ADMET-EvO: a 7-step evidence-gated architecture for self-evolving AI agents
MaryamMiradi · x · 2026-10-23
Dr. Maryam Miradi breaks down the ADMET-EvO paper (arXiv:2609.10121), calling out the dangerous loop in most 'self-improving' agents: the agent proposes a change, tests it, and grades its own homework. ADMET-EvO's key split: the LLM decides what evidence to acquire next, while deterministic components decide what that evidence actually proves.
Her 7-step roadmap highlights:
- Define the system contract: fix task, metrics, boundaries and allowed evidence upfront; specify what the agent may change and what stays protected.
- Diagnose before changing: locate weaknesses — data scarcity, retrieval failure, model limitation, distribution shift.
- Later steps follow a contract–diagnose–hypothesize–intervene–evaluate–revise loop with model freezing and blind certification, carrying supported/rejected/inconclusive verdicts forward.
Applied to ADMET prediction, the framework turns fragmented research into a persistent, evidence-gated research process, improving reliability and applicability by evolving the research process itself.
More from coding & agent
- Cognition open-sources Devin agent plugin to patch the day's CVE wave — Scobleizer · 2026-10-23
- Fine-tuning models solo with a Grok bot and rented GPUs — Rasmic · 2026-10-23
- One prompt found pre-auth PHP RCEs: agent-guided vuln hunting timeline — chrisrohlf · 2026-10-23
- AutoRouter launches beta: auto-routes each coding task across Claude, GPT and other models to cut token spend — floguo · 2026-10-23
- How Speak uses GPT-Live and a Decisions API for ultra-responsive live AI language tutoring — pbbakkum · 2026-10-23
- Google Live API adds decisions model triggers, faster than function calls for voice agents — pbbakkum · 2026-10-23