HiVe-Mem Study - M2A - WorldMemArena - LJ: leaderboard
HiVe-Mem v1 Table1, M2A, WorldMemArena LJ native0–100. Two unchanged answer backbones,6:2:2train/validation/test,adversarialcasesexcluded,GPT4omini commonjudge. Temperature0,maxoutput512,256tokenhistoryepisodes,Qwen3VLEmbedding2Bretrieval. BaselinesTop7;HiVe5seeds+2expanded,degree4,lambda0. Qwen-trainedselector frozen/transferredtoGPT. ExactGPTsnapshotunreported; no inferred new checkpoint; tokens/latency excluded.
Metric: Score (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5-mini (HiVe-Mem Run: memory=M2A;temperature=0;output=512) | 65.68 | |
| 2 | Qwen3-VL-4B-Instruct (HiVe-Mem Run: memory=M2A;temperature=0;output=512) | 37.73 |
Interactive version: theaggregate.ai/benchmark?slug=hive-mem-study-m2a-worldmemarena-lj · How It Works · Data refreshed daily, snapshot 2026-11-12.