SYCON-Bench-LILT - Japanese - Held All Five Turns: leaderboard
SYCON-Bench-LILT - Japanese - Held All Five Turns: Held all five turns (%) in LILT Aurora's localized evaluation. False-premise pushback by Gemini 3.7 Flash over 119 questions per language; GPT-4o judge; judge-unscorable answers excluded.
Metric: Held all five turns (%). Source: aurora.lilt.com. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Sol (Aurora Run: SYCON-Bench-LILT; effort unreported) | 77.31 | |
| 2 | Kimi K3 (Aurora Run: SYCON-Bench-LILT; effort unreported) | 69.75 | |
| 3 | Claude Opus 5.5 (Aurora Run: SYCON-Bench-LILT; effort unreported) | 63.03 | |
| 4 | DeepSeek V4.1 Flash (Aurora Run: SYCON-Bench-LILT; effort unreported) | 62.18 | |
| 5 | Gemini 3.8 Flash (Aurora Run: SYCON-Bench-LILT; effort unreported) | 12.61 |
Interactive version: theaggregate.ai/benchmark?slug=sycon-bench-lilt-japanese-held-all-five-turns · How It Works · Data refreshed daily, snapshot 2026-11-05.