LEAD-RL Delegation Harness (Swarm) - GAIA: leaderboard

Benchmarks: BrowseComp (full), BrowseComp-ZH, GAIA, HLE with tools, DeepSearchQA; splits/counts other than "full BrowseComp" not stated (§4.1 Benchmarks). Metrics: Final-answer accuracy (%) for BrowseComp, BrowseComp-ZH, GAIA, HLE; mean per-question answer-set F1 (%) for DeepSearchQA (§4.1 Evaluation metrics; S4.T1 caption). Judge: DeepSeek V4 Pro judges the native runs and validation checkpoints with temperature-0 prompts (§4.1). Appendix A describes the final-answer judge used for training rewards: BrowseComp judging protocol (answer extraction, semantic equivalence with small numerical tolerance), temperature 0, 16,384-token completion budget; the paper does not say whether evaluation used the identical prompt (§4.1; Appendix A). Harness: Efficient delegation harness: main agent routes to direct tools or parallel call_sub_agent briefs; subagents run in isolated contexts; local verifier on each report and global verifier on the final answer, both role-conditioned calls of the same model; no reference answer at inference (§3.2). Swarm on/off: S4.T2 compares matched runs with delegation enabled vs disabled under the same budget (S4.T2 caption). Decoding, execution budget, thinking/effort, access route, run count: not stated (only that controlled comparisons share them) (§4.1 Implementation details). Training-only settings not applied to evaluation: GSPO K=4, 32 questions/update, up to 16 subagents per main rollout, KL 0.001 are training hyperparameters (Appendix A Training hyperparameters). This board: Not stated (GAIA split and question count not printed); One score per model; number of runs/seeds not stated; not a mean over printed repeats. Not stated (GAIA split and question count not printed). Metric: Final-answer accuracy (%), judged by DeepSeek V4 Pro (temperature 0) Higher is better.

Metric: Final-answer accuracy (%), judged by DeepSeek V4 Pro (temperature 0). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked. Updated 14 Nov 2026.

Top models

#ModelScoreOverall rank
1DeepSeek V4 Pro97#191
2DeepSeek V4 Flash92.2#259
3Qwen 3.6 35B A3B81.6#473
4Ling-3.0-flash73.1#437

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=lead-rl-delegation-harness-swarm-gaia · How It Works · Data refreshed daily, snapshot 2026-11-16.