The essential ARC-AGI links, in one place.

2025 top score · 1st place · NVARC · 2025-12-05

• The team pivoted toward synthetic data after early reasoning-model experiments could not run within the offline time and hardware limits.
• Synthetic puzzles were generated from human H-ARC descriptions, with higher-cost code generation performed offline and distilled into a small submission model.
• Test-time fine-tuning was decisive: the TRM component moved from effectively zero to roughly ten percent in a constrained two-hour run.
• A modest standalone component can still help when its errors differ from the stronger ensemble member.
• Daily validation, bug fixing, and resistance to leaderboard overfitting were part of the method—not merely competition hygiene.

Evidence basis: Full YouTube transcript, reviewed 2026-07-29.

2025 top score · 2nd place · the ARChitects · 2025-12-05

• The 2025 system moved beyond the team's 2024 autoregressive, text-like grid formulation.
• Two-dimensional positional structure is part of the representation, not something the model must rediscover from a flattened sequence.
• Masked diffusion supports iterative answer construction rather than a single left-to-right commitment.
• Recursive refinement and perspective scoring turn generation into a propose, inspect, and repair loop.

Evidence basis: Verified ARC Prize video synopsis.

2025 top score · 3rd place · MindsAI · Tufa Labs · 2025-12-05

• Test-time fine-tuning remains the central adaptation mechanism.
• Augmentation ensembles expose the solver to multiple equivalent views of the same transformation.
• Tokenizer dropout reduces dependence on one brittle encoding of grids.
• Pretraining choices and inference-time machinery must be designed as one system.

Evidence basis: Verified ARC Prize video synopsis.

2025 paper award · 1st place · Tiny Recursive Model · 2025-12-05

• TRM uses roughly seven million parameters in a single recursive network.
• It maintains distinct latent-reasoning and answer states.
• Deep supervision teaches intermediate refinements rather than rewarding only the final output.
• Its ARC-AGI-2 drop, despite strong ARC-AGI-1 performance, helps locate the remaining difficulty in task generalization rather than model size alone.

Evidence basis: Verified ARC Prize video synopsis and on-page YouTube summary.

2025 paper award · 2nd place · SOAR · 2025-12-05

• Evolution supplies an explicit test-and-select loop over candidate programs.
• The model learns from its own search traces rather than relying on a hand-authored ARC DSL.
• Search and learning alternate: execution produces feedback, and feedback reshapes later proposals.
• The reported ARC-AGI-1 result reaches up to 52 percent without human solution datasets.

Evidence basis: Verified ARC Prize video synopsis.

2025 paper award · 3rd place · CompressARC · 2025-12-05

• Learning occurs from scratch on one puzzle rather than through a large pretrained prior.
• Minimum description length ranks hypotheses by how compactly they explain the observed transformation.
• Compression acts as an inductive bias against arbitrary memorization.
• The approach is evidence that meaningful ARC performance can emerge from task-local learning alone, although ARC-AGI-2 remains difficult.

Evidence basis: Verified ARC Prize video synopsis and on-page YouTube summary.

ARC-AGI-3 preview · StochasticGoose · Tufa Labs · 2026-07-01

• A useful preview heuristic—favor actions that change the frame—can confuse novelty with progress.
• Action-efficiency scoring and unseen games expose policies that explore broadly without learning which changes matter.
• Priors are necessary to bootstrap learning, but a strong prior can become a trap when the environment violates it.
• An agent needs to revise goals as well as tactics; optimizing the wrong inferred objective creates a self-reinforcing failure loop.
• The deeper target is a compact causal world model that supports counterfactual planning, not a larger store of remembered frames.

Evidence basis: Verified chaptered description and on-page YouTube summary; the interview discloses Tufa Labs as an MLST sponsor.