When should masked diffusion LMs adapt decoding? Selective adaptation captures 56.9% of oracle gains
HOLILAB · hf · 2026-10-01
HOLILAB's paper "Unmask the State" studies when state adaptation matters for masked diffusion language models (MDMs), where unmasking strategy is an inference decision.
Framework: MDM inference is organized into five axes — score, cardinality, region, commitment, and planning — with "adaptation opportunity" defined as the one-step utility advantage of the best candidate action over a validation-selected fixed action, characterized via strategy reversals.
Findings:
- Across 3 MDMs and 10 tasks, adaptation opportunities are highly heterogeneous, with some regimes showing concentrated, predictable one-step gains.
- Lightweight detectors calibrated on validation prompts enable selective adaptation: adapting only the top 10% of states on LLaDA-8B constrained JSON filling captures 56.9% of the candidate-set oracle opportunity.
Conclusion: state adaptation is most useful when applied selectively rather than uniformly.
More from Research
- OSWorld-Science Debuts: 146 Tasks Test How Well VLM Agents Handle Scientific Software — SciAILab · 2026-10-01
- Survey of Attention Evolution: Contextual Memory Becomes the Core of LLM Architecture Design — Zhentao Tan · 2026-10-01
- Hidden Dates in System Prompts Swing LLM Eval Scores by Up to 14% — Mario Sanz-Guerrero · 2026-10-01
- CheatBench Launches to Measure Reward Gaming and Cheating in AI Agents — cais · 2026-10-01
- KLS partially cracked: arXiv paper's core proof ideas generated by AI — burny_tech · 2026-10-01
- Newton's method: when it converges, barely converges, and fails entirely — burny_tech · 2026-10-01