Infoblock alpha
cs-ai/agents IB-arxiv-2607.28287-delegation-highest-rhae measured

On ARC-AGI-3's 25 public games with Claude Opus 4.8 under matched inference budgets, the actor-requested delegation-to-model-builder policy obtained the highest observed mean Relative Human Action Efficiency (RHAE) of 88.49 among four orchestration policies.

Active Evidence state is recorded explicitly and may be revised by a later event.

Observed quantities

mean_rhae 88.49 · public_games 25 · policies_compared 4

01

Primary source

  1. arxiv · 2607.28287 · v1 · abstract
Title
Tycho: Active Abstraction with Programmatic World Models for ARC-AGI-3
Authors
Jens Lehmann, Andrei Aioanei, Sahar Vahdati
Version
v1
Check
Support quote found word for word in the arXiv abstract v1, 2026-10-08
Supporting quote from the source
we compare four orchestration policies on all 25 public games using Claude Opus 4.8 under matched inference budgets. Actor-requested delegation to a model builder obtains the highest observed mean Relative Human Action Efficiency (RHAE), 88.49

02

Boundary of this block

No additional public limitation was recorded.

03

Connections

Outgoing

No outgoing public edges.

Incoming

No incoming public edges.