代理系統／評測
Tycho Lets Agents Write World Models During Interaction; Two Frontier Models Complete Every Public ARC-AGI-3 Level
Tycho translates unknown game rules into executable state-transition models, then uses validation results to decide whether to revise or bypass them. GPT-5.6 Sol and Claude Opus 5 completed all 183 levels in one public-set evaluation, though the results still depend heavily on frontier APIs, single runs, and public tasks.
ARC-AGI-3 no longer asks models merely to infer rules from static examples. Instead, it places agents in unfamiliar 64×64 interactive environments, where they must use a limited action budget to discover objectives, hidden states, and game mechanics. The newly released Tycho frames this type of problem as “active abstraction”: the agent records screens, actions, and state changes and, when necessary, delegates to another model to encode its hypotheses as a Python world model containing `State`, `transition`, `render`, and `outcome`. It then validates the model against existing trajectories and uses it to plan the next move.
The team compared four orchestration policies using Claude Opus 4.8. With no world model, relative human action efficiency (RHAE) was 79.07. When the primary agent decided for itself when to invoke the model builder, the score rose to 88.49. A fixed policy that automatically repaired the model after validation failures achieved only 83.07, showing that a simulator’s closer fit to observed transitions does not mean it understands the objective or will recommend a better next action. Using the same orchestration policy, GPT-5.6 Sol and Claude Opus 5 both achieved 100 RHAE across 183 levels in 25 public games; the latter used 61% fewer scored actions than the official aggregate human baseline.
Tycho’s engineering value lies in not treating the world model as a single fixed neural network. Instead, it lets the agent use the value of information to decide when to create, patch, query, or abandon an executable hypothesis. The public repository includes prompts, a planner, a validator, a replay viewer, a container sandbox, and OpenAI and Anthropic transport layers, making complete trajectories easier to inspect.
But this is not evidence that general reasoning has been solved. Policy selection was performed on the public set, the main comparison used a single run for each policy, and the results may be affected by randomness and tuning to public environments. Reproduction also requires costly calls to frontier models. The next questions are how the system performs on the hidden test set and across multiple seeds, what its costs and failure rates are, and whether smaller models can achieve similar gains from the same verifiable world-model framework.