Tycho: Active Abstraction with Programmatic World M...

Tycho: Active Abstraction with Programmatic World M... | AI Research

Key Takeaways

Tycho is a coding-agent system designed to solve ARC-AGI-3, a benchmark where agents must learn the rules of unfamiliar games through interaction.
ARC-AGI-3 turns abstraction into an interactive problem of skill acquisition.
A player must infer an unfamiliar game's rules, hidden state, and goal while maintaining action efficiency because every move counts.
We formalize these environments as parameterized rendered deterministic Moore machines and introduce Tycho, a coding-agent system that constructs and uses game-specific models during interaction.
Tycho separates actionable observations from intermediate animation, level-completion, and game-over frames.

Paper AbstractExpand

ARC-AGI-3 turns abstraction into an interactive problem of skill acquisition. A player must infer an unfamiliar game's rules, hidden state, and goal while maintaining action efficiency because every move counts. We formalize these environments as parameterized rendered deterministic Moore machines and introduce Tycho, a coding-agent system that constructs and uses game-specific models during interaction. Tycho separates actionable observations from intermediate animation, level-completion, and game-over frames. From this structured history, an agent can model, test, plan with, repair, or bypass a free-form executable hypothesis. In one matched public-set run per policy, 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. With this selected policy, GPT-5.6 Sol and Opus 5 both reach 100.00 RHAE and complete all 183 levels. Their game-balanced first-run human-replay midranks are 98.5 and 100.0. Opus 5 uses 61% fewer scored actions than the aggregate official human baselines. Automatic repair after verification failures produces models that reproduce observed transitions much more accurately, yet reaches only 83.07 RHAE. Transition match indicates whether a simulator reproduces observed dynamics, not whether it has identified the objective or improves the next action. Strong play also requires deciding when to construct, repair, use, or bypass a model. We call this joint problem active abstraction: generating a testable model from costly interaction and deciding when acquiring or using it is worth its cost.

Tycho is a coding-agent system designed to solve ARC-AGI-3, a benchmark where agents must learn the rules of unfamiliar games through interaction. Because every move in these games is scored and limited, the system focuses on "active abstraction"—the process of building, testing, and deciding when to use a programmatic model of the game’s hidden rules to conserve actions.

The Challenge of Interactive Abstraction

ARC-AGI-3 presents an agent with a sequence of levels where no rules or goals are provided. The agent must infer the game's mechanics, hidden state, and objectives while minimizing the number of actions taken. The researchers, Jens Lehmann, Andrei Aioanei, and Sahar Vahdati, formalize these environments as "rendered deterministic Moore machines." In this framework, the agent observes a grid, chooses an action, and receives a new state, with some frames representing transient animations that provide clues but do not allow for direct action.

How Tycho Works

Tycho functions as a coding-agent that constructs "executable hypotheses" in Python. These models are not mandatory; instead, the system allows the agent to decide whether to build, repair, use, or bypass a model based on the current interaction history. The architecture separates the agent’s reasoning from the model-building process, allowing the agent to:

Simulate potential outcomes before committing to a scored action.
Verify hypotheses against past transitions to ensure the model matches observed dynamics.
Refine models when discrepancies arise between the simulation and the actual game environment.

Performance and Results

In a matched study using Claude Opus 4.8, the researchers compared four orchestration policies. They found that "actor-requested delegation"—where the agent decides when to ask for a model to be built—achieved the highest Relative Human Action Efficiency (RHAE) of 88.49.
When testing with frontier models, GPT-5.6 Sol and Opus 5 both reached 100.00 RHAE, successfully completing all 183 levels of the public ARC-AGI-3 set. Notably, Opus 5 completed these levels using 61% fewer actions than the aggregate official human baselines.

Insights on Model Accuracy

The research highlights a distinction between a model’s accuracy and its utility. The team observed that automatically repairing models after every verification failure resulted in higher transition accuracy but lower overall game-playing performance (83.07 RHAE).
Franklin analysis: This suggests that simply having a simulator that reproduces game dynamics is not sufficient for success. The evidence indicates that "active abstraction" requires the agent to possess the metareasoning capability to determine when a model is worth the cost of construction and when it is better to rely on direct reasoning. A model that perfectly mimics the game but fails to identify the objective or guide the next optimal action provides less value than a more selective approach to modeling.