Gunshi Gupta

I joined Mistral AI a few months ago :) check out our new lineup of frontier large and tiny open-weights models. Happy to chat to people curious about open roles in the Science team.

This work was a fun collaboration with @KarmeshYadav @zsoltkira @yaringal and @AljundiRahaf. @KarmeshYadav will present this work at NeurIPS in San Diego. If you are interested in memory and long horizon RL, we'd love to chat.

We’re interested in trying Memo out at VLM-scale. As part of this, we’re releasing FindingDory, a new benchmark designed to evaluate long-context memory in both VLMs (QA) and VLAs (control), filling a gap we noticed during this work. findingdory-benchmark.github.io More on this soon.

How do you give an RL agent useful long term memory when it needs to act over thousands of steps? Storing everything in-context is expensive, text summaries lose detail and plain recurrence struggles with long horizons. Our NeurIPS Spotlight paper explores a simple idea 🧵:

We’re interested in trying Memo out at VLM-scale. As part of this, we’re releasing FindingDory, a new benchmark designed to evaluate long-context memory in both VLMs (QA) and VLAs (control), filling a gap we noticed during this work. More on this soon. Thanks for reading!

@karmeshyadav will present this work at NeurIPS in San Diego. If you are interested in memory and long horizon RL, the paper has many more experiments we tried related to summarization and long-context fine-tuning: Paper: arxiv.org/abs/2510.19732 Code: github.com/gunshi/memo