Thies Brockmoeller retweeted
Imagine giving Claude Opus 5 ten questions it reliably answers individually. Put them in one prompt, and it starts getting some wrong.
Excited to share our new work that investigates why. Our findings point to “implicit budgeting” in LLMs: as more questions share a prompt, reasoning per question shrinks, following an approximate power law (even with plenty of context space left). Thus, questions that need longer reasoning are hit hardest.
Understanding this also helps us batch questions more efficiently. For Opus 5 on GPQA-Diamond, that saves ~24% of tokens with roughly unchanged accuracy.
Huge credit to @chenpan_li for the thoughtful experiments and hard work behind this paper!
Thies Brockmoeller retweeted
📄 Paper: arxiv.org/abs/2610.27190
🌐 Project: chenpanli.github.io/fragilit…
💻 Code: github.com/chenpanli/fragili…
🤗 discussion: huggingface.co/papers/2610.2…
Thies Brockmoeller retweeted
“On the Fragility of Tasks under Composition: Implicit Budgeting in Language Models” (2610.27190)
Joint work with @ThiesJB, @StefanieJegelka, @sharut_gupta and @AspectStalence (@TU_Muenchen, @MIT, @MunichCenterML). Thank you all for a great collaboration!
#LLM #Reasoning
Thies Brockmoeller retweeted
The same mechanism can also save compute. Keep fragile tasks in small groups, batch the rest together, and you reduce token use by 43% without losing accuracy. It only needs the model’s outputs, so it works on closed models like Claude Opus 5 too.
Thies Brockmoeller retweeted
High-effort tasks are hit hardest. They end up with too little reasoning to succeed, so we call them fragile. This offers an explanation to why long-horizon tasks are hard, since they pack many steps into a single context.