InjazAI

A new arXiv technical report introduces AutoFyn, a non-parametric Expert Iteration harness that improves long-horizon agents by updating persistent state from verified rewards instead of retraining model weights, with results in olympiad math, data science and cybersecurity.

1 min read

AutoFyn: A Training-Free Agent Loop Beats Provider Coding Agents on IMO 2026

What is new in the AutoFyn paper?

FAQ

What is AutoFyn?

An agent harness inspired by the Expert Iteration algorithm that adapts a frozen model across rounds by updating persistent state from verified reward signals rather than model weights.

How does AutoFyn differ from fine-tuning or retraining?

It never changes model weights. Durable information is reintroduced only through explicit interfaces such as persistent memory files, reports and repository state, which lowers cost and improves traceability.

What results did AutoFyn report?

Higher scores on six fresh IMO 2026 problems, the top-ranked agent on the Spider 2.0 dbt benchmark, and 16 maintainer-confirmed vulnerability advisories in open-source projects.

Should MENA AI teams adopt it now?

It is worth piloting for long-horizon agent tasks such as data analysis and code review, but it remains an early technical report requiring internal validation before production use.

Source: arXiv cs.AI

AI-assisted content, human-reviewed.