しょう | AX

AIの最新情報をお届け | 法人向けAI活用支援MOOMENT代表 | 業務棚卸し•AI実装•運用支援など推進 | 数百名規模の全社AI活用を主導し利用率100%達成 | 上場企業も支援 | 松尾研認定スタートアップにてAI研修講師も兼務

[しょう | AX @SHO_A]

https://t.co/3Kgp9KfOgr

[しょう | AX @SHO_A]

AIが同じ方法を試し続けて、いつまでも終わらない。その改善を狙うAlibabaの研究「PoS」が紹介された。AIが今どんな前提で動いているかを記録し、行き詰まりを検知して条件を調整する。単に会話履歴を残すだけではない、長い作業の管理方法。

[DailyPapers @HuggingPapers]

Alibaba's PoS: Beyond memory for long-horizon agents

An inference-time framework that maintains explicit belief states, detects when agents stall (Belief Trapping), and recovers with targeted constraints. Achieves SOTA across 4 benchmarks, with gains up to 37.89%.

[しょう | AX @SHO_A]

AIに同じ失敗を繰り返させないための研究「SkillRefiner」。過去の作業記録から、失敗しやすい箇所を見つけて手順書を修正する。新しい試行を大量にやり直さず、手元にある記録を使うのが特徴。改善に必要な計算量を減らせるか検証している。

[alphaXiv @askalphaxiv]

Can agents improve their own skills without generating any new rollouts?

This paper introduces SkillRefiner, which learns entirely from historical agent traces. It turns past mistake patterns into targeted edits to the agent’s existing skill, by basically treat deployment history like a bug report database.

Additionally, repeated successful behaviors get reinforced, while repeated failures become new guardrails, with an extra evidence check to avoid learning the wrong lesson.

Across all eight settings, it improved the original skill while using 1.4-13x fewer refinement tokens than the "extracting lessons from each past run individually" (Trace2Skill) method, and 2.4-42x fewer than the "generating new runs just to test whether each skill rewrite works" (GEPA) method.

https://t.co/YhWHRRSe3E

[しょう | AX @SHO_A]

作業中のCanvaの資料に、自分専用のタイマーを置く。Canvaが、AIで休憩時間や見た目を指定して作る例を公開した。使った指示文も紹介されている。大きなアプリを作る前に、毎日の小さな不便を解消する使い方として真似しやすい。

[Canva @canva]

After tinkering with Canva Code, we built a few simple tools you can use for your workday. Prompts in the thread! 🔧👇🏻

First up: a Pomodoro timer with the intervals and design we wanted, added directly to the file we’re working on. #MadeWithCanvaAI

[しょう | AX @SHO_A]

設計を考えるAIと、実際にコードを書くAIを選び分けられる。Command Codeに新機能が追加。まず計画担当が既存コードを調べ、人が承認したら実装担当へ引き継ぐ。最初から最後まで同じAIに任せず、役割ごとに使い分けたい人向け。

[Command Code @CommandCodeAI]

Plan with one model
Implement with another

now live in Command Code

Your planning model explores the codebase and writes the plan. Approve it, and your implementation model takes over mid-run.

Set both in /config → Feature models

[しょう | AX @SHO_A]

同じAIでも、使う開発ツールが違うだけで成績が変わる。Hugging Faceが、複数の環境をまたいでAIを訓練する手順を公開した。実験では課題の成功率が42％から54％に改善。特定のツールだけで通用するAIにしないための取り組み。

[Hugging Face @huggingface]

The same model, with the same weights, scores 62% in one agent harness and 33% in another.

@adithya_s_k and the @huggingface team just released the ultimate guide to multi-harness RL, and it's one of the most practical RL write-ups this year, and everything open!

The trick is simple. Don't touch the harness. Point it at a proxy instead of the model. The proxy speaks all four API formats coding agents use (OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, Gemini). It records the exact token ids and logprobs vLLM sampled, and you train on that. You don't change a single line of Claude Code, Codex or OpenCode.

Results:
🔹 Trained across 4 harnesses at once, LFM2.5-2.6B by @liquidai went from 42% to 54%
🔹 31% fewer tool calls, thanks to a small bonus for solving tasks in fewer steps
🔹 Training in OpenCode alone took OpenCode from 34% to 58%, but the multi-harness model improved everywhere

They also tried the shortcut everyone reaches for: fine-tune on 3,189 successful rollouts from Qwen3.8-27B. Imitation plateaued at 47.5%, below both RL runs. Copying a bigger model doesn't get you there. Practice does.

The best part is that everything is open: the capture proxy in OpenEnv, the trainer in TRL, the tasks, the SFT data, the training code and all seven trained models.
Agents will run in dozens of harnesses.

Now open models can be trained for each of them, by anyone.

Read it here 👇
https://t.co/s2I8GI9SLS

[しょう | AX @SHO_A]

AIの不具合修正を、直したつもりで終わらせない。LangSmith Engine v2は、問題が起きた環境と入力を再現し、修正案を作ってテストまで繰り返す。最後は人が変更内容を確認して反映。再現と確認にかかる手間を減らす開発者向けの更新。

[LangChain @LangChain]

Validate agent fixes before shipping them with LangSmith Engine v2.

Engine now tests and validates fixes automatically:

✅ Replicates the issue with the same deployment environment and same inputs
✅ Builds a fix, tests it, and iterates on improving it until it has a satisfactory solution
✅ You review the fix and deploy it with a ready-made PR

[しょう | AX @SHO_A]

「きれいだけど、頼んだ絵と違う」を減らす画像AIの研究。Arenaは、人が好む見た目に加え、指定した内容や文字が守られているかも採点して学習させた。既存モデルの成績が改善。見栄えの良さだけを褒めると、指示漏れを見逃すという話。

[Arena.ai @arena]

How to design rewards for post-training frontier image models?

Our research suggests human preference reward is necessary, but insufficient: A preference model may still reward outputs that look appealing but miss details, introduce unrequested content, or exhibit other forms of reward-hacking.

We therefore optimize towards a composite reward:
- Bradley-Terry reward model trained on ~5.6M pairwise human votes
- Faithfulness reward from auto-generated prompt checklists evaluated by a vision-language model
- Constraint reward covering explicit and implicit user intent
- Anti-reward-hacking rubric rewards targeting failures such as garbled text and photorealism drift
This post-training recipe improves two already-strong open image models:
- Post-trained FLUX.2-dev gains 69 Elo points on our live T2I leaderboard, scoring 1202
- Post-trained Ideogram 4 gains 20 Elo points reaching a score of 1224 and surpassing all publicly listed open models (as of Sep 04, 2026).

Offline ablations with Gemini 3.5 Flash as the judge, show that these reward components are complementary: win rate against the base model increases as we add faithfulness and then constraint rewards on top of preference-only training, reaching 64.2%.

Finally, we ensemble policies trained with and without the anti-reward-hacking objective directly in weight space, further increasing win rate to 66.0%.

[しょう | AX @SHO_A]

契約書を読み比べて、取引が止まる原因を探すAIの実演。Boxが22件の資料を使い、資金の不足と顧客の同意漏れを見つける流れを公開した。追加書類で資金不足が解消しても、同意漏れは残ると指摘。資料の更新で結論がどう変わるかまで追っている。

[Box @Box]

We put @OpenAI dots to work on the final checks that could stop a $420 million acquisition from closing.

Using 22 deal documents in Box, the dot uncovers a $3.85 million funding gap at the planned closing date and a missing customer consent and saves a source-cited readiness brief back to Box.

Then we add a signed equity amendment and ask it to update the brief. The funding gap is resolved. The missing consent still blocks closing.

Connect Box to ChatGPT to put your business documents to work.

[しょう | AX @SHO_A]

社内AIに「今の情報」を調べさせる仕組みを作りやすく。CloudflareのAI Gatewayが、ExaなどのWeb検索に対応した。検索結果をAIへの入力に組み込める。古い学習内容だけで答えさせず、最近の記事や公開情報を参照させたい開発者向け。

[Cloudflare @Cloudflare]

Cloudflare AI Gateway now supports native web search API integration in partnership with Ceramic[.ai], Exa, and Linkup. Developers can now inject real-time web context into model inference calls via AI Gateway, REST APIs, or Workers bindings. https://t.co/0Q1bhexFGX #BirthdayWeek

[しょう | AX @SHO_A]

会議の音声を、手元のパソコンで文字にする機能が更新。AnythingLLM 1.17.0では、話者を分けるAIと音声認識を組み合わせた。ダウンロード容量は従来の3分の1に。文字起こしをそのまま自分の資料集に加え、後から質問して探せる。

[AnythingLLM @AnythingLLM]

AnythingLLM 1.17.0 is out. We overhauled the Meeting Assistant to run @NVIDIAAI Nemotron 3 Diarization + @moondreamai Parakeet Redux, on your device.

3x smaller download
70-200x real-time on your hardware (GPU or CPU)
Integrated into your knowledgebase.

https://t.co/YK7BU4g8Ts

[しょう | AX @SHO_A]

調べたいことと論文の抜粋を渡すと、出典付きのレポートにまとめるAI。Ai2が「AstaBrief 8B」を公開した。手元の機器で動かせるうえ、学習用データも公開。研究資料をどう読んで報告書にしたか、仕組みを確認しながら使える。

[Ai2 @allen_ai]

Introducing AstaBrief 8B, an open model that turns complex research questions + literature excerpts into cited reports.

Run it locally on your own hardware, with open weights + training data you can inspect & build on. 🧵

🤗 Download: https://t.co/KvTMbe1e9j

[しょう | AX @SHO_A]

大量の資料を読めるAIでも、長い作業を最後まで正確に続けられるとは限らない。NVIDIAの研究は、表や台帳を順番に更新する課題で検証。文章量だけでなく、入力の書式や1回ごとの処理の難しさでも成績が落ちた。長時間の自動作業を任せる際の盲点。

[DAIR.AI @dair_ai]

Banger paper from NVIDIA on long running agents.

A model can accept 128K tokens of context and still make more mistakes the longer it works through a task.

If your agent loses its place partway through a long table or ledger, this work measures what causes it.

The setup:

Long-Transduction asks a model to keep reading, updating and outputting state-dependent results over thousands of outputs, and varies three factors separately.

Results:

Across seven open-weight models, accuracy drops 62.8% when context grows from 4K to 128K, 36.5% when only the input format changes, and 39.9% when the per-step operation gets harder.

Paper: https://t.co/kjWSHzoegq

[しょう | AX @SHO_A]

有料のAI機能を、自分たちで作って公開するコンテスト。ComfyUIが10月5〜19日に開催し、最優秀賞は1万ドル。RTX 5090などの賞品も用意されている。既製品を使うだけでなく、自分の欲しい画像・動画機能を形にしたい開発者向け。

[ComfyUI @ComfyUI]

The Comfy Dev Platform Challenge is open!

Partners: @ltx_io & @bfl_ai

For two weeks (10/5–10/19), we're challenging the community to open-source a paid AI feature.

🏆 $10K grand prize
🖥️ 3x NVIDIA RTX 5090s
🎁 Bonus prizes from LTX & Flux
⚡ Free credits for the first 100 sign-ups

Register and learn more 👇

[しょう | AX @SHO_A]

会社で使うAI、誰が何にどれだけ使っているか把握できていますか。Factoryが分析画面を刷新。利用者別・AIモデル別に使用量や効率を確認できる。料金の合計だけではわからない「どこで使われ、どこに費用がかかるか」を見るための機能。

[Factory @FactoryAI]

Announcing our new and improved Analytics, giving engineering leaders transparency into consumption, model efficiency, and adoption, broken down by model and by user across every session in the organization.

[しょう | AX @SHO_A]

AIを自動で選んでもらうとき、「なぜそのモデル？」を確認できる。OpenRouter Chatに、Jev Routerが毎回選んだモデルと判断のスコアを示す画面。仕事の難しさや、大きなAIを使う利点などが見える。自動選択に任せつつ、料金や結果の理由を追いやすくなる。

[OpenRouter @OpenRouter]

You can see Jev Router's thinking process in OpenRouter Chat.

The routing insights panel shows the model it picked for each turn and the scores behind it, like task, difficulty, precision, and larger model benefit.

Test Jev Router here: https://t.co/kVWPuaDkgr

[しょう | AX @SHO_A]

便利なAIボットのひな型を作った人へ、報酬を払う動き。Grok Botの制作者向けプログラムが紹介された。公式規約では、米国の対象者への招待制の試験。報酬額はX側が決めるもので、売上の一定割合をもらう制度ではない。日本で誰でも参加できる案内とは分けて見たい。

[Polymarket @Polymarket]

JUST IN: SpaceXAI introduces Grok Bot Creator Rewards, paying creators based on how many people use their AI Bots & how frequently they return.

[しょう | AX @SHO_A]

ロボットに作業を教えるとき、見本を何百回も見せずに済むかもしれない。

NUSの研究「APPL」は、動きの構造を手がかりに技を覚えて組み合わせる方法。公開された実験では、2回の実演から動作を学習。見慣れない物への対応も調べている。

[DailyPapers @HuggingPapers]

NUS researchers release APPL on Hugging Face

APPL uses structural priors as the interface between skill learning and composition.
It boosts OOD generalization from few demos: 89.6% skill success with 2 demos,
50% on shifted objects vs 10% for DP.

[しょう | AX @SHO_A]

音声AIを作るための、大規模な学習データが公開された。YODAS v3は110万時間、100以上の言語を収録。音声に加え、時間情報付きの文字起こしや翻訳も含む。聞き取りや多言語の音声AIを研究する人が、共通の材料から開発を始められる。

[William Chen @chenwanch1]

Who wants more speech training data?

YODAS v3 is now available @huggingface

It’s 1.1M hours - the biggest audio dataset ever. And the first at this scale with stereo audio at 48kHz, with timestamped transcripts and translations. 100+ langs

https://t.co/KLS44NvDC5

CC-BY-3.0

[しょう | AX @SHO_A]

AIの育て方は、正解を教え続ける方法だけではない。Sakana AIの共同創業者らが、進化の仕組みを使う「Neuroevolution」の教科書を刊行。オンライン版は無料で読める。候補を変化させ、良いものを残していく考え方を学ぶための教材。

[hardmaru @hardmaru]

Our Neuroevolution textbook is finally in print!

Free online edition: https://t.co/3IDTb1remp
Pre-order: https://t.co/SUpqBz1HK8

I am incredibly grateful to my co-authors Sebastian Risi, Yujin Tang, and Risto Miikkulainen for making this happen. Neuroevolution is a subject very dear to my heart. It is the field that convinced me that nature has already figured out how to build intelligence: through evolution, collective behavior, and adaptation under constraints.

This idea, that intelligence emerges from evolution operating under constraints rather than unlimited resources, is what eventually led me to founding Sakana AI here in Japan.

The concepts in this book about open-ended creativity and self-organizing systems are exactly what we build at @SakanaAILabs. Our name and logo are inspired by schools of fish moving together, adapting as one. It is the core philosophy of what we build. This book captures the theoretical foundations of that belief.

[しょう | AX @SHO_A]

ゲームで人がどう動いたかを、AIの練習材料にする。WIREDが、ゲームの操作記録を、現実世界を移動するAI向けの学習データに加工する英国企業を紹介した。文章を読ませるだけでなく、「状況を見て、次にどう動いたか」から学ばせる取り組み。

[WIRED @WIRED]

A British startup is shaping video game inputs into training data for AI models that can navigate the physical world. https://t.co/IaghZXtmsE