TL;DR
ArcticSwarm is a novel multi-agent architecture designed for long-horizon web research that prevents "premature consensus"—the tendency of AI agents to copy each other's early, incorrect guesses. By isolating agents during evidence gathering and enforcing rigorous, multi-stage reviews before committing to an answer, it significantly outperforms existing deep-research baselines, achieving up to 88.3% accuracy on complex research benchmarks.
What Problem Does This Paper Address?
Multi-agent systems work beautifully when there is an automated way to check their work, such as running unit tests on generated code. But what happens when you ask a swarm of AI agents to conduct open-ended web research where no simple "verifier" exists?
Typically, developers rely on majority voting or allow agents to chat freely to reach a consensus. However, the authors identify a critical failure mode in this setup: premature consensus. When agents share partial findings too early, they anchor on the same initial (and often incorrect) hypothesis. This collapses the diversity of the search space; agents stop exploring alternative angles and collectively fail along the same path.
Methods
The authors introduce ArcticSwarm, an architecture that decouples evidence gathering from evidence integration using two main pillars:
- Gated Bulletin Board System (BBS): Instead of free-flowing peer-to-peer communication, all agent interactions are routed through a centralized BBS. Crucially, the orchestrator can dynamically assign agents to Isolation Mode (where they can write findings to the board but cannot read peer posts) or Collaboration Mode (where they can read and write). This ensures agents explore independent search paths without being biased by early group-think.
- Staged Review at Commitment Boundaries: Before any hypothesis is accepted, it must pass through three strict gates: an internal Self-Check, a Board Audit by a dedicated auditor agent, and a final Commit Gate. The commit gate requires a "VERIFIED" verdict from both the original investigating agent and an independent reviewer, alongside an explicit sweep for alternative candidates.
Datasets
The architecture was evaluated on two challenging long-horizon web search benchmarks:
- BrowseComp: A dataset containing 1,266 complex, multi-constraint identification questions evaluated against the live web.
- BrowseComp-Plus: An upgraded version containing 830 questions evaluated against a static, pre-defined corpus of ~100K curated web documents to ensure perfect evaluation replicability.
Key Insights and Ideas
- Communication Can Be Harmful: In open-ended research, dense communication actually hurts performance by accelerating convergence and reducing search diversity. Restricting what agents can read during active search forces them to find unique evidence.
- The Answer is Usually Found, Just Not Chosen: A failure analysis revealed that the correct answer is actually present somewhere in the agents' collective browser history in 58% of failed cases. The breakdown happens during the adjudication phase when choosing between candidates.
- Forcing Contrarian Thinking Works: Forcing the orchestrator to launch at least one "alternative/contrarian" search task dramatically boosts accuracy, breaking the candidate fixation loop.
Key Results
- State-of-the-Art Accuracy: ArcticSwarm sets a new benchmark on BrowseComp-Plus, achieving 82.6% accuracy using the open-weight Qwen 3.5-27B model, beating the strong MiroFlow baseline (70.6%).
- Frontier Model Performance: When scaled to GPT-5 on the live-web BrowseComp, ArcticSwarm hits 73.6%, vastly outperforming OpenAI's own deep-research baseline (54.9%) and MiroFlow (63.4%).
- Ablation Proof: Turning off gated isolation dropped accuracy by 3.8 percentage points, and disabling the review gates dropped it by another 4.3 percentage points, proving both components are vital.
Limitations and Future Work
- High Token Cost: The extensive multi-stage review process and independent search paths consume significantly more tokens than simpler architectures. Optimizing token and resource efficiency is left for future work.
- Live-Web Flakiness: Performance on the live web is heavily bottlenecked by upstream infrastructure issues. Gateway authentication errors, rate limits, and safety filters accounted for roughly 91% of subagent task failures.
- Heuristic Orchestration: The policy for when to switch an agent between isolation and collaboration mode is currently heuristic and could be made more dynamic or learned in future iterations.
Summary updated