TBNAI
Noise Out, Bias In: Targeted Bias Injection in Diffusion Language Models via Closed-Loop Activation Steering
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 04/10/2026, 20:00:00 UTC
Briefing created: 18/10/2026, 05:00:36 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 05:18:24 UTC.
Briefings
A 2026 study reveals targeted bias injection attacks on masked diffusion language models (dLLMs) using closed-loop activation steering. The attack employs a proportional-integral controller to track target-answer probabilities during denoising, dynamically adjusting steering vectors. On ambiguous BBQ questions, the attack increased LLaDA-8B-Instruct's preference for a targeted group by 16.7 percentage points, surpassing fixed-strength baselines. On SocialStigmaQA, it raised stigmatizing answer selection from 17.6% to 58.1%. The attack requires approximately 40 minutes per GPU run and relies on the model's evolving internal activations rather than a static frozen model.
Key points
- The attack uses a proportional-integral (PI) controller to track target-answer probabilities during denoising steps.
- On ambiguous BBQ questions, the attack increased LLaDA-8B-Instruct's preference for a targeted group by 16.7 percentage points.
Context
Masked diffusion language models generate text by iteratively denoising masked positions, exposing answers at every step before commitment. This allows adversaries to exploit the model's evolving internal activations to adjust interventions dynamically, unlike autoregressive models that commit answers at a single step.
Limitations
Coverage is limited to specific ambiguous questions and demographic targets; results may vary across other datasets. The attack duration of 40 minutes per GPU run is a constraint for practical deployment. The study relies on internal activation data rather than external ground truth labels, introducing potential measurement uncertainty.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 18/10/2026, 05:00:36 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeida3xgdpsuzastvrv6zt6w6cjyaxhciyglyayd224j4pfg4ydekru