WIth Lean-FIRE, we achieved the first end-to-end autoformalization for 13 Putnam problems, but also showed that conjecturing is still a massive reasoning challenge.

We're tackling this by treating conjecturing as its own task. We introduce ConjectureBench to measure performance, and Lean-FIRE, a new method to guide LLMs to generate better conjectures.

"Code-Optimise: Self-Generated Preference Data for Correctness and Efficiency" with Leonidas Gee,
Milan Gritta, Ignacio Iacobacci NAACL 2025 (Findings)

[Paper] arxiv.org/pdf/2406.12502

"Mixture of Attentions for Speculative Decoding" with
Matthieu Zimmer, Milan Gritta, Haitham Bou Ammar, Jun Wang
ICLR 2025

[Paper] arxiv.org/pdf/2410.03804
[Code] github.com/huawei-noah/...

"Human-like Episodic Memory for Infinite Context LLMs" with @zfountas.bsky.social, Martin Benfeghoul, Adnan Oomerjee, Fenia Christopoulou, Haitham Bou Ammar, Jun Wang
ICLR 2025

[Paper] arxiv.org/pdf/2407.09450
[Code] github.com/em-llm/EM-LL...