Diffusion large language models generate text in parallel rather than left-to-right, offering a different trade-off from autoregressive models. But they remain prone to hallucination. A new arXiv paper proposes TRACESTEER, a method that steers the model's internal activations to reduce such false outputs.

Activation steering is a known technique for influencing model behavior by modifying internal states. TRACESTEER makes this trace-driven, meaning it uses information from generation traces to decide how to adjust the activations. The work is aimed specifically at the parallel token generation used by diffusion LLMs.

Because the paper is new and the abstract is brief, details on evaluation and baselines are not yet available in the digest. The authors position the method as a promising mitigation approach, but independent verification would be needed.