Today's edition highlights systems that optimize their own training and execution environments, from evolving optimizer agents to unsupervised vetting of scientific discoveries.
The loop: A meta-agent rewrites the policy and code library of an optimizer agent based on evidence from earlier training runs. The improved optimizer agent then manages subsequent model training tasks more effectively.
A meta-agent recursively improves the prompts, procedures, and code library of an optimizer agent that monitors and adjusts neural network training dynamics.
The loop: A self-improving science agent uses a Minimum Description Length compression ratio to unsupervisedly vet its generated scientific relations. This filters the agent's discovery ledger to retain true laws, which guides its future research.
A deployed science agent uses compression ratios to automatically distinguish fundamental physical laws from statistical regularities among its own discoveries.
The loop: A model undergoing reinforcement learning post-training uses its own recent rollout history to construct an n-gram drafter. This drafter accelerates the autoregressive generation bottleneck, speeding up the model's own post-training process.
NgramRL accelerates reinforcement learning post-training by using the model's own recent rollouts to build a memory-efficient speculative decoding drafter.
The loop: A self-evolving alpha mining agent records reusable evidence of successful and failed edit motifs in a structured memory. The agent then uses this memory to suppress failure patterns and improve its own future discovery efficiency.
AlphaMemo improves financial alpha mining by recording successful and failed edit motifs in a structured memory to guide its future searches.
The loop: An agent learns from reference trajectories to evolve its own execution harness. The improved harness provides a more reliable environment, making the agent better at subsequent self-evolution tasks.
An agent uses reference trajectories to systematically evolve and improve the reliability of its own execution harness.
The loop: An automated system evolves agent harnesses based on performance feedback. These improved harnesses are then deployed to provide better environments and evaluation frameworks for future agents.
HarnessCompass automatically evolves agent harnesses to create more generalizable and effective environments for agent execution.
The loop: Opus 5.5 acts as an automated interpretability researcher to analyze and identify the function of neurons in GPT-2-XL. These insights into model internals provide a pathway for researchers to better align and develop future models.
A researcher uses Opus 5.5 to automatically interpret the activations of GPT-2-XL neurons, achieving a significant improvement over previous models.
