates to the final answer: under random remasking, once the reasoning names one of the two subjects, the model answers "none" less often (see our response to Reviewer YtFv, Q1.1).

W3.3

Thank you for raising this point. We computed 95% bootstrap intervals for the variation that remains in LLaDA tuned with a lexicon-based fairness term, and confirmed that part of it is not statistically significant.

What we intended to convey in §4.2 is that (1) “lowering the bias and reducing its sensitivity to the sampling configuration are distinct objectives in DLMs”, and (2) “improvements in sampling robustness are not consistent across models and training objectives”. We have revised the text so that this claim is clear.


References

[1] Parrish et al. “BBQ: A Hand-Built Bias Benchmark for Question Answering.” Findings of ACL 2022.

AuthorsOpenReview ↗

Response to Reviewer W54B (2/2)

This is the second part of our response, covering your questions.

Q3.1

Our results point to generation order and parallel token prediction (§2.1), rather than to the amount of refinement. We have added this analysis as §5.4.

Amount of refinement. LLaDA, LLaDA-1.5, and Dream-7B never remask a token once it is committed. The variation we observe therefore does not come from the amount of refinement.

Generation order and parallel token prediction. The trend differs between models.

On LLaDA and LLaDA-1.5, the bias score hardly changes with the remasking strategy in two settings: at TPS = 1, where each step commits a single token, and at B = 8 and TPS = 8, where each step commits a whole block and the strategy has no positions to choose from. In the other settings, the two strategies differ by up to .056 (Fig. 2). On these models, the bias score thus changes mainly when generation order and parallel token prediction interact.

On Dream-7B, in contrast, the gap is already present at TPS = 1 and is significant on high-validity configurations (see our response to Reviewer RTri, W2.1), so the order alone is sufficient there.

How the order changes the answer. Under random remasking, once the reasoning names one of the two subjects, the model follows it into the answer more often and answers "none" less often (see our response to Reviewer YtFv, Q1.1).

Q3.3

Differences in training data may well contribute to the differences in social bias among the three models. However, attributing these differences to their sources is outside the scope of this study. As stated in §1, our aim is to examine whether the sampling controls specific to DLMs affect social bias.

Q3.5

Yes. We repeated the sweep on two benchmarks without predefined answer options: CEB [1] (single-turn free-form generation) and FairMT-1K [2] (multi-turn free-form dialogue); see our response to Reviewer YtFv, W1.3. On both benchmarks, the bias score still depends on the sampling configuration (Fig. 6 in §5.3; Fig. 15 in App. G). The model ranking also changes, most clearly on FairMT-1K (Fig. 6, third row; Fig. 16). Our claim that "fairness in DLMs must be assessed across sampling configurations" therefore also holds for free-form generation.

Q3.6

As stated in §1, our aim is not to identify a single best sampling setting, but to show how much bias changes across settings. We then claim in the conclusion (§7) that "fairness in DLMs must be assessed across sampling configurations, both before and after mitigation".

On answer quality, Fig. 2 reports the box failure rate for every setting, and §5 evaluates the utility of the debiased models on GSM8K, IFEval, and MMLU. Sampling speed is outside the scope of this study.


References

[1] Wang et al. “CEB: Compositional Evaluation Benchmark for Fairness in Large Language Models.” ICLR 2025.

[2] Fan et al. “FairMT-Bench: Benchmarking Fairness for Multi-turn Dialogue in Conversational LLMs.” ICLR 2025.

Thank you for the detailed response. The new free-form experiments address my concern about whether the results depend on predefined answer options. The remasking analysis also helps clarify when token order and parallel prediction affect bias scores, even under greedy decoding.

My main concern is how much these findings add to prior work showing that measured bias depends on models and decoding settings. Is there any statistically significant consistent pattern found? I still find it unclear what additional findings or insights it adds to the existing literature. I am keeping my original rating.

AuthorsOpenReview ↗

Response to Reviewer W54B's follow-up comment

Thank you for reading our response and for acknowledging the additional analyses. We have accordingly revised the paper and indicate the changes in red text. In particular, we reorganized the related work (§6) and added §6.1, which compares our work with studies of decoding and bias in ARLMs.

To your question: yes, we find three consistent patterns. For each pattern, we state what we show and how far prior work had gone.


Pattern 1: The sampling controls specific to DLMs significantly change the bias score of a single model. We systematically evaluated the effect of the sampling controls specific to DLMs, taking confounding factors into account. This evaluation showed that, with the model weights and the decoding controls of ARLMs (temperature, top-p, and top-k) held fixed, the bias score of every model varies significantly across sampling configurations (§3.2; Fig. 2). We additionally verified that this is not specific to Open-BBQ and also holds on CEB and FairMT-1K (§5.3; Fig. 6). We also confirmed that it holds when restricted to high-validity configurations (Fig. 13a in App. E). Furthermore, we analyzed why the bias varies with the sampling configuration (§5.4).

Prior work had shown only that the bias of ARLMs varies with temperature, top-p, and top-k, which modulate sampling stochasticity. The sampling controls of DLMs instead change how generation itself unfolds and are chosen to set the speed–quality operating point, yet their effect on social bias had been examined only on a single task, primarily in the context of adversarial perturbations (§6.1, §6.3).

Pattern 2: The sampling configuration significantly changes the ranking of models by bias. On the practical question of which model is the most biased, we showed that the answer changes across sampling configurations on all three benchmarks (§3.2, §5.3; Fig. 13b and 13c in App. E). A fairness conclusion drawn from a single sampling configuration can therefore be misleading.

Prior work on DLMs had not examined whether a comparison between models is stable across sampling configurations.

Pattern 3: Debiasing lowers the bias but does not consistently reduce its sensitivity to the sampling configuration. We confirmed that mitigation strategies established for ARLMs transfer to DLMs and lower the bias, and observed that the sampling dependence can nevertheless persist after mitigation (§4.2; Fig. 3). Lowering the bias and reducing its sensitivity to the sampling configuration are thus distinct objectives. This changes what successful debiasing means: effective debiasing should maintain low bias across all sampling configurations.

Prior work had established these strategies for ARLMs only, where they are assessed by how much they lower the bias. To our knowledge, none had been applied to DLMs.


Together, these patterns lead to our message, which prior work on DLMs had not established, that fairness in DLMs must be assessed across sampling configurations, both before and after mitigation (§7). We appreciate that your review recognized these points among the strengths of the paper, and we hope this response conveys the value of our work.

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

This paper studies how sampling configurations affect measured social bias in diffusion language models. It evaluates LLaDA, LLaDA-1.5, and Dream on a modified OpenBBQ protocol, varying block length, remasking strategy, and tokens committed per denoising step. Both bias scores and the relative ranking of models change across configurations. The paper then evaluates supervised fine-tuning, diffusion on-policy self-distillation, and several fairness regularizers. These methods can substantially reduce measured bias, but sensitivity to sampling remains. Additional experiments examine downstream utility, category-level differences, and the construction of a lexicon used for probability equalization.

Critical strengths
  • The paper addresses a useful evaluation question. Comparing the same checkpoint across sampling configurations makes clear why an isolated fairness score may give an incomplete picture of a deployed model. Examining model rankings, rather than only within-model score changes, makes the practical consequence especially clear.

  • The experimental coverage is substantial: three models, a systematic sampling grid, two post-training objectives, and several regularizers. Reporting format failures alongside bias scores is helpful. The context-group separation between training and evaluation is also a sensible precaution, and the utility experiments provide useful evidence about the costs of mitigation.

Critical weaknesses
  • The main sampling comparison is confounded by configuration-dependent answer extraction. Bias and accuracy are computed only on outputs with a readable boxed answer, so different configurations can be scored on different subsets of the benchmark. This becomes particularly problematic for Dream, whose failure rate reaches 0.996 in one setting. Showing the failure rate makes this limitation visible, but does not establish how much of the reported bias variation or model reordering survives on comparable samples. The main ranking figures also do not quantify uncertainty in the rank changes. I would find the central conclusion more convincing with comparisons restricted to high-validity configurations, paired analyses on shared valid items, and uncertainty estimates for the relevant score differences. Some sampling dependence may well remain; its magnitude and effect on rankings are what need stronger support.

  • The interpretation of the mitigation results is broader than the validation supports. The lexicon regularizes groups formed partly by frequency and part of speech, and Appendix A.3 explicitly acknowledges groups such as “old, young, healthy” and “gay, boyfriend, girlfriend, plumber, ace, queer.” These are not straightforward demographic alternatives. Lower OpenBBQ bias therefore does not by itself establish that this regularization improves demographic fairness more generally. The ablation usefully shows improved ambiguous-context accuracy, but evaluation remains within the same benchmark, and probability equalization is measured on the training lexicon itself. The utility claim also needs qualification: the abstract states that mitigation does not degrade downstream performance, whereas Section 5.1 reports GSM8K degradation for dOPSD, and Appendix C.2 reports additional IFEval degradation on Dream.

Questions for authors
  • Which model-ranking reversals remain when comparisons use high-validity configurations and shared valid items? Can you provide paired, template-clustered confidence intervals for these differences and a sensitivity analysis for unreadable outputs?

  • Does the lexicon-based mitigation remain effective with semantically audited demographic groups or on a separate bias benchmark? This would help distinguish broader debiasing from adaptation to the particular benchmark and vocabulary.

  • Can you clarify the exact Dream sampling implementation? Appendix A.1 says generation reuses the released samplers, while Appendix C.1 states that the grid imposes a block structure absent from Dream’s published sampler. Please distinguish native and modified settings and specify the stochastic sampling settings and repetitions used for the ranking comparisons.

3 comments
AuthorsOpenReview ↗

Response to Reviewer RTri

We thank you for recognizing both the practical value of comparing model rankings across sampling configurations and the breadth of our experimental coverage. We respond to each comment below. In the revised manuscript, the changes that directly relate to our responses are shown in red.

W2.1, Q2.1

Thank you for raising this point.

First, the definition of acc in §3.1 did not match our implementation. The official implementation counts an unreadable output, one that does not follow the answer format, as none. We count it as incorrect, so it does not enter the numerator of acc. We have corrected the definition (Eq. 4 in the revision). No value, figure, or claim changes.

We then ran the checks you suggested on high-validity configurations and on shared valid items (Fig. 13 in App. E). All intervals are 95% paired bootstrap intervals that resample context groups (the question pair of one context).

High-validity configurations. We define a configuration as high-validity when its box failure rate on the ambiguous samples is at most 15%. Restricted to these configurations, the score of each model still varies significantly (a). The sampling dependence is therefore not a by-product of failed extraction.

At the five configurations of twenty-four where all three models are high-validity, the ranking still changes (b). Dream is significantly less biased than LLaDA and LLaDA-1.5 under low-confidence remasking and significantly more biased under random remasking.

Shared valid items. On the samples that all three models answered in the required format (c), the ranking still changes. The reversal is therefore not produced by scoring different subsets.

Sensitivity to unreadable outputs. Fig. 14 redraws the Dream column of Fig. 2 with unreadable outputs scored in five ways. Where the failure rate is low, the choice does not matter. Where it is high, the rank follows the choice, and each alternative to ours has a drawback there.

  • As none (the Open-BBQ official implementation). A failure counts as the correct answer in an ambiguous context, so a model looks fairer the more often it fails.
  • Dropped. Each configuration is scored on a different subset, which is the concern you raise.
  • All stereotyped or all anti-stereotyped. These assign an answer the model did not give, and the ranking then follows the failure rate rather than the answers.

Counting a failure as incorrect, separate from none, avoids all three.

W2.2, Q2.2

Manual filtering of the lexicon groups As noted in App. A.3, the groups are formed by part of speech and corpus frequency rather than by meaning, so some groups hold a term that is not, strictly speaking, an identity term. However, we consider the harm of such terms to be small relative to the benefit of the fairness term, for two reasons:

  1. Filtering the mined pairs by hand changed the ambiguous bias score little (Tab. 9 in App. B.1).
  2. The fairness term is applied to ambiguous samples only (§4.1), and tuning barely moves the disambiguated bias score (Tab. 12 in App. C.4).

Since the lexicon depends on the tokenizer, hand filtering must be redone for each model (App. B.1). It may still improve the result, and we leave it to future work.

Evaluation on other benchmarks We agree. We have evaluated the base models on CEB and FairMT-1K (Fig. 6 in §5.3; see our response to Reviewer YtFv, W1.3). The evaluation of the debiased models is in progress, and we will revise our claims according to the results.

Equalization measured on the training lexicon This is by design. The measurement checks that the fairness term changes the distribution it targets; we do not use it as evidence of generalization. We have stated this explicitly in App. B.1.

Utility cost Our intended claim is that a DLM can be debiased at no utility cost with a mitigation strategy established for ARLMs. The degradations you point out do exist, but each model has at least one strategy with no significant loss on any benchmark (Fig. 5 and Fig. 12), so this claim holds. We have revised the related text (Abstract and §5.1) to make this intent clearer.

Q2.3

Dream sampler The official Dream sampler has no block length: the block is fixed to the full generation window. To study the effect of block length, we modified the official sampler only so that a block length can be specified. All Dream configurations in the sweep use this sampler. We have revised App. A.1 to make this explicit, and removed the sentence in App. C.1 that conflicted with it.

Dream’s stochastic sampling As noted in App. A.1, Dream returns an empty answer for most prompts at temperature 0, so we adopted the stochastic settings of the Dream evaluation script released with dOPSD (temperature 0.1, top-p 0.9). LLaDA and LLaDA-1.5 are decoded greedily and were therefore evaluated once per configuration; for consistency, each Dream configuration was also evaluated once on the 10,000 test samples (§3.1).

Reviewer_RTriOpenReview ↗

Beyond Model Weights: Sampling-Dependent Fairness in Diffusion Language Models

Thanks for adding the high-validity and shared-item analyses. Seeing the ranking reversal persist in both checks addresses my main concern about answer extraction. The Dream sampler clarification is also helpful. My remaining concern is with the mitigation claim: the cross-benchmark results are for the base models, so they don’t yet show whether the debiasing gains transfer beyond OpenBBQ. I’m keeping my original rating.

AuthorsOpenReview ↗

Response to Reviewer RTri's follow-up comment

Thank you for the continued discussion. Following your comment, we evaluated the debiased model on CEB [1] and FairMT-1K [2].

The debiasing effect generalizes to free-form generation. Bias is significantly reduced on 2 of the 3 tasks, with no significant degradation on the third. We evaluated the debiased LLaDA-8B (dOPSD+log-prob. var.) under the default sampling configuration.

Benchmark (lower is better)BaseDebiasedΔ [95% CI]
FairMT-1K, bias rate (%)44.038.0−6.1 [−9.1, −3.0]
CEB Conversation-S, bias score17.012.7−4.4 [−6.8, −2.1]
CEB Continuation-S, bias score29.130.3+1.2 [−0.7, +2.9]

Confidence intervals are from a paired bootstrap over prompts (5,000 resamples).


References

[1] Wang et al. “CEB: Compositional Evaluation Benchmark for Fairness in Large Language Models.” ICLR 2025.

[2] Fan et al. “FairMT-Bench: Benchmarking Fairness for Multi-turn Dialogue in Conversational LLMs.” ICLR 2025.

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

This paper examines social bias in diffusion language models and how sampling settings affect fairness. It evaluates three models, LLaDA, LLaDA 1.5, and Dream, on Open BBQ. Bias scores and fairness rankings vary a lot across sampling settings. Several debiasing methods adapted from autoregressive models are also tested. These methods reduce bias, but the variation across sampling settings remains. Sampling settings should therefore be considered when evaluating and reducing bias in DLMs.

Critical strengths
  1. The paper addresses an underexplored fairness issue in DLMs and highlights the importance of considering sampling configurations when evaluating model fairness.

  2. The authors conduct a systematic evaluation of 24 sampling configurations across three representative DLMs on OpenBBQ. The results reveal noticeable variations in bias scores and model fairness rankings, providing useful empirical observations of sampling-dependent fairness.

  3. The paper explores the applicability of existing ARLM debiasing methods to DLMs and examines the relationship between bias reduction and sampling robustness. The observation that sampling sensitivity persists after debiasing is interesting and provides a useful direction for future research.

Critical weaknesses
  1. The current fairness metric conflates stereotype preference with the model's ability to correctly abstain, making it unclear whether lower bias scores reflect reduced stereotype preference or simply more frequent selection of the "None" option.

  2. While the paper demonstrates that fairness scores vary across sampling configurations, it does not sufficiently quantify the relative importance of sampling-induced variation compared to the model's parameter bias. This raises questions about the practical significance of the observed sampling sensitivity.

  3. The fairness evaluation relies exclusively on OpenBBQ, making it unclear whether the observed sampling-dependent bias generalizes to other fairness benchmarks (FairMT [1]) and evaluation settings.

[1] Fan, Zhiting, et al. "Fairmt-bench: Benchmarking fairness for multi-turn dialogue in conversational llms." International Conference on Learning Representations. Vol. 2025. 2025.

Questions for authors

The paper shows that remasking strategies have a substantial impact on bias scores. Can the authors provide further analysis of why different remasking strategies lead to different fairness outcomes, particularly how the token commitment order influences biased predictions?

1 comment
AuthorsOpenReview ↗

Response to Reviewer YtFv

We thank you for recognizing sampling-dependent fairness in DLMs as an underexplored issue, and for the interest in our finding that sampling sensitivity persists after debiasing. We respond to each comment below. In the revised manuscript, the changes that directly relate to our responses are shown in red.

W1.1

In ambiguous contexts, "none" is both the fair answer and the correct one. BBQ therefore scales the ambiguous bias score by (1 − acc), so that a biased answer counts for more the more often it is given [1]. We follow this definition (§3.1).

In disambiguated contexts, in contrast, the correct answer is the subject identified by the context. Debiasing lowers the ambiguous bias score of LLaDA from .123 to between .000 and .064, while its disambiguated accuracy is maintained: .942 before tuning and between .934 and .991 after (see also Tab. 9 in App. B.1). The tuned models thus answer "none" where the context is insufficient and name the subject where it is sufficient.

To make this explicit, we have added the accuracy of every model to Tab. 17 in App. H and stated the argument above in the revision.

W1.2

We understand "the model's parameter bias" as the bias that a model acquires from its training data.

Our scope. The model's parameter bias may exist. However, as stated in §1, our aim is to examine how the sampling configuration changes the bias score of a single model, with its weights held fixed (see also our response to Reviewer W54B, Q3.3).

Practical significance. We do examine one question of direct practical relevance: which model is the most or the least biased. The answer depends on the sampling configuration (§3.2; Fig. 2, last row). This also holds on high-validity configurations: Dream is significantly less biased than LLaDA and LLaDA-1.5 under low-confidence remasking and significantly more biased under random remasking (see our response to Reviewer RTri, W2.1). The variation caused by sampling is therefore large enough to change the conclusion of a model comparison.

W1.3

Open-BBQ is built on BBQ [1], which remains one of the most widely used fairness benchmarks [2, 3, 4]. Other common benchmarks, CrowS-Pairs [5] and StereoSet [6], require likelihood-based candidate scoring methods that cannot be directly applied to DLMs. We therefore added two free-form fairness benchmarks: FairMT-1K [7], which you suggested, and CEB [8] (Fig. 6 in §5.3; Figs. 15 and 16 in App. G).

Bias score. We evaluated LLaDA-8B at the same 24 sampling configurations as in Fig. 2. The bias score varies on both benchmarks. On FairMT-1K, 14 of the 23 non-default configurations differ significantly from the default one (95% paired bootstrap). At large TPS, part of the outputs degenerate into repetition; the variation remains when these outputs are excluded.

Model ranking. Because of limited computation and time, we evaluated the three models at B = 128 only (eight configurations). The model ranking also changes, most clearly on FairMT-1K (Fig. 6, third row; Fig. 16): Dream-7B is significantly more biased than LLaDA-8B under low-confidence remasking at TPS = 1, and significantly less biased than both LLaDA models under random remasking. The reversal also holds on the dialogues for which all three models give a readable response.

The sampling dependence of the bias score is therefore not specific to Open-BBQ. We have added these results to the revision.

Q1.1

Thank you for the insightful questions.

Our additional analysis shows that, under random remasking, once the reasoning names one of the two subjects, the model follows it into the answer more often and answers "none" less often:

Share of "none" answersrandomlow-confidence
reasoning names a subject42%62%
reasoning names no subject69%70%

This explains why the base model tends to score higher under random remasking than under low-confidence remasking (Fig. 2). For this analysis, we regenerated the 5,000 ambiguous samples with LLaDA-8B (B=128, TPS=2) while recording the step at which every token is committed (§5.4 and App. F).


References

[1] Parrish et al. “BBQ: A Hand-Built Bias Benchmark for Question Answering.” Findings of ACL 2022.

[2] Xu et al. “BiasFreeBench: a Benchmark for Mitigating Bias in Large Language Model Responses.” ICLR 2026.

[3] Zhang et al. “Evaluating the Effect of Retrieval Augmentation on Social Biases.” EACL 2026.

[4] Feng et al. “C2PO: Diagnosing and Disentangling Bias Shortcuts in LLMs.” Findings of ACL 2026.

[5] Nangia et al. “CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models.” EMNLP 2020.

[6] Nadeem et al. “StereoSet: Measuring Stereotypical Bias in Pretrained Language Models.” ACL 2021.

[7] Fan et al. “FairMT-Bench: Benchmarking Fairness for Multi-turn Dialogue in Conversational LLMs.” ICLR 2025.

[8] Wang et al. “CEB: Compositional Evaluation Benchmark for Fairness in Large Language Models.” ICLR 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.