$ cat wiki/papers/2026/2610.34491-reasoning-effort.md
Reasoning effort: cost and task-level reliability
TL;DR
Higher reasoning effort can improve aggregate accuracy while failing tasks that lower effort solves. The authors find that starting low and escalating after verified failure matches or improves on fixed top effort in their tested settings. (source)
Authors & Org
Kaiser Sun, Yongchan Kwon, Zain Hasan, Shang Zhu, Kaitlyn Zhou and Federico Bianchi. Affiliations are not established by the abstract page. First submitted 2026-10-25. (source)
Method
The study compares counterfactual effort settings for proprietary and open-weight models on four benchmarks covering single-call reasoning and long-horizon agent tasks. It examines accuracy, spending and trajectory behavior, treating effort as a purchase whose marginal benefit a successful outcome alone cannot reveal. (source)
Results
On easy tasks solved at every effort level, the highest setting costs up to 42x as much as the lowest. Higher effort produces more verification and solution searching without necessarily changing the outcome. Escalation after verified failure matches or improves performance across the evaluated models and benchmarks, with lower costs for most proprietary models. These are author-reported findings, not a universal cost reduction. (source)
Significance
For Test-Time Compute (Inference-Time Compute Scaling) and Agents (LLM Agents), more inference is not a guarantee of instance-level reliability. Eval Harness Configuration should preserve the effort policy alongside the model and task. (source)
Open Questions
The abstract does not specify the full model/benchmark inventory or how verified failure is obtained in every setting. Whether the escalation policy transfers to tasks without reliable outcome checks remains open. Full paper and code were not reviewed. (source)
Cite
What Effort Level Should I Choose? On the Economics of Reasoning Effort in LLM Agents, arXiv:2610.34491, 2026. (source)