Modern language models are increasingly asked to handle several tasks within a single prompt. This requires not only solving each sub‑task correctly but also weaving the answers into a coherent response. We identify a fundamental limitation: when tasks are composed, models make substantially more errors even though they can solve every constituent task reliably in isolation. This degradation cannot be fully explained by the accumulation of independent errors. Experiments on open‑weight models and the frontier model Claude Opus 5 point to an implicit thinking budget. Under composition, models allocate far fewer tokens per task than they do when the tasks are presented in isolation, even though the full sequence remains well within the context window. Consequently, tasks that need longer solution traces in isolation become more fragile under composition, exhibiting higher error rates. Conversely, non‑fragile tasks receive more reasoning than necessary when solved alone, wasting compute. By grouping tasks according to their fragility, we can manipulate the implicit budget to cut token usage by about 43 % across a range of math and linguistic tasks while preserving accuracy.
Review: The paper uncovers a hidden resource‑allocation mechanism in multi‑task prompting, offering practical guidance for crafting more efficient prompts. Exploiting task fragility enables significant compute savings without sacrificing performance, a finding that is highly relevant for real‑world deployments.