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

Koshiro Saito; Daisuke Oba; Danushka Bollegala; Naoaki Okazaki

cs.CLOct 20, 2026

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

Authors: Koshiro Saito, Daisuke Oba, Danushka Bollegala, Naoaki Okazaki

Organizations: Institute of Science Tokyo · University of Liverpool · Amazon · AIST ·

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 strategies 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. Project page is available at https://pakapaka333.github.io/Beyond-Model-Weights/.

Figures & tables

Figure 1: One rollout of a masked diffusion language model. A row is the generation window after a denoising step, a cell is one position and [M] is the mask symbol. Red outlines mark the positions a step commits, and the pale cells are those the current block puts out of reach.

Figure 2: Bias varies with the sampling configuration. Disambiguated bias score (first row), ambiguous bias score (second row), box failure rate (third row), and rank of the model by ambiguous bias score (fourth row) for the three models across 24 sampling configurations.

Figure 3: Effect of debiasing with SFT and dOPSD. Disambiguated bias score (upper), ambiguous bias score (middle), and box failure rate (lower) for the base, SFT, and dOPSD LLaDA models, across 24 sampling configurations. The bottom panels give the rank of LLaDA among the three models by the ambiguous bias score of their SFT and dOPSD models.

Figure 4: Effect of the three fairness terms on LLaDA. Each bar is the range of the ambiguous bias score over the 24 sampling configurations, drawn per remasking strategy; the marker is the default configuration, and the dashed rule is the base model at the default configuration.

Figure 5: Effect of debiasing tuning on utility. Ambiguous bias score on Open-BBQ against accuracy on GSM8K, IFEval, and MMLU for LLaDA tuned on Open-BBQ.

\dagger

\ddagger

: significant improvement/degradation in accuracy over the base model under a paired cluster bootstrap (

5

k resamples,

95\%

interval excluding zero; clusters are prompts for IFEval and questions otherwise). Every tuned model lowers the bias score significantly (tested on

|\mathrm{bias}|

with BBQ templates as clusters).

			SFT				dOPSD
Category	$n$	Base	—	log-ratio	JS div.	var.	—	log-ratio	JS div.	var.
Age	314	$-$ .465	$-$ .118	$-$ .145	$-$ .003	$-$ .016	$-$ .301	$-$ .232	$-$ .062	$-$ .182
Physical appearance	134	$-$ .366	$-$ .030	$-$ .030	$-$ .000	$-$ .008	$-$ .261	$-$ .358	$-$ .082	$-$ .172
SES	586	$-$ .361	$-$ .012	$-$ .004	$-$ .000	$-$ .000	$-$ .165	$-$ .156	$-$ .010	$-$ .056
Disability status	134	$-$ .261	$-$ .082	$-$ .022	$-$ .000	$-$ .000	$-$ .187	$-$ .190	$-$ .046	$-$ .093
Religion	102	$-$ .157	$-$ .108	$-$ .049	$-$ .010	$-$ .000	$-$ .147	$-$ .098	$-$ .020	$-$ .059

Table 1: Bias and the effect of debiasing by social category. Ambiguous bias scores on Open-BBQ by social category for the LLaDA models, sorted in descending order by the values of the base model.

\mathrm{Mean}=\Sigma(\mathrm{Category\ bias})/11

\mathrm{SD}

is the standard deviation of the 11 category biases.

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	LLaDA-8B / LLaDA-1.5	Dream-7B
generation window	256 tokens	256 tokens
denoising steps	256 (before TPS)	256 (before TPS)
temperature	0 (greedy)	0.1
top- $p$	—	0.9
batch size per GPU	8	8

Appendix

Table 2: Sampling settings held fixed across the Open-BBQ evaluation. Block length, remasking strategy, and TPS are the sampling configurations of § 3.1 and are omitted; every other setting follows the reference configuration named in the text.

Benchmark	Items	Wall time
Open-BBQ, base model	10,000	127 min
Open-BBQ, with adapter	10,000	155–157 min
MMLU, IFEval and GSM8K, base model	15,902	101 min
MMLU, with adapter	14,042	126–128 min
IFEval, with adapter	541	9 min
GSM8K, with adapter	1,319	21–22 min

Appendix

Table 3: Wall time of one evaluation run on eight H200 GPUs. Rows with an adapter give the range over the three tuned models compared in § 5 . The utility benchmarks of the base model were run as a single script, so only their combined time is reported.

Category	Samples	Category	Samples
age	628	race $\times$ gender	2,728
disability status	268	race $\times$ SES	1,908
gender identity	972	religion	204
nationality	528	SES	1,172
physical appearance	268	sexual orientation	148
race/ethnicity	1,176	total	10,000

Appendix

Table 4: Samples of the 10k evaluation split by social category. The split holds 10,000 samples in 2,415 context groups, drawn with seed 0 from the 58,492 samples of Open-BBQ. Every category is split

1{:}1

between ambiguous and disambiguated samples.

ID	Instruction	Tokens	Box failure rate
P1	Put the number of your answer in \boxed{}.	31	$11.4\%$
P2	Finish with \boxed{N} -- just the option number, nothing after it.	36	$9.7\%$
P3	Reason briefly, then finish with \boxed{N} -- the option number, not its text. Nothing after it.	43	$9.2\%$
P4	At the very end of your response, give your final answer as the number of the option you choose, inside \boxed{}, for example \boxed{<number>}. Write nothing after it.	59	$10.6\%$
P5	Finish your answer with a line in exactly this form: \boxed{N} / N is the number of the option you chose. Put only that digit inside the braces, not the option text, and write nothing after that line.	66	$5.5\%$
P6	Give your chain of thought first, then finish your answer with a line in exactly this form: \boxed{N} / N is the number of the option you chose. Put only that digit inside the braces, not the option text, and write nothing after that line. $\bigstar$ Adopted	74	$3.8\%$

Appendix

Table 5: The eight candidate instructions , by their box failure rate on the 660 items. Each instruction follows the options list. “Tokens” counts the whole appended block with the LLaDA-1.5 tokenizer. Under the same setting, P6 gives box failure rates of

5.7\%

on LLaDA-8B and

14.2\%

on Dream-7B.

Objective	Wall time
SFT	5 min
dOPSD	951 min
dOPSD with a fairness term	956 min

Appendix

Table 6: Wall time of one training run on eight H200 GPUs. Each run trains LLaDA-8B for 938 steps.

Setting	Value	Source
adapter	LoRA, $r=32$ , $\alpha=32$ , no dropout	Dat et al. (2026)
LoRA targets	q, k, v, up (LLaDA); all (Dream-7B)	Dat et al. (2026)
learning rate	$4\times 10^{-5}$ , linear decay to 0, no warm-up	sweep
training length	938 steps ( $=3$ epochs of 10k samples)	Dat et al. (2026)
effective batch size	32 ( $4$ per GPU $\times$ 8 GPUs)	Dat et al. (2026)
seed	1234	—

Appendix

Table 7: Training settings shared by the tuned models. “Source” says whether a value was selected by a sweep in this work or taken from prior work or its released implementation. Settings not listed are left at the defaults of the released code.

Category	groups	terms	Category	groups	terms
Age	13	39	Race $\times$ gender	28	342
Disability status	9	36	Race $\times$ SES	21	193
Gender identity	21	263	Religion	10	39
Nationality	13	223	SES	12	57
Physical appearance	12	55	Sexual orientation	7	23
Race/ethnicity	13	132

Appendix

Table 8: Size of the bias group lexicon. 159 groups over the 11 social categories of Open-BBQ. groups counts the term sets a fairness term equalizes and terms the terms they hold. A term can be registered to more than one category, but never to two groups of the same category.

Figure 6: How the bias group lexicon is built. Each box is one step, with a term or group that passes it (blue) and one that it filters (red); all examples are taken from the build of the lexicon the main experiments train on. Steps marked “C4” read the corpus statistics gathered from the subset of C4 corpus.

Category	$\|g\|$	example terms
Age	3	old , young , healthy
Age	5	infant , boyfriend , toddler , teenager , teenage
Disability status	4	healthy , normal , smart , slow
Disability status	5	dumb , diabetic , lame , autistic , handicapped
Gender identity	3	he , she , him
Gender identity	9	John , Jesus , David , Paul , James , Michael , Mark , George , Mary

Appendix

Table 9: Two examples from each social category of the lexicon the main experiments train on. Terms are listed in descending corpus frequency, which is the order the bands are cut in, and are reproduced verbatim: casing is part of the vocabulary entry.

|g|

is the number of terms in the group.

		ambiguous		disamb.	lexicon
groups	pair filter	acc $\uparrow$	bias $\downarrow$	acc $\uparrow$	skew $\downarrow$	mass
Base	—	.703	.123	.942	.886	1.13
dOPSD, no term	—	.771 †	.072 †	.937	.888	1.52
option pair only	—	.759 †	.062 †	.932 ‡	.864	1.42
$+$ source groups	auto	.869 †	.047 †	.936	—	—
$+$ source groups	manual	.859 †	.052 †	.939	—	—

Appendix

Table 10: Effect of the lexicon, and of manual against automatic pair filtering (LLaDA-8B, dOPSD with the log-prob. var. , 300 steps). groups names the groups the term is given; pair filter says whether the mined terms were filtered by hand (manual) or not (auto). mass is scaled by

10^{3}

. Bold is the best in each column.

\dagger

\ddagger

: significant improvement/degradation over the base model under a paired cluster bootstrap (

5

k resamples,

95\%

interval excluding zero; clusters are BBQ templates).

			ambiguous		disamb.
Objective	fairness term	$\lambda$	acc $\uparrow$	bias $\downarrow$	acc $\uparrow$
Base	—	—	.703	.123	.942
SFT	—	—	.965 †	.015 †	.965 †
	JS div.	10	.986 †	.007 †	.986 †
	JS div.	30	.999 †	.000 †	.990 †
	log-prob. var.	1	.998 †	.001 †	.991 †

Appendix

Table 11: Effect of the weight

\lambda

of the JS div. (LLaDA-8B, 938 steps).

\dagger

\ddagger

: significant improvement/degradation over the base model under a paired cluster bootstrap (

5

k resamples,

95\%

interval excluding zero; clusters are BBQ templates).

Figure 7: Effect of debiasing with SFT and dOPSD on LLaDA-1.5. Disambiguated bias score (upper), ambiguous bias score (middle), and box failure rate (lower) for the base, SFT, and dOPSD LLaDA-1.5 models, across 24 sampling configurations. The bottom panels give the rank of LLaDA-1.5 among the three models by the ambiguous bias score of their SFT and dOPSD models.

Figure 8: Effect of the three fairness terms on LLaDA-1.5. Each bar is the range of the ambiguous bias score over the 24 sampling configurations, drawn per remasking strategy; the marker is the default configuration, and the dashed rule is the base model at the default configuration.

Figure 9: Effect of debiasing with SFT and dOPSD on Dream-7B. Disambiguated bias score (upper), ambiguous bias score (middle), and box failure rate (lower) for the base, SFT, and dOPSD Dream-7B models, across 24 sampling configurations. The disambiguated panel is cut at

-0.2

; two dOPSD+log-prob. var. configurations with box failure rates close to one fall below it. The bottom panels give the rank of Dream-7B among the three models by the ambiguous bias score of their SFT and dOPSD models.

Figure 10: Effect of the three fairness terms on Dream-7B. Each bar is the range of the ambiguous bias score over the 24 sampling configurations, drawn per remasking strategy; the marker is the default configuration, and the dashed rule is the base model at the default configuration.

Figure 11: Effect of debiasing tuning on utility on LLaDA-1.5 and Dream-7B. Ambiguous bias score on Open-BBQ against accuracy on GSM8K, IFEval, and MMLU for LLaDA-1.5 (upper) and Dream-7B (lower) tuned on Open-BBQ.

\dagger

\ddagger

: significant improvement/degradation in accuracy over the base model under a paired cluster bootstrap (

5

k resamples,

95\%

interval excluding zero; clusters are prompts for IFEval and questions otherwise).

LLaDA-1.5
			SFT				dOPSD
Category	$n$	Base	—	log-ratio	JS div.	var.	—	log-ratio	JS div.	var.
Age	314	$-$ .446	$-$ .134	$-$ .121	$-$ .022	$-$ .140	$-$ .350	$-$ .247	$-$ .040	$-$ .198
Physical appearance	134	$-$ .388	$-$ .045	$-$ .022	$-$ .000	$-$ .060	$-$ .267	$-$ .276	$-$ .066	$-$ .119
SES	586	$-$ .326	$-$ .019	$-$ .014	$-$ .000	$-$ .037	$-$ .209	$-$ .148	$-$ .024	$-$ .068
Disability status	134	$-$ .254	$-$ .075	$-$ .060	$-$ .000	$-$ .015	$-$ .194	$-$ .218	$-$ .043	$-$ .060

Appendix

Table 12: Bias and the effect of debiasing by social category on LLaDA-1.5 and Dream-7B. Ambiguous bias scores on Open-BBQ by social category for the LLaDA-1.5 and Dream-7B models, each block sorted in descending order by the values of its base model.

\mathrm{Mean}=\Sigma(\mathrm{Category\ bias})/11

\mathrm{SD}

is the standard deviation of the 11 category biases. Excluding the exceptions named in the text, the Spearman correlation between the category ranking of each tuned model and that of its base model is at least

.72

on LLaDA-1.5 and

.63

on Dream-7B (ties given the average rank).

LLaDA-8B
			SFT				dOPSD
Category	$n$	Base	—	log-ratio	JS div.	var.	—	log-ratio	JS div.	var.
Age	314	$-$ .013	$-$ .003	$-$ .006	$-$ .048	$-$ .010	$-$ .013	$-$ .010	$-$ .026	$-$ .007
Physical appearance	134	$-$ .134	$-$ .104	$-$ .149	$-$ .173	$-$ .158	$-$ .105	$-$ .076	$-$ .105	$-$ .098
SES	586	$-$ .040	$-$ .032	$-$ .010	$-$ .034	$-$ .038	$-$ .051	$-$ .045	$-$ .040	$-$ .049
Disability status	134	$-$ .084	$-$ .015	$-$ .000	$-$ .045	$-$ .030	$-$ .030	$-$ .023	$-$ .015	$-$ .098

Appendix

Table 13: Disambiguated bias by social category. Disambiguated bias scores on Open-BBQ by social category for the three models, every block sorted as in Tab. 1 .

\mathrm{Mean}=\Sigma(\mathrm{Category\ bias})/11

\mathrm{SD}

is the standard deviation of the 11 category biases.

Disambiguated bias score, low-confidence remasking
Method	$B=128$				$B=32$				$B=8$
$\mathrm{TPS}$	1	2	4	8	1	2	4	8	1	2	4	8
Base	$\scriptstyle-$ .008	$\scriptstyle-$ .010	$\scriptstyle-$ .018	$\scriptstyle-$ .009	$\scriptstyle-$ .010	$\scriptstyle-$ .010	$\scriptstyle-$ .011	$\scriptstyle-$ .018	$\scriptstyle-$ .013	$\scriptstyle-$ .009	$\scriptstyle-$ .014	$\scriptstyle-$ .005
SFT	$\scriptstyle-$ .011	$\scriptstyle-$ .007	$\scriptstyle-$ .014	.006	$\scriptstyle-$ .006	$\scriptstyle-$ .001	$\scriptstyle-$ .009	$\scriptstyle-$ .011	$\scriptstyle-$ .006	.000	$\scriptstyle-$ .005	$\scriptstyle-$ .016
dOPSD	$\scriptstyle-$ .012	$\scriptstyle-$ .012	$\scriptstyle-$ .017	$\scriptstyle-$ .007	$\scriptstyle-$ .016	$\scriptstyle-$ .014	$\scriptstyle-$ .012	$\scriptstyle-$ .013	$\scriptstyle-$ .013	$\scriptstyle-$ .003	.002	$\scriptstyle-$ .002
dOPSD+log-prob. var.	$\scriptstyle-$ .014	$\scriptstyle-$ .014	$\scriptstyle-$ .014	.000	$\scriptstyle-$ .004	$\scriptstyle-$ .014	$\scriptstyle-$ .006	.001	$\scriptstyle-$ .012	$\scriptstyle-$ .006	$\scriptstyle-$ .007	.010

Appendix

Table 14: Sampling sweep on LLaDA-8B. Disambiguated bias score, ambiguous bias score, and box failure rate for each method at each of the 24 sampling configurations. The values are those plotted in Figs. 2 , 3 , and 4 .

Disambiguated bias score, low-confidence remasking
Method	$B=128$				$B=32$				$B=8$
$\mathrm{TPS}$	1	2	4	8	1	2	4	8	1	2	4	8
Base	$\scriptstyle-$ .015	$\scriptstyle-$ .003	$\scriptstyle-$ .007	$\scriptstyle-$ .004	$\scriptstyle-$ .013	$\scriptstyle-$ .008	$\scriptstyle-$ .019	$\scriptstyle-$ .017	$\scriptstyle-$ .015	$\scriptstyle-$ .006	$\scriptstyle-$ .018	$\scriptstyle-$ .013
SFT	$\scriptstyle-$ .015	$\scriptstyle-$ .009	$\scriptstyle-$ .010	$\scriptstyle-$ .011	$\scriptstyle-$ .017	$\scriptstyle-$ .006	$\scriptstyle-$ .011	$\scriptstyle-$ .010	$\scriptstyle-$ .006	$\scriptstyle-$ .016	$\scriptstyle-$ .007	$\scriptstyle-$ .003
dOPSD	$\scriptstyle-$ .015	$\scriptstyle-$ .017	$\scriptstyle-$ .016	$\scriptstyle-$ .006	$\scriptstyle-$ .010	$\scriptstyle-$ .009	$\scriptstyle-$ .015	$\scriptstyle-$ .010	$\scriptstyle-$ .015	$\scriptstyle-$ .011	$\scriptstyle-$ .003	$\scriptstyle-$ .006
dOPSD+log-prob. var.	$\scriptstyle-$ .015	$\scriptstyle-$ .013	$\scriptstyle-$ .010	$\scriptstyle-$ .021	$\scriptstyle-$ .013	$\scriptstyle-$ .014	$\scriptstyle-$ .003	$\scriptstyle-$ .015	$\scriptstyle-$ .011	$\scriptstyle-$ .011	$\scriptstyle-$ .019	.004

Appendix

Table 15: Sampling sweep on LLaDA-1.5. Disambiguated bias score, ambiguous bias score, and box failure rate for each method at each of the 24 sampling configurations. The values are those plotted in Figs. 2 , 7 , and 8 .

Disambiguated bias score, low-entropy remasking
Method	$B=128$				$B=32$				$B=8$
$\mathrm{TPS}$	1	2	4	8	1	2	4	8	1	2	4	8
Base	$\scriptstyle-$ .018	$\scriptstyle-$ .008	$\scriptstyle-$ .005	$\scriptstyle-$ .108	$\scriptstyle-$ .018	.008	$\scriptstyle-$ .021	$\scriptstyle-$ .088	$\scriptstyle-$ .017	$\scriptstyle-$ .012	$\scriptstyle-$ .034	.067
SFT	$\scriptstyle-$ .007	$\scriptstyle-$ .006	$\scriptstyle-$ .009	$\scriptstyle-$ .050	$\scriptstyle-$ .007	$\scriptstyle-$ .005	$\scriptstyle-$ .011	$\scriptstyle-$ .043	$\scriptstyle-$ .009	$\scriptstyle-$ .003	$\scriptstyle-$ .053	$\scriptstyle-$ .041
dOPSD	$\scriptstyle-$ .007	$\scriptstyle-$ .007	$\scriptstyle-$ .017	.545	$\scriptstyle-$ .002	$\scriptstyle-$ .022	.039	.444	$\scriptstyle-$ .019	$\scriptstyle-$ .017	.058	.000
dOPSD+log-prob. var.	$\scriptstyle-$ .015	$\scriptstyle-$ .011	.011	.000	$\scriptstyle-$ .012	.004	.032	$\scriptstyle-$ .333	$\scriptstyle-$ .011	$\scriptstyle-$ .029	.021	.000

Appendix

Table 16: Sampling sweep on Dream-7B. Disambiguated bias score, ambiguous bias score, and box failure rate for each method at each of the 24 sampling configurations. The values are those plotted in Figs. 2 , 9 , and 10 .

LLaDA-8B
	Open-BBQ bias score $\downarrow$		Accuracy $\uparrow$
Method	ambiguous	disambiguated	GSM8K	IFEval	MMLU
Base	.123	$\scriptstyle-$ .010	.754	.619	.522
SFT	.015	$\scriptstyle-$ .007	.751	.621	.602 †
SFT+log-ratio	.017	$\scriptstyle-$ .014	.751	.631	.576 †
SFT+JS div.	.000	$\scriptstyle-$ .004	.756	.628	.576 †

Appendix

Table 17: Fairness and utility at the default sampling configuration. Ambiguous and disambiguated bias score on Open-BBQ and accuracy on GSM8K, IFEval, and MMLU for each method on the three models. The values are those plotted in Figs. 5 and 11 . GSM8K is sampled with

B=8

instead of

B=32

(App. A.1 ).

\dagger

\ddagger

: significant improvement/degradation in accuracy over the base model under a paired cluster bootstrap (

5

k resamples,

95\%

interval excluding zero; clusters are prompts for IFEval and questions otherwise).

Social Bias in Language Models

Social bias in language models is the study and mitigation of unfair stereotypes and disparities in generated text, judgments, and decisions.

See all 199 papers

Who Gets Unmasked? Sampler-Induced Bias in Diffusion Language Models
Oct 17, 2026Xujun Che, Farsheed Haque, Depeng Xu +2Diffusion Language Models Language Model Bias Evaluation
University of North Carolina at Charlotte, Charlotte, NC, USA
Noise Out, Bias In: Targeted Bias Injection in Diffusion Language Models via Closed-Loop Activation Steering
Oct 5, 2026Sarim Hashmi, Mukul Ranjan, Abdelrahman Elsayed +3Adversarial Attacks on Diffusion Models Diffusion Language Models
Mohamed bin Zayed University of Artificial Intelligence
FairLMs: A Turnkey Library for Fairness in Language Models
Sep 18, 2026Jiale Zhang, Michael Larionov, Zichong Wang +2Language Model Bias Evaluation Social Bias in Language Models
Department of Biostatistics and Health Data Science Indiana University Indianapolis, IN, United States · Department of Statistics and Data Science Carnegie Mellon University Pittsburgh, PA, United States · Knight Foundation School of Computing and Information Sciences Florida International University Miami, FL, United States

Oct 17, 2026J/K move · Enter open · S save

Who Gets Unmasked? Sampler-Induced Bias in Diffusion Language Models

Xujun Che, Farsheed Haque, Depeng Xu, Xi Niu+1

University of North Carolina at Charlotte, Charlotte, NC, USA

Diffusion Language Models Language Model Bias Evaluation Gender Bias in Language Models

Open paper AskPDF arXiv

Diffusion language models (dLLMs) generate by gradually unmasking a canvas, so their samplers decide not only which token to write but also when to commit it. We show that this second decision changes who appears in generated text, a sampler-induced bias on top of the model's own. Common commit rules, including LLaDA's default, commit the positions whose sampled tokens are most probable; this peeking favors majority options. To separate the sampler from the model, we read the model's belief about a fork, such as a protagonist's gender, in one forward pass, and show that this belief matches exact sampling from the model. Measured against it, peeking samplers on LLaDA and Dream amplify occupational gender stereotypes and collapse opinion distributions, and confidence-based ordering shifts gender-balanced occupations toward men, more so under faster parallel decoding. We propose Fork-First Sampling, which decides the fork before anything else is written: it follows the belief or any target distribution, improves fluency, and avoids the stories that remark on the protagonist's gender, which prompt-based fixes produce, at the cost of one forward pass. Fairness evaluations of dLLMs should report the commit rule.

Abstract

Social Bias in Language Models

Explore similar work

Who Gets Unmasked? Sampler-Induced Bias in Diffusion Language Models

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

FairLMs: A Turnkey Library for Fairness in Language Models