@VmaxAI 🤝 @modal
Very happy to finally share this! We worked with @modal to make coding agents better at using the Modal Python SDK with our in-house self-play algorithms, and the results were 🔥🔥
All the details in the blog post below! ⬇️
1/ As a teaser for Campaign, Vmax’s infrastructure for automated task generation, we are sharing work we did earlier this year for @modal using self-play to generate tasks that target weaknesses of models when using the Modal-SDK.
"including training in games [...] strategies learned there transfer to research mathematics and other fields"
we also showed with GlyphBench that training LLMs in games can yield general reasoning gains!
x.com/VmaxAI/status/…
Rumors I’ve been hearing. I’ll start with OpenAI, then Anthropic.
OpenAI: The internal model’s reasoning jump comes from more aggressive RL, including training in games and simulated environments. Strategies learned there transfer to research mathematics and other fields.
🚨 Excited to release GlyphBench, our new paper on games, RL, and LLM reasoning!
big, big thanks to @MavorParker, @matthewjsargent, @Cote_Marc, @GlenBerseth & @pcastr
and shoutout to @wwwjim for the beautiful blog design!
preprint, code, and blog post are all public! ✨
1/ 🎮 Can training on games improve LLM reasoning?
Introducing GlyphBench: 360+ game tasks for language-model RL.
We use it as a benchmark, a research tool, and a post-training suite.
🧵⬇️
Want to train your SWE agents 2× faster 🏎️💨?
We introduce KV-streams, which, rather than flushing the KV cache on agentic compaction, we preserve it and keep generating. This avoids the re-prefill cost on both the inference engine and the trainer.
Unlike other works that
