Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

Tycho: Active Abstraction with Programmatic World Models for ARC-AGI-3

Source

arXiv

arxiv.org/abs/2607.28287v1

Read the source

Source

Plain-Language Brief

The research paper introduces Tycho, a system that helps an agent interact with and understand unfamiliar games by constructing game-specific models during the gameplay. The system separates actionable observations from intermediate frames and compares four orchestration policies on all 25 public games using Claude Opus 4.8 and GPT-5.6 Sol. The policy that deploys actor-requested delegation to a model builder yields the highest mean Relative Human Action Efficiency (RHAE) of 88.49.

What the Document Says

  • ARC-AGI-3 focuses on interactive abstraction as a skill acquisition problem in games.
  • Games are formalized as parameterized rendered deterministic Moore machines.
  • Tycho is an introduced coding-agent system that constructs game-specific models during interaction.
  • Tycho separates actionable observations from intermediate frames.
  • The agent can model, test, plan with, repair, or bypass a free-form executable hypothesis from the structured history.
  • Four orchestration policies are compared on all 25 public games using Claude Opus 4.8 and GPT-5.6 Sol under matched inference budgets.
  • Actor-requested delegation to a model builder obtains the highest observed mean RHAE, 88.49.
  • With the selected policy, both GPT-5.6 Sol and Opus 5 reach 100.00 RHAE and complete all 183 levels.
  • Their midranks are 98.5 and 100.0 for game-balanced first-run human-replay.
  • Opus 5 uses 61% fewer scored actions than the aggregate official human baselines.
  • Automatic repair after verification failures produces models that reproduce observed transitions more accurately but 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 requires deciding when to construct, repair, use, or bypass a model, referred to as active abstraction.

Key Concepts Mentioned

  • ARC-AGI-3
  • interactive abstraction
  • games
  • Tycho
  • game-specific models
  • orchestration policies
  • Claude Opus 4.8
  • GPT-5.6 Sol
  • Relativ Human Action Efficiency (RHAE)
  • actionable observations
  • actor-requested delegation

Claims Made by the Document

  • Tycho helps an agent understand unfamiliar games by constructing game-specific models.
  • Actor-requested delegation to a model builder yields the highest mean RHAE.

Evidence Used

  • The description of Tycho and its capabilities.
  • The comparison results of four orchestration policies using Claude Opus 4.8 and GPT-5.6 Sol.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

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.