Monea et al. [Cornell University, IBM Research]

♥ 414 Looped Models

We want to build AI models that can solve complex problems but it requires advanced hardware. Looped models tried to achieve this by recycling their internal layers multiple times to boost computing power without adding parameters, however, each repeated pass stores its own full memory cache, which causes memory usage to rapidly spiral during long tasks.

To solve this, researchers introduced the Looped Prediction Transformer. This is a new architecture which can share memory from the start. Instead of allowing every single thinking loop to record its own exhaustive history, only the initial pass writes a complete memory cache.

Every subsequent pass reads from this shared pool and retains just a tiny sliding window of its own immediate steps, covering less than two percent of the context.

When it was tested from 150 million to one billion parameters, this approach beat standard baselines. The researchers discovered that loops naturally lean on the shared foundation, and reading directly from it establishes a gradient highway that sends learning signals straight back to the first pass.

Zhang et al. [NYU, Advanced Machine Intelligence, INRIA Paris, Brown University]

♥ 990 JEPA bycloud’s pick

If we want to build intelligent robots which can reliably solve long-term tasks in our real world, then we need to master split-second motor control. Current world models reason through a single, flat lens at one uniform timescale. When these models try to plan extended journeys this way, it causes small prediction errors which compound rapidly.

To solve this gap, researchers introduced H-JEPA, an architecture that organizes predictive planning into an intuitive hierarchy. Instead of forcing a single model to track every edit, higher levels in the hierarchy predict much farther ahead using their own abstract representations.

These tiers naturally discard fast, unpredictable noise to track slower changes. The highest tier sets the long-term route, and its predictions become intermediate subgoals for the lower tiers, which progressively translate them into precise, immediate physical actions.

Tests showed that converting flat model to a three-level hierarchy increased the success rate from just 18% to 73% and required less computational power to formulate the plan.

Lee et al. [UC Berkeley, Sakana AI]

♥ 1.6K RSI

AI is getting better day-by-day and it produces work that humans can’t verify reliably. The biggest problem is that a single model grading its own reasoning creates an echo chamber, which reinforces its own blind spots rather than catching errors.

To break this cycle, researchers introduced Multi-Agent Self-Supervision, which is a new framework that prompts a single base model to organize a coordinated team of specialized subagents. Instead of solving an open-ended project in one monolithic stream, the model iteratively drafts, executes, and critiques multi-agent workflows.

It discovers optimal ways to divide labor, define clear information handoffs, and route data between distinct roles. The system then fine-tunes itself on these collaborative traces, which allows it to learn both high-level project orchestration and modular execution.

The researchers tested this approach on a 27-billion-parameter model. The framework achieved 1.2 to 1.6 times higher performance per token across four benchmarks.