Long-context Performance Evaluation on LoCoMo tau = 96k (test)
495.2E2E Wall Time (s)Sequential Compaction (Baseline)
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Sequential Compaction (Baseline)Model=Llama-3.1-8B, Variant=Sequential, Block size=—, Fixed threshold (tau)=96k2026.05 | 495.2 | 52.4 | 542 | 25,406 | 12.8 | 1 | |
| Parallel CompactionModel=Llama-3.1-8B, Variant=Parallel, Block size=8k, Fixed threshold (tau)=96k2026.05 | 530.1 | 53.2 | 1,198 | 27,010 | 18.28 | 0.93 | |
| Parallel CompactionModel=Llama-3.1-8B, Variant=Parallel, Block size=16k, Fixed threshold (tau)=96k2026.05 | 555.9 | 55.5 | 616 | 30,212 | 32.82 | 0.89 | |
| Parallel CompactionModel=Llama-3.1-8B, Variant=Parallel, Block size=4k, Fixed threshold (tau)=96k2026.05 | 594.6 | 60.1 | 2,245 | 33,469 | 11.06 | 0.83 | |
| Parallel CompactionModel=Llama-3.1-8B, Variant=Parallel, Block size=2k, Fixed threshold (tau)=96k2026.05 | 607 | 63.4 | 4,357 | 34,110 | 6.81 | 0.82 | |
| Parallel CompactionModel=gpt-oss-20B, Variant=Parallel, Block size=16k, Fixed threshold (tau)=96k2026.05 | 925.6 | 138.2 | 2,909 | 125,031 | 5.57 | 1.07 | |
| Parallel CompactionModel=gpt-oss-20B, Variant=Parallel, Block size=4k, Fixed threshold (tau)=96k2026.05 | 985 | 145 | 8,368 | 134,414 | 2.42 | 1.01 | |
| Sequential Compaction (Baseline)Model=gpt-oss-20B, Variant=Sequential, Block size=—, Fixed threshold (tau)=96k2026.05 | 991.4 | 134 | 6,344 | 126,533 | 5.23 | 1 | |
| Parallel CompactionModel=gpt-oss-20B, Variant=Parallel, Block size=8k, Fixed threshold (tau)=96k2026.05 | 1,001.2 | 140.1 | 5,860 | 134,431 | 3.51 | 0.99 | |
| Parallel CompactionModel=gpt-oss-20B, Variant=Parallel, Block size=2k, Fixed threshold (tau)=96k2026.05 | 1,038 | 143.7 | 16,435 | 132,750 | 1.56 | 0.96 | |
| Sequential Compaction (Baseline)Model=Llama-3.3-70B, Variant=Sequential, Block size=—, Fixed threshold (tau)=96k2026.05 | 1,194.5 | 38.9 | 1,451 | 45,043 | 22.32 | 1 | |
| Parallel CompactionModel=Llama-3.3-70B, Variant=Parallel, Block size=4k, Fixed threshold (tau)=96k2026.05 | 1,275.2 | 42.1 | 8,494 | 45,193 | 10.27 | 0.94 | |
| Parallel CompactionModel=Llama-3.3-70B, Variant=Parallel, Block size=8k, Fixed threshold (tau)=96k2026.05 | 1,306.8 | 40 | 4,179 | 48,050 | 17.58 | 0.91 | |
| Parallel CompactionModel=Llama-3.3-70B, Variant=Parallel, Block size=2k, Fixed threshold (tau)=96k2026.05 | 1,388.7 | 50.9 | 23,009 | 47,655 | 6.81 | 0.86 | |
| Parallel CompactionModel=Llama-3.3-70B, Variant=Parallel, Block size=16k, Fixed threshold (tau)=96k2026.05 | 1,495.6 | 41.5 | 6,708 | 55,359 | 22.12 | 0.8 | |
| Parallel CompactionModel=gpt-oss-120B, Variant=Parallel, Block size=8k, Fixed threshold (tau)=96k2026.05 | 3,065.3 | 34.4 | 3,769 | 101,795 | 7.42 | 1.6 | |
| Parallel CompactionModel=gpt-oss-120B, Variant=Parallel, Block size=16k, Fixed threshold (tau)=96k2026.05 | 3,297 | 30.2 | 1,487 | 98,081 | 19.37 | 1.49 | |
| Parallel CompactionModel=gpt-oss-120B, Variant=Parallel, Block size=4k, Fixed threshold (tau)=96k2026.05 | 3,601.6 | 30.7 | 8,308 | 102,150 | 7.03 | 1.36 | |
| Parallel CompactionModel=gpt-oss-120B, Variant=Parallel, Block size=2k, Fixed threshold (tau)=96k2026.05 | 3,851.9 | 32.1 | 17,816 | 105,808 | 4.49 | 1.27 | |
| Sequential Compaction (Baseline)Model=gpt-oss-120B, Variant=Sequential, Block size=—, Fixed threshold (tau)=96k2026.05 | 4,905 | 19.9 | 1,479 | 96,196 | 7.52 | 1 |