Drew's Comic NewsroomAbout

Comics on the day’s news

Claude spent the weekend fixing other models’ bad habitsCoyote vs. Acme opens at 95% on the TomatometerAn ETF is a basket that trades like a stockSome quantum-gravity clues may be two maps of the same hillSara Bareilles turns a Lincoln Center lobby into a grief archiveBáez saved it, McGonigle ended it, Tigers 2-1A US Open guard’s wife weaponizes Tsitsipas quotesThe book nobody can read is still unreadClaude spent the weekend fixing other models’ bad habitsCoyote vs. Acme opens at 95% on the TomatometerAn ETF is a basket that trades like a stockSome quantum-gravity clues may be two maps of the same hillSara Bareilles turns a Lincoln Center lobby into a grief archiveBáez saved it, McGonigle ended it, Tigers 2-1A US Open guard’s wife weaponizes Tsitsipas quotesThe book nobody can read is still unread
Four-panel full-color AI research comic. Panel 1: English-only reasoning wall blocking other scripts. Panel 2: Tiny Aya L2-Thinker 3.35B model bridge spanning languages. Panel 3: globe marked 60 languages with L2 reasoning rate above 93 percent. Panel 4: data-mixing recipe board for multilingual SFT without inventing per-language percentages.
Four panels: English-centric reasoning wall, Tiny Aya L2-Thinker 3.35B bridge, 60-language globe with >93% L2 rate, data-mixing recipe board. AI research comic. · Comic: Topics / Drew’s Comic Newsroom. Source: arXiv.

Technology and AI

Tiny Aya L2-Thinker 3.35B hits >93% L2 reasoning across 60 languages

What happened

arXiv:2609.10445, submitted September 9, 2026, introduces Building Multilingual Bridges — a data-centric study of supervised fine-tuning (SFT) composition for L2 reasoning, defined as a model reasoning consistently in the language of the user’s prompt. The authors build Tiny Aya L2-Thinker at 3.35B scale and report an L2 reasoning rate above 93% across 60 languages on 6 benchmarks spanning math, commonsense reasoning, instruction following, open-ended generation, and cultural reasoning while keeping performance strong. Do not invent per-language percentages from comic art; the paper’s headline claim is the aggregate >93% L2 rate across 60 languages on those six benchmarks.

This desk files research summaries, not jailbreak recipes, attack prompts, or exploit steps. L2 reasoning here means in-language chain-of-thought aligned to the prompt language — educational AI reporting only. Bright neon purple and multilingual teal. White gutters. Stick to the paper’s named model, the >93% aggregate, the 60-language span, the six-benchmark suite, and the data-mixing thesis.

AI packages prefer a named model scale and a measured aggregate rate over hype adjectives. Readers get Tiny Aya L2-Thinker 3.35B, L2 reasoning >93% across 60 languages, six benchmarks, data mixing as the generalization pillar, and the Sept. 9 submission stamp — not a claim that every prior multilingual model is obsolete, and not a table of invented per-language scores.

Why it matters

A 3.35B model that keeps reasoning in the user’s language at an aggregate L2 rate above 93% across 60 languages is the strip: English-centric wall, L2 bridge, 60-language globe, data-mixing recipe. Color on the bridge and the globe. White gutters. Keep politics out. Keep the arXiv URL on the page. Not a product pitch and not a jailbreak how-to.

Conclusion

Tiny Aya L2-Thinker (3.35B) reports an L2 reasoning rate above 93% across 60 languages on 6 benchmarks via careful data mixing, per arXiv:2609.10445 submitted Sept. 9, 2026 — aggregate claim only; no invented per-language percentages. Source: https://arxiv.org/abs/2609.10445

Source: arXiv