Stipe Frković, Metod Jazbec, Dan Zhang, Christian A. Naesseth, Ilija Bogunovic, Eric Nalisnick
Masked diffusion language models (dLLMs) have recently emerged as a competitive alternative to autoregressive language models, with the promise of faster inference via parallel token generation. A notable limitation of the masked formulation, however, is that once a token has been unmasked it can no longer be revised, leaving dLLMs vulnerable to early sampling mistakes. Confidence-based remasking methods aim to address this in a training-free manner by revising low-confidence tokens, with encouraging early results. In this work, we compare the accuracy-compute Pareto frontiers of three such methods (WINO, Saber, and COVER) to those of confidence-based unmasking alone (Fast-dLLM). We find that under greedy decoding, all three perform on par with Fast-dLLM at commonly used shorter block lengths; gains of remasking appear only at longer block lengths, where confidence-based unmasking is known to degrade. Under stochastic decoding, WINO reduces generation diversity beyond that of unmasking alone, while it helps recover from errors introduced by learned Bernoulli unmasking policies. Our results show that the benefits of confidence-based remasking are highly setting-dependent, underscoring the need for more comprehensive evaluation. Finally, our analyses point to the dLLM, not the remasking rule, as the bottleneck, suggesting that effective remasking may require changes to training, not only to inference.
Tianjian Li, Jingyu Zhang, William Jurayj, Xi Wang, Chuanyang Jin, Mehrdad Farajtabar, Eric Nalisnick, Daniel Khashabi
AI 速览：提出SelfCompact框架，让语言模型智能体在推理时自主决定何时及如何压缩上下文，以降低token成本并提升性能。
Long agent traces composed of chains of thought and tool calls accumulate stale content that anchor subsequent generations, and eventually outgrow the context window. Existing scaffolds mitigate it with fixed-interval compaction triggered at a token threshold. Such triggers pay no heed to trajectory structure, risking discard of partial results mid-derivation or mid-search. We propose SelfCompact, a scaffold that allows the model itself to decide when and how to compact. Specifically, it pairs two inference-time elements: (i) a compaction tool the model invokes to summarize the accumulated context, and (ii) a lightweight rubric specifying when to fire (a sub-task has resolved, or the trajectory is converging) and when to suppress (mid-derivation, or when stuck). Both are needed. The tool alone is unevenly used across open-weight models, often invoked at unhelpful moments or not at all; the rubric alone cannot act. Together, they elicit effective adaptive compaction without any fine-tuning or external supervision. We present empirical results on six benchmarks (competitive math and agentic search) and seven models. Our results show that SelfCompact matches or exceeds fixed-interval summarization at a fraction of the token cost, improving over a no-summarization baseline by up to 18.1 points on math and 5-9 points on agentic search at 30-70% lower per-question cost. Our results expose a meta-cognitive gap: although unprompted models cannot reliably tell when their own context is rotting, a lightweight rubric closes this gap, reframing when to compact as a capability that scaffolds can supply without training.