Fei Ding's preprint proposes LLM 'hyper-threading' hypothesis

On an AIME 2025 development set, a setup the paper calls Concurrent Functional Loading beat both a baseline and a serial task-scheduling condition. Author Fei Ding's arXiv preprint, posted 23 August 2026, reads that as preliminary evidence for what it names the Model Hyper-Threading Hypothesis: that a language model, while generating one token at a time, can carry more than one task inside that step.

The framing pushes back on a common reading of attention patterns. The paper argues that "more dispersed attention can coexist with higher accuracy," offering that against the default view, which it says treats attention dispersion as "a signal of interference or error." Concurrent Functional Loading produced output lengths similar to the serial condition, and shorter on most problems, while showing greater attention dispersion and higher task-relevant coverage, "albeit with a heavier output-length tail."

Ding is careful about what has actually been shown. "Within-step concurrency and its causal mechanism still require direct tests," the paper says, describing its own evidence as "preliminary behavioral and correlational." The abstract names one benchmark, AIME 2025 dev, and reports no per-condition accuracy figures. Two researchers we track picked up the link the day it went up.