Mainstream large language models almost universally use the autoregressive (AR) generation paradigm, outputting tokens one after another. This approach has a natural inference-speed bottleneck — the longer the sequence, the more generation time. In the past year, diffusion language models (dLLM) have attracted attention as a disruptive alternative; Google Gemini Diffusion, Inception Mercury, and others have launched, promising the higher throughput of parallel decoding.
But can speed and accuracy really coexist? The ICML 2026 paper d3LLM (pseuDo-Distilled Diffusion LLM) tackles this question head-on, proposing a systematic solution.
Core innovation: Pseudo-Trajectory Distillation
d3LLM, from the Hao-AI Lab team, was published on arXiv (2601.07568). The team found that existing diffusion models learn inefficiently under random masking strategies — the student model doesn't know which tokens the teacher model prefers to decode first. So they extracted Pseudo-Trajectories from the teacher model's real decoding order, using them to guide the student model toward a generation order closer to that of autoregressive models; this design alone brought an 18% TPF (Token Per Forward) improvement.
Additionally, the team designed a Progressive Noise Schedule and a Progressive Window Sizing strategy, contributing an extra 12% and 8% TPF improvement respectively.
Inference side: multi-block parallel decoding
At decoding, d3LLM introduces entropy-based multi-block parallel decoding — each forward pass handles multiple blocks simultaneously, with entropy values used to filter high-confidence tokens. Combined with periodic KV-Cache refresh, throughput in long-context scenarios is boosted another 35%. Layered with the EOS early-stop mechanism, the overall system achieves up to 5× speedup (vs. Qwen-2.5-7B-it on H100 GPU).
Accuracy-speed trade-off: introducing the AUP metric
One of the paper's key contributions is pointing out the trade-off law between accuracy and parallelism: processing more tokens per forward pass almost inevitably comes with accuracy loss. To address this, the team proposes the AUP (Accuracy Under Parallelism) metric, which holistically measures the speed-accuracy balance, helping the community more fairly evaluate the cost-performance of diffusion language models versus autoregressive models.
Why this is worth attention
d3LLM's significance is not in dethroning autoregression, but in proving that the diffusion paradigm can, in engineering terms, very closely match or even exceed the autoregressive baseline. With 5× speedup plus native support from mainstream inference engines like SGLang, dLLM's path from lab to production is now clear. For cost-sensitive deployment scenarios, this is a direction worth continued attention.