Moran Mizrahi

How can we help LLMs move beyond the obvious toward generating more creative and diverse ideas? In our new TACL paper, we propose a novel approach to enhance LLM creative generation! arxiv.org/abs/2504.20643 @ChenShani2 @GabiStanovsky @jurafsky @HyadataLab @stanfordnlp @nlphuji

Can LLMs reliably predict program termination? We evaluate frontier LLMs in the International Competition on Software Verification (SV-COMP) 2025, directly competing with state-of-the-art verification systems. @AIatMeta @HebrewU @Bloomberg @imperialcollege @ucl @jordiae @pascalkesseli @jvanegue @HyadataLab @adiyossLC @PeterOHearn12 Paper: arxiv.org/pdf/2601.18987 Website: orensultan.com/llms_halting_p… 🧵👇 1/n

New preprint 💥 Can a general-purpose model achieve results comparable to medically pre-trained models? 🤔 We show that lightweight fine-tuning of a general-purpose LVLM and an LVLM-aware retriever can. 🚀 🔗 GitHub: github.com/Nirmaz/CLARE 📄 Paper: arxiv.org/pdf/2508.17394

Finally taking a moment to post after #EMNLP2025 in Suzhou, China! Had a wonderful time presenting our TACL paper on using structured recombination to push LLMs toward more creative and diverse idea generation 🍳✨ What an amazing experience to conclude my PhD journey! :)

@nadav_bas ממליצה בחום! מדריך מס' 1! :)

עברתי ליד מאפייה בשכונה ושמתי לך שבחזית יש פסלונים קטנים של כל תהליך הכנת הלחם ״הזורעים בדמעה ברינה יקוצרו״ סיורים בפריז איתי: 0033783913015

New benchmarking task for LLM judges: Evaluating debate speeches! 🗣️ We find that judges align with human orderings but remain miscalibrated. Presenting tomorrow at @emnlpmeeting, 14:00-15:30, Hall C. Come by to chat, and even debate (in the spirit of the poster) 🤝 #EMNLP2025

Excited to present our paper "ReliableEval: A Recipe for Stochastic LLM Evaluation via Method of Moments" at #EMNLP2025 🎉 Everyone knows LLMs are prompt-sensitive, yet we still report single-prompt scores. Our work suggests a method to make evaluation statistically reliable!

Looking forward to presenting our TACL paper on enhancing LLM creativity at #EMNLP2025 tomorrow (Wed, Nov 5)! 📍 Room A108 🕝 14:30–16:00 (Linguistic Theories, Cognitive Modeling & Psycholinguistics) Details below ⬇️ #NLP #LLMs #Creativity

👣 Looking forward to sharing our paper on Trajectory Analysis at #EMNLP2025 🚀 this Wednesday, Nov. 5, during Gather Session 2 - Join us virtually! 💻✨ @EitanWagner, @AbendOmri, @RKeydar Details in thread ⬇️ #stayinghome #virtualconferencing

We're proud of our team's 11 papers accepted to #EMNLP2025! See you next week in Suzhou✈️

I can't make it to #EMNLP2025, but @EliyaHabba and @GiliLior will present our PromptSuite🌈(demo!): a framework tackling prompt sensitivity by generating benchmark variations for any task. Try it with just a few lines of code or the web interface! eliyahabba.github.io/PromptSuite/

9/ We believe this approach can extend creative and diverse generation beyond cooking — wherever you can define suitable structures and value criteria (details in the paper!)

8/ Moreover, domain experts judged our recipes mostly coherent / feasible (valuable) and significantly more novel than GPT-4o’s – i.e., more creative overall! ✨

6/ Experiments show that our model’s generations are significantly more diverse compared to GPT-4o outputs: GPT-4o often made shallow merges (preparing two dishes separately and just putting them together), and fixated on certain “creative” ingredients.

5/ Finally, we sample from the space of recombinations, evaluating candidates for novelty and value, and those deemed most promising are realized into full recipes using the LLM (leveraging its commonsense and world knowledge to fill in missing details and fix inconsistencies).

4/ Then, inspired by human creativity, we manipulate these structures to explore new concepts by recombining them.

3/ Our approach starts by using an LLM to parse existing recipes into structured representations. A common way to represent recipes (which are procedural texts) is as trees.

How can we help LLMs move beyond the obvious toward generating more creative and diverse ideas? In our new TACL paper, we propose a novel approach to enhance LLM creative generation! arxiv.org/abs/2504.20643 @ChenShani2 @GabiStanovsky @jurafsky @HyadataLab @stanfordnlp @nlphuji

2/ We demonstrate our approach in the culinary domain, introducing DishCOVER, a model that generates creative recipes.

We present DyPE, a framework for ultra high resolution image generation. DyPE adjusts positional embeddings to evolve dynamically with the spectral progression of diffusion. This lets pre-trained DiTs create images with 16M+ pixels without retraining or extra inference cost. 🧵👇