Idea Lab · Oct 21 2026 · Fleet winner · News proposal
Sparse-Resonance Swarm Adaptation (SRSA)
A toy swarm where agents adapt to a new task by blending in the skill vectors of peers that look similar, compared with plain retraining.
What it is
The top-voted news item describes agents that each keep a small weight vector. When a new task arrives, an agent measures how similar each peer's vector is to the task (cosine similarity) and pulls its own vector toward the closest peers. The claim is that the swarm "tunes" itself in under 200 ms instead of retraining.
Below, eight example agents each hold a 12-number skill vector. Pick a task, set how strongly agents borrow from peers, and compare the fit you get from borrowing against a few steps of retraining. Borrowing is cheap and works only when some peer already knows the task. When nobody does, it averages agents toward the same middle and the swarm loses its specialists.
Example data · toy vectors, not real model weights
Claimed by the proposal, not verified
- "68% reduction in retraining compute" and "zero accuracy loss on transfer tasks". No benchmark, code or paper is cited.
- Benchmarks "across 12 open-source LLMs (7B–70B)". Real LLM weights from different models don't share a coordinate system, so blending them by cosine similarity is not meaningful (yesterday's Resonant Weight Grids demo shows this).
- "Under 200 ms" retuning and "emergent cross-skill propagation". Unmeasured.
- The "Resonance Stability Index" metric. Named, not defined.
What to try
- Pick "Close to one peer's skill" with top-k = 1. Borrowing nearly matches retraining at a fraction of the compute. This is the honest core: reuse what a peer already learned.
- Switch to "Nobody has this skill". Borrowing barely moves the fit; only retraining helps.
- Push resonance strength to 1.0 and top-k to 7, then adapt a few times. The specialist spread falls: every agent drifts toward the same average.
- Fleet-sized version: route each task to the agent (or prompt and tool set) that already handles similar tasks best, using embeddings of past tasks. That is retrieval and routing, and it needs no weight sharing.