Karten et al. [Princeton University, Prime Intellect, MIT]

♥ 573 Self Improving RL

LLMs often struggle with complex, long-horizon tasks because they work sequentially and have a limited context window. To solve this, researchers built Prime Agent, which is an open-source harness designed to give language models the structured information management of a full computer.

Instead of relying purely on static context windows, Prime Agent provides a persistent coding environment where the model can write and execute code, test intermediate hypotheses, and spin up recursive subagents that coordinate through direct messaging. The framework also incorporates a continual memory system that preserves trajectory histories, saves factual notes, and turns successful multi-step procedures into reusable, versioned skills across sessions.

This flexible architecture allows models to construct their own problem-solving strategies, unlocking significant performance gains using their existing capabilities. In benchmark evaluations, Prime Agent dramatically elevated reasoning scores on the ARC-AGI-3 benchmark from 30% to 95.5%, which outperforms native harnesses across long-context programming and GPU kernel optimization.

Chen et al. [Westlake University, Nanyang Technological University]

♥ 342 World Models bycloud’s pick

Video-based world models try to learn the hidden mechanisms of a world by watching pixels, which makes it difficult to track changes that occur off-screen. To solve this, researchers introduced the Code World Model, a framework that separates a world's logical evolution from its visual rendering.

Instead of forcing a generative video model to deduce all the hidden rules from visual pixels alone, the system uses a language-based coding agent as the world's "brain." This agent handles high-level reasoning and writes executable code to govern the world. The underlying code updates low-level mechanics, such as positions, object boundaries, and causal events, which ensures that actions have lasting, rule-consistent consequences over long horizons.

To turn this running code into actual images without using costly 3D graphics engines, the researchers developed an intermediate layer called a proxy. The proxy compiles the abstract world state into a coarse visual layout that defines frame-by-frame object trajectories and camera motions.

A fine-tuned video model, such as MiniMax-H3, takes this proxy alongside text prompts and fills in rich textures, complex lighting, and natural physical dynamics. By combining the logic of code with modern video models, we can use this approach to create highly detailed virtual worlds.

Kuhn et al. [German Cancer Research Center, German Cancer Consortium, Goethe University Frankfurt, Mila, Université de Montréal, Brown University, Courant Institute, Advanced Machine Intelligence (AMI Labs)]

♥ 1.2k JEPA

LeVJEPA is a streamlined framework that relyies on a single neural network for training AI with video data. Instead of juggling target networks or complex predicting modules, the model simply learns to align cropped, local snippets of a clip with the broader global view.

To make sure that the network does not generate meaningless constants, the system enforces a regularizing constraint called SIGReg. This mechanism shapes the model's internal data representations into an isotropic standard bell-curve distribution, which mathematically guarantees stability with just a single, fixed tuning parameter.

The researchers found that discarding up to 95% of visual patches at random during training did not degrade understanding; instead, it acted as an effective learning stimulus that boosted accuracy while reducing computational costs by up to twentyfold compared to prior methods.

Additionally, the network adopts a causal structure where each video frame is understood purely through current and past observations, incurring no performance penalty compared to systems that look into the future.

This built-in temporal order allows models to process live, incoming video streams smoothly at a constant incremental cost, and even enabling effective pretraining on a single consumer graphics card.

Schmidgall et al. [Harvard University]

♥ 771 Real-World AI

AI can make unverified claims and this makes it unsuitable for for scientific research. To solve this, researchers developed an extended version of Co-Scientist, a multi-agent system designed to serve as an execution-grounded research partner.

The system orchestrates an end-to-end scientific loop: it generates and refines hypotheses, writes and runs experimental code, and compiles findings into complete research manuscripts.

To ensure scientific integrity, the architecture cross-checks every quantitative claim in its writing directly against raw machine execution logs, which prevents the fabrication of positive outcomes while using continuous ethical screening to filter out hazardous experimental plans.

When this system was deployed in real-world environments, it produced validated results across several disciplines. In materials science, it identified a safe, non-toxic chemical pathway to synthesize novel two-dimensional layered materials using chemical vapor deposition. It also made sure to tailor growth recipes directly to custom hardware to achieve single-attempt crystal growth of multiple semiconductors.

In biology, the system used sparse imaging data to accurately forecast how colonies of engineered bacteria swarm and change shape across different chemical concentrations, closely matching physical wet-lab measurements.