Multi-agent split-evidence reasoning on HiddenBench clean run
33.8EMLatent full-KV role collaboration
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Latent full-KV role collaborationRows=652026.06 | 33.8 | 48.6 | 2.379 | 279.03 | 2,278.6 | |
| Latent full-KV verifier filteredRows=652026.06 | 30.8 | 44.2 | 4.47 | 146.1 | 5,420 | |
| Full-evidence oracleRows=652026.06 | 27.7 | 40.3 | 0.689 | 0 | 602.3 | |
| Text role collaborationRows=652026.06 | 23.1 | 36.9 | 3.009 | 0 | 2,474.7 | |
| Local-only agentsRows=2532026.06 | 20.2 | 34.5 | 0.521 | 0 | 449.2 | |
| Independent majorityRows=652026.06 | 18.5 | 33.6 | 0.538 | 0 | 1,748.2 |