JP-TL-Bench - English to Japanese - Hard: leaderboard
Hard English-to-Japanese items of JP-TL-Bench (Lin and Lensenmayer, Shisa.AI, 2026): 19 English passages including fiction, game dialogue, essays and technical how-tos, where register and nuance decide quality, translated into Japanese and compared pairwise and reference-free against the frozen Base Set v1.0 anchor translations by gemini-2.5-flash. Bradley-Terry strength fitted on this slice alone, mapped to a 0-10 LT score; higher is better.
Metric: LT Score (0-10; Hard items only; anchored pairwise Bradley-Terry, gemini-2.5-flash judge). Source: arxiv.org. Saturation forecast: Estimated already saturated. 64 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 10 | #141 |
| 2 | Gemini 2.5 Pro | 9.99 | #262 |
| 3 | Gemini 3 Pro (Preview) | 9.99 | #111 |
| 4 | Gemini 2.5 Flash | 9.98 | #452 |
| 5 | DeepSeek V3.1 Terminus | 9.94 | #442 |
| 6 | GPT-OSS-120B | 9.91 | #699 |
| 7 | Kimi K2 0905 | 9.9 | #555 |
| 8 | Qwen 3 235B A22B 2507 Instruct | 9.84 | #619 |
| 9 | Qwen 3 30B A3B 2507 Instruct | 9.72 | #1078 |
| 10 | Ministral-3-14B-Instruct-2512 | 9.71 | #1595 |
| 11 | GPT-4o (2024-08-06) | 9.68 | #731 |
| 12 | Gemma 3 27B (IT) | 9.66 | #1306 |
| 13 | ELYZA-Shortcut-1.0-Qwen-32B | 9.26 | #1420 |
| 14 | ABEJA-Qwen2.5-32B-Japanese-v1.0 | 9.08 | #1304 |
| 15 | Llama-3.3-Swallow-70B-Instruct-v0.4 | 9.03 | #1312 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=jp-tl-bench-english-to-japanese-hard · How It Works · Data refreshed daily, snapshot 2026-11-12.