← @bgSponge

Dream-RSI: Recursive Self-Improvement through Evolving Worlds

Summary

Dream-RSI uses accumulated discovery history as a replay simulator to enable rapid, low-cost offline evaluation and refinement of autonomous AI exploration policies.

  1. Dreaming enables rapid policy refinement.

    By converting historical discovery trees into replayable simulators, the system avoids expensive online rollouts during the offline dreaming phase.

  2. Accumulated history scales performance.

    Using the full discovery history for replay achieves a 1.90 performance score on ConvDiv, compared to 1.59 when restricted to current history.

  3. Semantic guidance can constrain exploration.

    Explicitly providing semantic guidance may limit the agent's search space, potentially leading to lower final performance on tasks like ConvDiv.

  4. Batches must be dynamic portfolios.

    Effective batches prioritize exploration and justified recovery over simple exploitation, using prefix evidence rather than fixed quotas to fill slots.

  5. Dream-RSI discovers high-performance kernels.

    The system produced a Lasso-path solver using 4x register-blocked lazy Gram precomputation and a complete ConvDiv kernel with fused convolution and pooling.

    reduces column loads by 75%

    reduces residual vector loads by 75%

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →