On the Fragility of Tasks under Composition: Implicit Budgeting in Language Models

On the Fragility of Tasks under Composition. Implicit Budgeting in Language Models. By Chenpan Li and four others, from the Technical University of Munich and MIT CSAIL. Published on the nineteenth of October. Its code repository is worked out: about one hundred and sixty source files, with a readme and a licence.

Language models are more and more often handed several tasks inside a single prompt. That asks two things of them: solve each task correctly, and then fold the answers into one coherent response. The authors identify what they call a fundamental limitation here. Models make substantially more errors when tasks are composed, even when every one of the constituent tasks is something they solve reliably on its own. And the drop is larger than the accumulation of independent errors can fully explain. Something beyond compounding chance is going on.

Their account is an implicit thinking budget. The evidence comes from open-weight models and from the frontier model Claude Opus 5. Under composition, models spend substantially fewer tokens on each task than they do in isolation, and this happens even when the full sequence sits well inside the context window. Running out of room is not the reason. The consequence is a pattern of fragility. Tasks that need longer solution traces when solved alone are the ones that suffer most when combined, meaning that prompts containing them draw more errors.

The other side of the budget shows up in isolation. There, tasks that are not fragile receive more reasoning than they need, and that extra compute is wasted.

The claim rests on the open-weight models studied and a single frontier model, and the task families the abstract names are math and linguistic ones. The authors say their findings point to the budget; they do not say it is the whole story, only that independent error accumulation cannot fully account for the effect. It is an implicit pattern inferred from token usage, not a setting anyone wrote down.

It is also something that can be steered. By grouping tasks according to how fragile they are, the implicit budgeting can be manipulated to cut token usage by forty-three percent across those math and linguistic tasks, while accuracy is preserved.

Here is a failure that looks like a capacity problem and turns out to be an allocation problem. A model given several jobs at once seems to ration its thinking across them, starving the ones that need long traces and overfeeding the ones that do not. Once that rationing is understood, it stops being only a weakness: arranging tasks with it in mind saves tokens without losing answers.