Today's edition highlights methods for making recursive self-improvement more budget-efficient, alongside novel applications in semantic communications and long-video memory.
The loop: A coding agent proposes and scores edits to its own scaffold, which improves its solve rate. This higher solve rate then guides the selection of the next scaffold edit.
A coding agent uses a linear readout to score proposed edits to its own scaffold before running them, improving its solve rate on held-out instances.
The loop: A large language model modifies its own communication program and training policy. The accepted code and textual feedback then inform the model's subsequent code proposals.
A framework that recursively evolves a semantic communication program through language-model-assisted code edits and textual feedback.
The loop: An automated multi-agent framework formulates memory-policy development as a multi-objective code search. It records past proposals, traces, and failures in a persistent meta-memory to guide its subsequent policy revisions.
An automated multi-agent framework that optimizes memory policies for long-video agents by learning from its own past traces and failures.
The loop: A model acts as its own teacher by re-scoring its trajectories with retrospective context. The resulting teacher distribution is then distilled back into the same policy over multiple rounds.
A multi-round framework where a model acts as its own teacher by re-scoring trajectories with retrospective context and distilling the knowledge back into itself.
The loop: A recursive self-improvement workflow evolves target agents and its own system architecture. This improves the capability of the system to perform further evolution on itself and other agents.
An open-source workflow that evolves both target agents and its own recursive self-improvement system.
The loop: An automated interpretability tool uses the Jacobian lens to evaluate functional global workspace representations in language models. This identifies structural deficiencies that can be addressed in the design of successor architectures.
A critique of Anthropic's global workspace tests showing that even GPT-2 passes most of them, suggesting the need for more rigorous structural evaluations.
The loop: An orchestration model named Fugu dynamically switches between multiple models to manage geopolitical risks and cultural alignment. This improves the reliability of the overall agent system it operates within.
An event report detailing Sakana AI's approach to AI sovereignty, including an orchestration layer that dynamically switches between models.
