SID-2 Unveils 12-Agent Search Swarm: 0.720 NDCG Outperforms GPT-6 Astra, But the Benchmark Is SID's Own
SID AI has released SID-2, which scores 0.585 NDCG at 3.11s latency for a single agent, and hits 0.720 NDCG with its 12-agent Swarm configuration, outperforming GPT-6 Astra on SID's in-house benchmark. The metrics are notable, but independent third-party validation is still pending.
Reported by Michal Sutter from MarkTechPost, SID AI has launched SID-2, an agentic search model trained with reinforcement learning. Key performance metrics: A single agent achieves 0.585 NDCG in 3.11 seconds; the SID-2 Swarm composed of 12 agents reaches 0.720 NDCG, surpassing GPT-6 Astra on SID's proprietary benchmark.
NDCG is a widely adopted metric for search ranking performance, with a score closer to 1 indicating better performance. While 0.720 is far from perfect, in the agentic search scenario, it demonstrates that multi-agent parallel operation has substantially improved ranking quality. The 3.11s latency for SID-2's single agent is a practically meaningful reference metric. However, the Swarm mode does not publish its latency or token cost. With 12 agents running simultaneously, the ranking quality improves, but the hidden tradeoffs may emerge later.
Agentic search can be roughly understood as a search paradigm where the model does not just return results in a single pass: it autonomously breaks down the query, generates follow-up searches, reads retrieved content, and adjusts its outputs iteratively. SID-2 uses reinforcement learning to train this entire workflow, while the Swarm design enables multiple agents to split tasks or run searches in parallel. This core concept is not entirely new, but the performance jump from 0.585 to 0.720 at least proves that scaling effects have not hit a bottleneck for this task yet.
A healthy dose of skepticism is still warranted. The 0.720 NDCG score was measured on SID's self-developed benchmark, and the original article did not provide details for independent reproduction, full dataset specifications, or the exact comparison configuration for GPT-6 Astra. GPT-6 Astra is used as a well-known reference point, but the fact that the benchmark is owned by SID inherently limits the generalizability of this conclusion. We can fairly assess whether SID-2 is a meaningful step forward for agentic search only after third-party evaluation results or tests on real-world query datasets are released.
Original link: https://www.marktechpost.com/2026/11/17/sid-ai-sid-2-agentic-search-swarm/
发布时间: 2026-11-18 15:10