TIL: When distilling reasoning capability from a teacher LLM to a smaller LLM, you should use Agent traces instead of CoT traces. Advantages are:
1. Increased generalization
Intuitively, this is because your agent can encounter more "surprising" results by interacting with its
@aymericroucher
-
Agent traces better than CoT for distilling reasoning
By
–
-
Datadog’s new open model tops forecasting benchmarks with BOOM
By
–
Who said Transformers couldn't be good at forecasting? Datadog's new open model tops forecasting benchmarks! And boy did they cook. They followed the playbook to build the best model: 1. The best benchmark They release a new benchmark named BOOM, based on observability
-
OpenAI Codex Agent Shows Net-Positive Coding Gains, Future Bright
By
–
I've tried OpenAI's Codex, the autonomous SWE agent integrated into ChatGPT. -> While certainly not producing one-shot perfect PRs, it's already a net-positive. And the gain will only improve with better coder models! I think that in the next era of coding, most will be done
-
ByteDance’s small Seed1.5 model beats Claude and Gemini in vision tasks
By
–
ByteDance takes the lead on vision models: small (~21B) Seed1.5 beats behemoths like Claude 🔥
— m_ric (@AymericRoucher) 16 mai 2025
ByteDance just released the technical report for their new model Seed 1.5, that achieves performance on vision tasks on par with the current king Gemini-2.5-Pro (leaving OpenAI o1 and… pic.twitter.com/gx631UzBm7ByteDance takes the lead on vision models: small (~21B) Seed1.5 beats behemoths like Claude ByteDance just released the technical report for their new model Seed 1.5, that achieves performance on vision tasks on par with the current king Gemini-2.5-Pro (leaving OpenAI o1 and
-
Exploring application for agent runs with scarce agentic traces
By
–
Indeed! We'll have to look into applying it for our own agent runs, since agentic traces are hard to find!
-
LLMs train without external data, breaching the “data wall”?
By
–
Absolute Zero: LLMs can train without any external data Has the "data wall" just been breached? Recent RL paradigms often relied on a set of questions an answers that needs to be manually curated. Researchers from @Tsinghua_Uni went like "why though". Indeed, why learn
-
PyMuPdf drops figures; keeping images requires switching to VLM
By
–
Just PyMuPdf! So figures are dropped, a good direction for improvement would be to keep images, but then we'd need to switch to a VLM instead of LLM.
-
Open source NotebookLM alternative uses Llama 3.3-70B for podcasts
By
–
I've made an open and free version of Google's NotebookLM, and it shows how high the open source tech task has risen! The app's workflow is simple. Given a source PDF or URL, it extracts the content from it, then tasks @AIatMeta
's Llama 3.3-70B, with writing the podcast -
Smolagents 1.15 release adds streaming outputs
By
–
¡Releasing smolagents 1.15, with streaming outputs! ⏩
— m_ric (@AymericRoucher) 8 mai 2025
✨ We've just added streaming outputs in smolagents, meaning that all interactions can feel much smoother.
Just pass stream_outputs=True to your CodeAgent upon initialization to try it out!
[Insert feels Good man meme] pic.twitter.com/EB05A19bgj¡Releasing smolagents 1.15, with streaming outputs! We've just added streaming outputs in smolagents, meaning that all interactions can feel much smoother. Just pass stream_outputs=True to your CodeAgent upon initialization to try it out!
[Insert feels Good man meme] -
Struggling to find a short package name due to abandoned projects
By
–
I've had the issue when trying to launch smolagents: I wanted a short, impactful name, and many names in something*agent*something were already taken. Mot of these packages had had 1 initial release, then went unmaintained for 10 years. But they still kept the name, because