Explorative modeling: Train on the best of K guesses

5 hours ago

Explorative Modeling (XM) introduces exploration (K guesses) to solve the many-answers problem by modifying the training objective so that only the best match trains.
XM acts as a third pretraining axis alongside parameters and data, monotonically improving existing generative models across images, video, and language with gains growing at scale.
XM achieves 6.2× sample efficiency, 4.1× FLOP efficiency, and 47% better parameter efficiency, and enables end-to-end generation matching diffusion with up to 256× less inference compute.
The concept of generative expressivity (number of distinct answers a model can capture) is introduced and shown to be scalable via exploration, directly impacting performance.
Adding exploration to existing models (e.g., RAE, SiT) improves data efficiency, convergence speed, and generalization, with benefits increasing as models and data scale.
End-to-end XMs match or exceed Diffusion Policy and Diffuser on control tasks with a single forward pass instead of hundreds.

Schema harness enables frontier models to achieve ~99% performance on ARC-AGI-3 Public by treating latent world representations as executable programs, allowing interpretable, verifiable, and searchable modeling.
ARC-AGI-3 challenges agents to infer game rules, objects, and goals from raw 64x64 color grids without explicit instructions, using the Relative Human Action Efficiency (RHAE) metric for evaluation.
The approach combines state grounding (identifying objects and variables from observations) and mechanism discovery (learning transition rules), solved jointly within an editable program that mirrors physicist-like reasoning.

Ring-Zero scales zero reinforcement learning (RL) to one trillion parameters for improved reasoning.
A stable training pipeline with algorithmic and system optimizations addresses challenges like poor readability and redundancy.
Key findings: scaling enhances sample efficiency and performance; training involves discovery and sharpening phases; emergent cognitive behaviors like self-verification develop.

Inkling-Small is a general-purpose multimodal model that accepts text, image, and audio inputs and generates text outputs.
It can be used with libraries like Transformers, vLLM, SGLang, and Docker, as well as through APIs and third-party providers.
The model uses a 42-layer decoder-only transformer with a sparse Mixture-of-Experts (MoE) architecture, totaling 276B parameters but activating only 12B per token.

Transformer Transformer is a unified model that generates a complete robot embodiment (links, joints, motors, inertia) optimized for a given manipulation demonstration.
The model uses RoboTokens, a unified tokenization of robot embodiments, states, and actions, enabling it to span diverse embodiment spaces like wheeled bimanual, quadrupeds, and humanoids.
It is a diffusion transformer trained with DDIM, and by masking different tokens it can act as a generator, cross-embodiment controller, or motion-conditioned designer.

Pantograph develops a method for learning goal-directed behavior through pretraining on internet-scale video, improving ability to achieve complex goals.
The Pan model (4B parameters) in Minecraft shows strong performance in fighting, exploration, building, and handling diverse environments.
Goal-conditioned reinforcement learning uses future video frames as goals, side-stepping the need for explicit reward functions.

Schema harness enables frontier models to achieve ~99% performance on ARC-AGI-3 Public by treating latent world representations as executable programs, allowing interpretable, verifiable, and searchable modeling.
ARC-AGI-3 challenges agents to infer game rules, objects, and goals from raw 64x64 color grids without explicit instructions, using the Relative Human Action Efficiency (RHAE) metric for evaluation.
The approach combines state grounding (identifying objects and variables from observations) and mechanism discovery (learning transition rules), solved jointly within an editable program that mirrors physicist-like reasoning.

Ring-Zero scales zero reinforcement learning (RL) to one trillion parameters for improved reasoning.
A stable training pipeline with algorithmic and system optimizations addresses challenges like poor readability and redundancy.
Key findings: scaling enhances sample efficiency and performance; training involves discovery and sharpening phases; emergent cognitive behaviors like self-verification develop.

Inkling-Small is a general-purpose multimodal model that accepts text, image, and audio inputs and generates text outputs.
It can be used with libraries like Transformers, vLLM, SGLang, and Docker, as well as through APIs and third-party providers.
The model uses a 42-layer decoder-only transformer with a sparse Mixture-of-Experts (MoE) architecture, totaling 276B parameters but activating only 12B per token.

Transformer Transformer is a unified model that generates a complete robot embodiment (links, joints, motors, inertia) optimized for a given manipulation demonstration.
The model uses RoboTokens, a unified tokenization of robot embodiments, states, and actions, enabling it to span diverse embodiment spaces like wheeled bimanual, quadrupeds, and humanoids.
It is a diffusion transformer trained with DDIM, and by masking different tokens it can act as a generator, cross-embodiment controller, or motion-conditioned designer.

Pantograph develops a method for learning goal-directed behavior through pretraining on internet-scale video, improving ability to achieve complex goals.
The Pan model (4B parameters) in Minecraft shows strong performance in fighting, exploration, building, and handling diverse environments.
Goal-conditioned reinforcement learning uses future video frames as goals, side-stepping the need for explicit reward functions.

Hasty Briefsbeta