Gallil Maimon

Speech LMs, Code LMs; PhD @CseHuji, Ex - Meta (FAIR)

[Gallil Maimon @GallilMaimon]

🚨New paper🚨 Self-Execution Simulation Improves Coding LLMs

Current reasoning LLMs <think> before giving a code solution.

We post-train them to explicitly simulate test execution, in order to verify and fix their own code leading to additional gains!

Details 👇🏽🧵

[Gallil Maimon @GallilMaimon]

PhD Submitted🎉🎓

What a ride🎢 who'd have thought over a decade ago coding my (crappy) ASR system😅

So grateful to my amazing supervisor @adiyossLC for always focusing on good science, and letting me make mistakes. And of course my collaborators, lab mates, family and wife❤️

[Gallil Maimon @GallilMaimon]

@MichaelHassid @HebrewU @royschwartzNLP @adiyossLC @AIatMeta Totally well deserved! Lucky to have had chances to work together :)

[Gallil Maimon @GallilMaimon]

@DimitrisPapail Very cool! You might also find our recent work interesting :) We explicitly post-train models to simulate the execution of their own code, and use that for self-verification and iterative self-refinement 👇

https://t.co/kK9SgmnEGF

[Gallil Maimon @GallilMaimon]

🚨New paper🚨 Self-Execution Simulation Improves Coding LLMs

Current reasoning LLMs <think> before giving a code solution.

We post-train them to explicitly simulate test execution, in order to verify and fix their own code leading to additional gains!

Details 👇🏽🧵

[Gallil Maimon @GallilMaimon]

@xwang_lk Agreed! Plus getting the paper uploaded to Arxiv after “on hold” is starting to take as long as a review cycle 😂😅

[Gallil Maimon @GallilMaimon]

@LChoshen Sent a DM :)

I also think it is interesting to understand if your experiment represents a local minima (I.e. sharing a representation would have lower loss), or if under the setting (data, model size) having separate representations would lead to lower train/val loss

[Gallil Maimon @GallilMaimon]

@LChoshen We saw cases of knowledge transfer (answering in speech facts only seen in text) when looking at interleaved speech-text LMs like the ones from our paper - https://t.co/tgzJ3X4hhC

We are working now on understanding how this works, what is needed, how to boost:)

[Michael Hassid @MichaelHassid]

Have LLMs become supervised learners (once again)?!

In our new paper, we argue that current LLMs’ post-training methods have effectively reverted to the "pre-train then fine-tune" era, explicitly tailoring models to desired behaviors.

1/n

[Alon Wolf | Researcher @AlonWolfy]

[1/5] Is Text Enough for Control? 🐇

Text-driven video editing lets you describe *what* to change. But what about *how much*?

We introduce Adaptive-Origin Guidance (AdaOr).

A joint work with @DecartAI and @TelAvivUni 🧪

accepted to #SIGGRAPH2026.

[Iddo Yosha @iddoyosha]

StressTest is heading to #ACL2026! ✌️

Updates include:
☕️ StressPresso, new benchmark for stress tasks generalization.
🧠 More models and more complex reasoning tasks!

Thanks to the GREAT collaborators @GallilMaimon @adiyossLC !

Project and paper: 🔗 https://t.co/pA7FAtaHOJ

[Boyuan (Nemo) Chen @boyuan_chen]

Coding models still have a weird blind spot: they can write code, but often can't faithfully predict what that code will do.

"Self-Execution Simulation Improves Coding Models" attacks that gap by training models to mentally execute programs, then using that skill to rank and repair solutions.

[fly51fly @fly51fly]

[CL] Self-Execution Simulation Improves Coding Models
G Maimon, O Yoran, F Kreuk, M Hassid… [FAIR team, Meta] (2026)
https://t.co/uTaHky6YCm

[Gallil Maimon @GallilMaimon]

@MiniMax_AI Really interesting! “training coding agents relies heavily on execution... which brings massive computational overhead... We are exploring building a World Model capable of predicting code execution results”

Our new work is highly related, check it out👇
https://t.co/kK9SgmnEGF

[Gallil Maimon @GallilMaimon]

🚨New paper🚨 Self-Execution Simulation Improves Coding LLMs

Current reasoning LLMs <think> before giving a code solution.

We post-train them to explicitly simulate test execution, in order to verify and fix their own code leading to additional gains!

Details 👇🏽🧵

[Gallil Maimon @GallilMaimon]

@arxiv FINALLY, after almost a month 🤯 It's out :)

https://t.co/kK9SgmnEGF

[Gallil Maimon @GallilMaimon]

🚨New paper🚨 Self-Execution Simulation Improves Coding LLMs

Current reasoning LLMs <think> before giving a code solution.

We post-train them to explicitly simulate test execution, in order to verify and fix their own code leading to additional gains!

Details 👇🏽🧵

[Gallil Maimon @GallilMaimon]

@arxiv upload "on hold" for over a week and not for the first time🥲 Guess the world isn't ready for the paper😅

No reason provided, reach out responded with generic "wait". Would be useful to get reasons so that next time we can avoid overloading the moderators

[Gallil Maimon @GallilMaimon]

📜Paper - https://t.co/LrbSjr0nnF

Thanks to my amazing collaborators! @OriYoran @FelixKreuk @MichaelHassid Gal Cohen @PierreChambon6 @adiyossLC

Done during internship @FAIR @Meta

[Gallil Maimon @GallilMaimon]

🚨New paper🚨 Self-Execution Simulation Improves Coding LLMs

Current reasoning LLMs <think> before giving a code solution.

We post-train them to explicitly simulate test execution, in order to verify and fix their own code leading to additional gains!

Details 👇🏽🧵

[Gallil Maimon @GallilMaimon]

🎓💯 Inspired by RL with execution feedback, we also train the model in multi-turn RL to fix/submit its own code given simulated execution feedback.
Turns out you can get boosts relative to strong reasoning models even without tools and real execution!