The efficiency fix: judge, schedule, route
An arXiv preprint (cross-listed in CS.CL and CS.AI) from an unnamed author team proposes AIBuildAI-2.5, an agentic system that builds AI models by framing model development as a code search problem solved through tree search with LLM agents. The authors identify three efficiency weaknesses in existing agents of this type: few candidates can be executed within a realistic budget, so reward rankings rest on sparse, noisy scores; no resource-aware strategy schedules training jobs, hurting hardware utilization; and every agent call is served by one powerful model, inflating inference cost. The system addresses these with an LLM-guided tree search in which a judge scores each candidate on expected improvement, grounding, and feasibility, and a selector ranks candidates from those scores plus search state; a scheduler that launches training jobs according to hardware resource status; and a router that assigns lower-cost LLMs to less demanding tasks while reserving the most capable model for the hardest sub-tasks. [1][2]
Reported results
The preprints state AIBuildAI-2.5 ranks first on MLE-Bench with a medal rate of 73.3% and outperforms a strong baseline on six autonomous AI research tasks from AIRS-Bench. These are the authors' own reported figures from the preprint, not independent evaluations, and the supplied evidence includes no third-party replication, methodology detail, or baseline naming. Treat the ranking and medal rate as claims pending external verification. [1][2]
Sources
- AIBuildAI-2.5: Efficient Autonomous AI Model Development Through LLM-Guided Tree Search
ArXiv CS.CL (Computation and Language) · Reporting ·
- AIBuildAI-2.5: Efficient Autonomous AI Model Development Through LLM-Guided Tree Search