Diffusion LMs allocate compute by task: novel outputs cost more forward passes
uclanlp · hf · 2026-10-24
Researchers from UCLA and LMU Munich analyze how diffusion language models allocate computation across 10 tasks. Since these models can emit multiple tokens per forward pass and decide mid-generation how many passes to spend, outputs of the same length can vary widely in cost.
Key findings:
- Forward passes per token depend heavily on the task: copying or lightly editing input text is cheap, while open-ended writing is far more expensive.
- Models automatically spend more computation on novel, rare, or incorrect outputs.
- Decoding speed itself can be exploited to improve inference efficiency and reliability.
The study covers two diffusion models and provides an empirical basis for cost estimation and adaptive inference in diffusion LMs.
More from Research
- Reuse, Don't Recompute: A 4 A.M. ICLR Story Meets Meta² Agent Evolution — furongh · 2026-10-24
- Mapping Economic Theory With Agentic AI: LLM Agents Turn Theory Papers Into Graphs — mattbeane · 2026-10-24
- Robotics model leaderboards are premature, argues practitioner: task-specific evals first — paigeinsf · 2026-10-24
- UTokyo/RIKEN lab brings CAROT cross-lingual alignment method to EMNLP, +11.2 accuracy points — Bollegala · 2026-10-24
- GPT-6 Astra solves 49 of 50 research-level physics problems, catches experts' own errors — MoonL88537 · 2026-10-24
- Rice & NVIDIA's Sensorimotor Alignment predicts robot policy success without training (r=0.717) — CSProfKGD · 2026-10-24