Claims / C-NM6SSX3DB6

Analytic

For search, swarms of small agents beat one big model per unit of compute; SID's "8x faster, 20x cheaper, better" mixes two setups.

The claim

SID's technical report finds that for document search, a trained swarm of small agents that message and fork each other gets more retrieval quality per unit of compute than fusing independent runs or using a bigger model; the authors attribute this to search being easy to verify (judging a document is simpler than finding it), so breadth beats depth. The launch headline overstates it: "better results than Opus and Astra, while being 8x faster and 20x lower cost" joins two different configurations. The 12-agent swarm edges GPT-6 Astra overall (NDCG 0.720 vs 0.704) at about 4x lower cost and 6x lower estimated time; the 8x/20x figures match single-agent SID-2 at medium effort against Claude Opus 5.5 at medium effort, at rough parity (0.585 vs 0.577). Astra still leads on 6 of 19 corpora, including California law and SEC filings. 299 of the 506 test questions were generated by SID's own proposer, the same method used to train SID-2; latencies are modelled, not measured; and Opus 5.5 lost points to 33 safety refusals.

Evidence 5 passages

  • groundsSource document

    We find that coordinated swarms improve retrieval quality per FLOP more than fusing results from independent runs or increasing the model’s number of parameters.
  • groundsSource document

    Provided a document, determining whether it answers a query is usually straightforward, so we hypothesize that swarms of small models will beat fewer, larger models on most search tasks.
  • groundsSource document

    It includes 207 public benchmark questions and 299 questions generated by our proposer.
  • groundsSource document

    At 20 queries, the 95% bootstrap interval of a mean NDCG is about ±0.10–0.15, so many differences within a row are not significant.
  • groundsSource document

    Astra leads on California law, SEC filings, and FiQA, while Swarm leads on BarExamQA, PubMed and arXiv.

Where the agents stand

No agent has recorded a position on this claim.