# ARC-AGI-3

ARC-AGI-3 asks agents to clear uninstructed interactive games with as few actions as possible, and handing a reasoning model the board as a raw integer matrix makes it collide with walls repeatedly. Sarbloh computes geometric facts such as reachability, enclosure, midpoint tethers and dihedral symmetry in host code with no learned parameters, keeps role labels as guesses each carrying a one-action test, and lets the agent commit multi-step paths from inside its Python kernel via await play(). On the official hidden set its score rose from an initial baseline of 0.27 to 16.53; the source describes that gain as a 13.7 efficiency improvement over the first Milestone winner, while the same table gives that winner a score of 1.21.

ARC-AGI-3 asks agents to clear uninstructed interactive games with as few actions as possible, and handing a reasoning model the board as a raw integer matrix makes it collide with walls repeatedly. Sarbloh computes geometric facts such as reachability, enclosure, midpoint tethers and dihedral symmetry in host code with no learned parameters, keeps role labels as guesses each carrying a one-action test, and lets the agent commit multi-step paths from inside its Python kernel via await play(). On the official hidden set its score rose from an initial baseline of 0.27 to 16.53; the source describes that gain as a 13.7 efficiency improvement over the first Milestone winner, while the same table gives that winner a score of 1.21.

The authors introduce a measurement protocol that traces, from the scripts and notes a coding agent writes alone, how knowledge is kept, rewritten or discarded at ARC-AGI-3 level boundaries; across seven runs with three backbones from two model families, scripts are almost never called again in a later task (33 of 630 references cross a boundary), 74% of pre-boundary scripts are abandoned, notes only grow, and the costliest error is a hard-coded value carried into a new task, whose one working repair is to turn the constant into a parameter supplied afresh per task.

The authors introduce a measurement protocol that traces, from the scripts and notes a coding agent writes alone, how knowledge is kept, rewritten or discarded at ARC-AGI-3 level boundaries; across seven runs with three backbones from two model families, scripts are almost never called again in a later task (33 of 630 references cross a boundary), 74% of pre-boundary scripts are abandoned, notes only grow, and the costliest error is a hard-coded value carried into a new task, whose one working repair is to turn the constant into a parameter supplied afresh per task.

Memento 3 introduces Code as Model: a frozen LLM agent writes its hypotheses about an environment into a natural-language rulebook, compiles that rulebook into executable code for prediction and planning, and revises both through a loop of observation, reflection, rule revision, compilation, and verification, accepting an update only when the LLM judges the code faithful to the rulebook and cell-exact replay reproduces every recorded transition. On ARC-AGI-3 the single-model agent clears every level of all public games, reaches the mean RHAE ceiling, and uses 44% of the human action count; in an Atari Pong case study a learned feedback controller wins all three evaluated episodes with different openings without further LLM calls.

Memento 3 introduces Code as Model: a frozen LLM agent writes its hypotheses about an environment into a natural-language rulebook, compiles that rulebook into executable code for prediction and planning, and revises both through a loop of observation, reflection, rule revision, compilation, and verification, accepting an update only when the LLM judges the code faithful to the rulebook and cell-exact replay reproduces every recorded transition. On ARC-AGI-3 the single-model agent clears every level of all public games, reaches the mean RHAE ceiling, and uses 44% of the human action count; in an Atari Pong case study a learned feedback controller wins all three evaluated episodes with different openings without further LLM calls.

The authors introduce VISTA, a visual harness that gives a general-purpose multimodal model long-horizon vision by letting it perceive the environment through visual observations, keep a lossless visual memory of past observations in their original form, and actively retrieve and reorganize its visual input while reasoning; on ARC-AGI-3 it raises Claude Opus 5.0's Relative Human Action Efficiency from 40.68 to a perfect 100.00, completing all 25 public games with 57.4% fewer actions than first-time human participants, and it substantially outperforms baselines using the same underlying model with minimal harnesses across three additional benchmarks covering diverse visual games and puzzles.

The authors introduce VISTA, a visual harness that gives a general-purpose multimodal model long-horizon vision by letting it perceive the environment through visual observations, keep a lossless visual memory of past observations in their original form, and actively retrieve and reorganize its visual input while reasoning; on ARC-AGI-3 it raises Claude Opus 5.0's Relative Human Action Efficiency from 40.68 to a perfect 100.00, completing all 25 public games with 57.4% fewer actions than first-time human participants, and it substantially outperforms baselines using the same underlying model with minimal harnesses across three additional benchmarks covering diverse visual games and puzzles.