This tiny AI startup just 鎮圧するd Google’s Gemini 3 on a 重要な推論する／理由ing 実験（する） ― here's what we know

Amanda Caswell

Mon, December 8, 2025 at 4:51 PM UTC

When you buy through links on our articles, 未来 and its syndication partners may earn a （売買）手数料，委託（する）／委員会／権限.

Adobe Firefly image of superintelligence. — Credit: Adobe Firefly/未来 AI

Since Gemini 3 made its debut, it has 首尾よく held the 最高の，を越す位置／汚点／見つけ出す on the LMArena leaderboard. This leaderboard is a crowdsourced 最高位の where thousands of real 使用者s compare AI models 長，率いる-to-長，率いる across a wide 範囲 of 仕事s, 投票（する）ing on which 返答 is better. But when it comes to reaching the toughest 推論する／理由ing （判断の）基準s, there's a new kid on the 封鎖する, and it's already pulled ahead of Google ― and it did it without training its own model.

A six-person team startup known as Poetiq says it has taken the 最高の，を越す位置／汚点／見つけ出す on the ARC-AGI-2 半分-私的な実験（する）始める，決める, a 悪名高くも difficult 推論する／理由ing challenge created by AI 研究員 Fran?ois Chollet. The startup’s system 得点する／非難する／２０d 54 パーセント, 辛勝する／優位ing out what Google 以前報告（する）／憶測d for Gemini 3 深い Think at around 45 パーセント.

To put that in 視野, most AI models were stuck under 5 パーセント on this （判断の）基準 just six months ago. 割れ目ing 50 パーセント is something 研究員s 広範囲にわたって assumed was years away.

宣伝

And the most surprising part: Poetiq’s 打開 wasn’t 力／強力にするd by a new frontier model ― but by a smarter way of orchestrating 存在するing ones.

How Poetiq pulled this off

Instead of building a massiv e transformer from scratch, Poetiq developed what it calls a meta-system; essentially an AI 監査役 that 監督するs, critiques and 改善するs the 生産（高）s of whatever model you plug into it. For their ARC-AGI-2 work, the team used Gemini 3 プロの／賛成の as the base model.

Poetiq 述べるs the system as a tight optimization 宙返り飛行: 生成する > critique > 精製する > 立証する.

Here’s what makes it stand out:

No retraining 要求するd: The system adapts to new models within hours
Built 完全に on off-the-shelf LLMs: No custom 罰金-tuning
Lower cost: Google’s 深い Think 報道によれば costs ~$77 per 仕事; Poetiq’s system ran closer to $30
Open source: The solver is public and inspectable
Self-auditing: The system 評価するs its own answers before returning a final result

On the company website, Poetiq’s team says the approach 作品 by squeezing more 推論する／理由ing 力／強力にする out of 存在するing LLMs ― not by 規模ing brute-軍隊計算する.

Why ARC-AGI-2 事柄s

Artificial intelligence concept image — Credit: Shutterstock

While most （判断の）基準s 手段狭くする技術s like coding or math, ARC-AGI-2 is designed to 実験（する） something deeper: pattern 承認, analogy, abstract 推論する／理由ing, and the 肉親，親類d of generalization humans learn in 早期に childhood.

宣伝

It’s 故意に hard and famously unfriendly to today’s LLMs. Even many frontier models fail spectacularly.

That’s why the leap from 選び出す／独身-digit 得点する／非難する／２０s to 54 パーセント in half a year has turned 長，率いるs. It 示唆するs 進歩 in 推論する／理由ing methods, not just raw model 規模.

However, Poetiq’s result 適用するs 特に to the 半分-私的な実験（する）始める，決める, which is not fully open to the public. The company 場所／位置 says the result has been 立証するd by the （判断の）基準’s 組織者s ― but 独立した・無所属 third-party replication is still 未解決の, which is important for a （判断の）基準 this 影響力のある.

Perhaps the next 打開 won’t come from bigger models as Poetiq’s work 最高潮の場面s a growing 傾向 in AI: 進歩 doesn’t always 要求する billion-dollar 組織／基盤／下部構造 or a 抱擁する研究 lab.

宣伝

If systems like this generalize beyond （判断の）基準s, to planning, coding, 研究 or real-world 決定／判定勝ち（する）-making, it could reshape how AI is developed. Instead of waiting for the next 打開 model, companies might build 層d 知能 that makes today’s models smarter, cheaper and more 一貫した.

底（に届く） line

Poetiq has open-sourced its ARC-AGI solver so 研究員s can 実験（する）, 延長する or challenge the results. The （判断の）基準 has a hidden 実験（する）始める，決める, and history shows results can 転換 once more people run 独立した・無所属 evaluations.

If Poetiq’s numbers 持つ／拘留する, this could 示す a turning point in AI 推論する／理由ing 研究. A six-person team may have just shown that orchestrating models can 競争相手, or even （警官の）巡回区域，受持ち区域, training bigger ones. Poetiq just 証明するd you don’t need a 巨大（な） lab to 勝利，勝つ a 一連の会議、交渉／完成する.

News

Life

Entertainment

財政／金融

Sports

New on Yahoo

傾向ing:

This tiny AI startup just 鎮圧するd Google’s Gemini 3 on a 重要な推論する／理由ing 実験（する） ― here's what we know

How Poetiq pulled this off

Why ARC-AGI-2 事柄s

底（に届く） line

More from Tom's Guide