Harmonic Flow Models for ARC-AGI-3

ARC-AGI-3 turns abstract reasoning into interaction. An agent must explore a small visual world, infer its rules from feedback and choose what to do next under an action budget. A correct move is not supplied in advance, so state representation and exploration matter as much as the final answer.

I implemented an HFM-inspired exploration layer inside the official ARC-AGI-3 agent scaffold. It builds canonical keys for observed states, extracts connected components as possible objects, proposes legal coordinate actions and remembers the transitions that follow. This is the original layer; the surrounding runner and toolkit come from the official scaffold.

Actions are scored online using novelty, frontier value, repeated-state penalties and no-op detection. When the agent enters a loop, a stuck detector changes the exploration policy rather than spending the rest of the budget repeating an action that has already failed. Every decision keeps its provenance so a run can be inspected afterwards.

Recorded scorecards reached two levels on the public game ls20 across two seeds with a 120-action cap. That result used an explicitly configured public-game route; it is not evidence of route-free generalisation. Three harder games, m0r0, ft09 and vc33, remained unsolved in the same sweep.

Those failures were useful. They exposed the limits of component heuristics when objects change identity, rewards are delayed or an action only becomes meaningful after a longer sequence. The project therefore includes a claim audit and experiment log alongside the agent, treating negative results as part of the system rather than something to edit out.

A poster describing the work was externally selected and presented at Pre-ICML @ London 2026 at UCL. The project is ongoing; it is an auditable exploration experiment, not a claim to have solved ARC-AGI-3 or built a general ARC solver.