Conference · Budapest, Hungary
EMNLP 2026
I presented Rethinking the Multilingual Reasoning Gap with Layer Swap at EMNLP 2026 in Budapest.
-
-
Model release
ColGrep-Agent-2B
We released a small code-search agent that runs locally and helps coding agents find the right files with semantic search.
-
Preprint
TrustMI: Causally controlling how assistants trust their users
We study how activation steering can control an assistant’s trust in its users and how that trust shapes safety-relevant behavior.
-
Preprint
Small Agents with Semantic Search: Efficient Multilingual Code Localization
We train compact agents to find relevant files with semantic search, improving code localization across programming languages while reducing latency and token usage.
-
Paper acceptance
Accepted to Findings of EMNLP 2026
Rethinking the Multilingual Reasoning Gap with Layer Swap was accepted to Findings of EMNLP 2026.
-
Model release
Luth-2
We released Luth-2-0.8B and Luth-2-2B: small models specialized for French through supervised fine-tuning, reinforcement learning, and multi-domain on-policy distillation.
-
Conference · Rabat, Morocco
EACL 2026
I presented Luth, our work on efficient French specialization for small language models and cross-lingual transfer, at the EACL Student Research Workshop.
