Sahaj.blogs

Day page

10 Sept 2026

Paper summary

Paper summary: Scaling Automatic Research Agents via World Models

RL for automatic-research agents is bottlenecked by slow, sandboxed experiment execution; World Model RL replaces execution with a learned world model plus debiasing and denoising, training 3–4× faster.

  • Generating solutions shares compute through batching, but running each experiment takes its own machine time, so execution becomes the bottleneck.
  • Because the world model's rewards are biased and noisy, two corrections are added: online debiasing and inverse-variance denoising.

Read the post →

Study

Agentic AI · Agent memory

Short-term memory (the conversation in the context window) versus long-term memory (facts stored outside and retrieved when relevant), with a small long-term store built on embeddings.

  • Short-term memory is just the message history; long-term memory is a searchable store the agent writes to and reads from.
  • Retrieval-based memory means the agent only pulls in what's relevant now, instead of everything it ever saw.

Read the post →