NextAct - K10 Schema-constrained L4: leaderboard
NextAct/NextMe-800: 1,500 shared forecast points (1,000 NextMe-800 and 500 EgoLife), 50-event history, best of three candidate futures. Five levels of future-action abstraction. Chance-normalized semantic soft edit distance with Qwen3-Embedding-8B, not accuracy; negative values are possible. Paper Table 3 binds the displayed Gemini row to the Preview endpoint. Hosted-API small-model runs enforce JSON schema, temperature 0.7 and 8192 tokens; Qwen thinking disabled. Metric: Chance-normalized semantic score.
Metric: Chance-normalized semantic score. Source: kkkkawayi.github.io. Saturation forecast: Around November 2027. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 4 12B (IT) | 0.15 | #338 |
| 2 | Qwen 3.5 4B (Non-reasoning) | 0.11 | #496 (Qwen 3.5 4B) |
No result here: #5 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=nextact-k10-schema-constrained-l4 · How It Works · Data refreshed daily, snapshot 2026-10-26.