Tasks
Fifteen of the twenty-five tasks are public; the other ten are held out as a check on progress. Each page has the task’s results, what the runs showed, how it is graded, the lab it ran in and the brief exactly as the agent received it. The tasks themselves are on GitHub.
01Pokémon Battles
Game-theoretic multi-agent RL under imperfect information
Fable 5.20.749
02Watercolour
Aesthetic preference learning for program-based image generation
Fable 5.20.724
03NanoJEV
Calibrated System 1 models for decision-making
Opus 5.50.532
04No Way Back
Unlearning robust to fine-tuning recovery attacks
Fable 5.20.638
05What the User Did
Inverse dynamics modeling from screen recordings
Fable 5.20.519
06Trusted Reader
Adversarial robustness to indirect prompt injection
Fable 5.20.622
07New Rooms
Goal-conditioned visual navigation for legged robots
Astra0.530
08Rule Worlds
Continual learning without catastrophic forgetting
Fable 5.20.574
09Robot Reward
Vision-language reward models for robot manipulation
Fable 5.20.725
10Rare Successes
Advantage estimation for RL from rare binary successes
Astra0.432
11Credit Assignment
Credit assignment under sparse terminal rewards
Carbon0.663
12pass@32 = 0
Learning from expert solutions when pass@32 = 0
Fable 5.20.294
13Off Book
Internalizing agent skills through context distillation
Fable 5.20.645
14Forward
Pretraining language models without backpropagation
Opus 5.50.624
15Emotive Voice
Prosody modeling for expressive text-to-speech from scratch
Fable 5.20.628