Program synthesis, neuro-symbolic systems and ARC-style generalization

Combine neural intuition with discrete program search or refinement loops to learn new tasks from few examples; ARC Prize tracks progress.

Pattern interpolation generalizes poorly to novel tasks. Chollet's Ndea bets program synthesis, guided by deep learning, is the missing paradigm; ARC Prize 2025 named iterative 'refinement loops' the year's theme. ARC-AGI-3 now tests interactive exploration, modeling, goal-setting and planning.

Where it stands. ARC-AGI-3 went from frontier 0.51% (March) to 62.7% standard and 99.9% adapter-harness for GPT-6 Astra (September), at $19-26k a run. ARC Prize 2026 ($2M+) is open.

Evidence

For

ARC Prize 2025 (2025-12-05): 1,455 teams; top Kaggle score 24% on ARC-AGI-2 from NVARC, combining test-time-trained models with TRM components.
Poetiq's Gemini 3 Pro refinement harness scored 54% on ARC-AGI-2 versus 37.6% for Claude Opus 4.5 Thinking (ARC Prize, Dec 2025).
CompressARC, 76K parameters trained only at test time with no pretraining, won third paper prize in 2025; TRM won first.
GPT-6 Astra (2026-09-03): 62.7% on ARC-AGI-3 semi-private ($26k, standard harness) and 99.9% ($19k, provider harness); fewer actions than the median human on 96% of levels.
Astra converted unseen game mechanics into compact symbolic notes it invented, a program-like world model built in context (ARC Prize).

Against

ARC Prize says saturating ARC-AGI-3 would not prove AGI: the games are deterministic, closed-ended and tightly bounded.
Inference: the jump from 0.43% (GPT-5.5, 2026-05-01) to 62.7% (Astra) came from a general frontier LLM with large pretraining and RL, not from program-synthesis systems.
Results depend on harness and cost: 62.7% vs 99.9% across harnesses, at $19-26k per full run, while ARC pays humans ~$12.78 per attempted game.
Sources differ at launch: Fortune cites 66% on the standard harness, while ARC Prize reports 62.7% on its semi-private set (2026-09-03); OpenAI's own launch page gives only the 99.9% figure.
HRM's ARC result came mostly from an outer refinement loop and task-specific training, not its hierarchical design (ARC Prize, 2025-08-15).
Not found: a frontier lab shipping a program-synthesis system as its main product; Ndea has published no frontier results.

Milestones

GPT-6 Astra scores 62.7% (standard) and 99.9% (provider harness) on ARC-AGI-3 OpenAI / ARC Prize Foundation · 3 September 2026
ARC Prize 2026 ARC-AGI-3 Milestone Prize #1 awarded ARC Prize Foundation · 6 July 2026
ARC analysis shows GPT-5.5 scores 0.43% and Opus 4.7 0.18% on ARC-AGI-3 ARC Prize Foundation · 1 May 2026
ARC-AGI-3 launched (interactive games); frontier AI 0.51% vs humans 100%; $2M+ prizes ARC Prize Foundation · 25 March 2026
ARC Prize 2025 results show refinement loops as the theme, and NVARC tops Kaggle at 24% ARC Prize Foundation · 5 December 2025
Tiny Recursive Model has 7M parameters and scores 45% on ARC-AGI-1 and 8% on ARC-AGI-2 Samsung SAIL Montreal · 6 October 2025
ARC Prize replicates and dissects HRM: outer refinement loop drives the score ARC Prize Foundation · 15 August 2025
Hierarchical Reasoning Model (HRM) paper, 27M parameters Sapient Intelligence · 26 June 2025
ARC-AGI-2 and ARC Prize 2025 announced ARC Prize Foundation · 24 March 2025

Who is working on it

Francois Chollet, Ndea / ARC Prize Foundation
Greg Kamradt, ARC Prize Foundation
Poetiq team (refinement harnesses), Poetiq
Gary Marcus (neurosymbolic advocate), Marcus on AI
Alexia Jolicoeur-Martineau (TRM), Samsung SAIL Montreal

The labs with the most milestones here are ARC Prize Foundation (6), Samsung SAIL Montreal (1), Sapient Intelligence (1) and OpenAI (1).

Sources

This research bet was checked and corrected against its sources on 6 October 2026. How we check