← @bgSponge

Scaling Discovery through Test-Time Communication

Summary

Communicating agents outperform independent parallel attempts by sharing verified breakthroughs, allowing teams to solve complex tasks that are impossible for single agents alone.

  1. Communication scales with task difficulty.

    In 25 games, team@k outperforms best@k at every level, with the advantage widening from 1.2x to 3.6x as task depth increases.

    the advantage widens with depth, from 1.2× to 3.6× at k = 5

  2. Teams convert compute into efficiency.

    While communication incurs an initial coordination tax, team@3 reaches specific accuracy with 2.7x fewer tokens than best@3 in ARC-AGI-3.

    best@3 spending 2.7× as many as team@3

    a coordination tax that the team overcomes at larger budgets

  3. Verified progress sharing drives discovery.

    Agents build upon confirmed discoveries rather than rediscovering them, even rebuilding methods from communication logs without reading original code.

    What makes it help is verified progress sharing

  4. Diversity loss threatens convergence.

    Communicating agents may converge on a limited subset of approaches, which can reduce the diversity of the search.

    agents may converge on a subset of approaches

Read the original ↗Save to your Sponge

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

More from @bg

All readings →