iiinova — where your digital self keeps evolving

marccote.bsky.social @marccote.bsky.social · 09/09/2026

Froggy Team: Minseon Kim, Eric Yuan, Alessandro Sordoni, Zhengyan Shi, Emiliano Penaloza, Roger Creus Castanyer, Jeonghye Kim, Christopher Cui, Maryam Hashemzadeh, Isadora White, Matheus Pereira, Chinmay Singh, Baolin Peng, Darya Moldavskaya, Fabio Vera @msftresearch.bsky.social

000

marccote.bsky.social @marccote.bsky.social · 09/09/2026

Scaling RL pushes boundaries of the model. We observe a consistent gap in pass@1 and pass@8 between FrogNano and the Qwen3.5-4B. Using a verifier to rank the 8 rollouts increases the success rate from 61.5% -> 62.8%. FrogNano still has a lot left in the pond!

100

marccote.bsky.social @marccote.bsky.social · 09/09/2026

For models at 4B scale, the harness can be essential. We designed LEAF, a lightweight harness, which provides us a 37.2% SWE-bench Verified resolve rate on Qwen3.5-4B compared to 8.3% using the R2E-Gym harness. This made the model start from a much more learnable initial state.

100

marccote.bsky.social @marccote.bsky.social · 09/09/2026

Online calibrated tasks help both maintaining learnability and preserving entropy in RL training.

110

marccote.bsky.social @marccote.bsky.social · 09/09/2026

Throughout iterative training, TaskPilot shifts the task distribution to build stronger coding ability into FrogNano.

100

marccote.bsky.social @marccote.bsky.social · 09/09/2026

FrogNano is trained on tasks synthesized online by TaskPilot, a pipeline that uses LLMs to generate a targeted task sets for the current checkpoint. At each iteration we sample tasks within the checkpoint's learnability zone, and once the gains saturate we repeat the process.

100

marccote.bsky.social @marccote.bsky.social · 09/09/2026

We observe similar gains on SWE-bench Pro, Terminal-Bench 2.0, and PatchEval-Verified. FrogNano competes with models 6–8× its size while being substantially cheaper to serve and train.

100

marccote.bsky.social @marccote.bsky.social · 09/09/2026

4B parameters. ZERO distillation. 61.5% on SWE-bench Verified 🤯 Meet FrogNano 🐸: Qwen3.5-4B post-trained purely with RL on synthetic tasks from TaskPilot. Just 5 iterations × 300 tasks. Who said coding agents have to be huge? 🐸

100

Reposted by @marccote.bsky.social

Akhil Arora @akhilarora.bsky.social · 03/06/2026

AI agents like @openclaw 🦞 are everywhere, answering emails, managing calendars, doing our chores for us. 📣 REALM is back for year 2! Workshop for Research on Agent Language Models at #EMNLP2026, Budapest 🇭🇺 🌟 Stellar lineup ⬇️ 📅 Submit by July 17, 23:59 AoE #AgenticAI #AIagents #LLMAgents #NLProc

REALM 2026 workshop invited speakers and panelists: Maria Brbić (EPFL), Vishrav Chaudhary (Meta Superintelligence Labs), Marc-Alexandre Côté (Microsoft Research), Manling Li (Northwestern University), Baptiste Rozière (Mistral AI), Graham Neubig (Carnegie Mellon University), Perouz Taslakian (ServiceNow AI Research).

195

Reposted by @marccote.bsky.social

Lara J. Martin @laramartin.net · 29/08/2025

Congratulations to the winners of the Commonsense Personal Grounded Dialogue Challenge! discourse.aicrowd.com/t/winners-ca... we look forward to hearing about the details of your systems at the EMNLP 2025 Wordplay Workshop!

discourse.aicrowd.com

🏆 Winners & Call for Paper

Hello all, Thank you for being part of the Commonsense Persona Grounded Dialogue Challenge 2025! Whether you finished at the top or joined for the first time, we’re glad you participated in this seco...

031

marccote.bsky.social @marccote.bsky.social · 04/08/2025

Excited to be presenting aka.ms/debug-gym at IVADO's Bootcamp: Focusing on the Current State of Agents ivado.ca/en/events/bo...

ivado.ca

Bootcamp: Focusing on the Current State of Agents | IVADO

000

Reposted by @marccote.bsky.social

LawZero - LoiZéro @law-zero.bsky.social · 03/06/2025

Tout système d'IA de pointe devrait reposer sur un engagement essentiel : celui de préserver la joie et les aspirations humaines. Aujourd’hui, nous lançons LoiZéro, une organisation à but non lucratif visant à concevoir des systèmes d'IA sécuritaires. loizero.org

01910

Reposted by @marccote.bsky.social

LawZero - LoiZéro @law-zero.bsky.social · 03/06/2025

Every frontier AI system should be grounded in a core commitment: to protect human joy and endeavour. Today, we launch LawZero, a nonprofit dedicated to advancing safe-by-design AI. lawzero.org

04623

marccote.bsky.social @marccote.bsky.social · 12/04/2025

Join us on April 17th for Mila TechAide 2025! 🌟 Hear from amazing speakers in #AI like Sara Hooker, Kyunghyun Cho, Golnoosh Farnadi, David Adelani, Derek Nowrouzezahrai, and Tegan Maharaj. All proceeds support #Centraide. Let's make a difference together! 💡🤝 Get your tickets now! t.co/EBsatAqGTA

mila.quebec

Mila Techaide 2025 | Mila

Conférence IA d'une journée au profit de Centraide du Grand Montréal

054

Reposted by @marccote.bsky.social

Microsoft Research @msftresearch.bsky.social · 10/04/2025

Developers spend a lot of time debugging code. Learn how debug-gym can equip AI agents to help, enabling them to set breakpoints, navigate the codebase, and print runtime variable values on demand, so they better understand the code and its execution flow: msft.it/6017qF6RT

A graphic with a gradient background transitioning from blue on the left to pink on the right. The graphic features a white outline of a computer monitor with code brackets on the screen, an arrow pointing downwards into the monitor, and another arrow curving around to point upwards towards a magnifying glass with a bug icon inside it.

021

Reposted by @marccote.bsky.social

Neil Tenenholtz @ntenenz.bsky.social · 17/11/2024

Looking to find your MSR colleagues? I put together a starter pack. Msg me if I missed you (or you just joined). go.bsky.app/NxtpELZ

6157