Jeff Shen

PhD Candidate @Princeton, researcher @PolymathicAI
working on generative and multimodal models for science

[Jeff Shen @jeff_shen_]

@JulienBlanchon @Jay41969851 @_ryantlee @jonathanmay @cosmo_shirley we are currently running tests at different batch sizes, will let you know how it holds up. for policy drift, no, we did not measure that here. i expect that given that this drafter is trained offline and not cotrained with the target model performance would degrade over training

[Jeff Shen @jeff_shen_]

Announcing FlashMap: a new speculative decoding algorithm based on flow maps! >6x faster inference on Qwen3-8B 🧵

Project page: https://t.co/DZcgcCM04a
arXiv: https://t.co/VBwQgu3gsv

w/ @Jay41969851 @_ryantlee @jonathanmay @cosmo_shirley

[Jeff Shen @jeff_shen_]

@lollipop_stat @Jay41969851 @_ryantlee @jonathanmay @cosmo_shirley yes! would definitely work. havent done it (yet) because of compute limitations :)

[Jeff Shen @jeff_shen_]

@EmircanErkulX @Jay41969851 @_ryantlee @jonathanmay @cosmo_shirley we do :) on Qwen3–8B we are ~7% faster than DFlash and ~60% faster than EAGLE-3 across code, math, and chat benchmarks. see paper for more details

[Jeff Shen @jeff_shen_]

@kartikb753 @Jay41969851 @_ryantlee @jonathanmay @cosmo_shirley to some extent! we are roughly 11% faster than a matched DFlash baseline there. in general speculative decoding algorithms are a bit weaker there though, we get ~2.8x speedup at T=0 and ~2.3x at T=1 on chat-style text benchmarks (MT-Bench, Alpaca, Arena-Hard)

[Jeff Shen @jeff_shen_]

@Jay41969851 @_ryantlee @jonathanmay @cosmo_shirley Try it at https://t.co/hWeTNlRIyT

• Drafters for Qwen3-8B and Qwen3-4B on Hugging Face
• Integrations with Transformers, vLLM, and SGLang
• Training recipes for speculators and SpecForge

[Jeff Shen @jeff_shen_]

@Jay41969851 @_ryantlee @jonathanmay @cosmo_shirley FlashMap uses a flow map (from Flow Map Language Models), which jumps from Gaussian noise to the whole block in a single network evaluation, modeling the block jointly rather than position by position. The target then verifies the block, so the output is exactly its own.

[Lindsay Smith @LindsayMSmith3]

DiscoverPhysics leaderboard update: Claude Fable 5 solves the majority of our current worlds! Pass@5 rate: 73%, ~1.5x next best model (Opus 4.7, 50%). It even passes difficult worlds with latent structure, which all other models struggle with.
🌐 https://t.co/OCdfxhqRlU

[Rudy Morel @rdMorel]

Partial observability is a key challenge in predicting physical systems, where only part of the state is observed.

Check out our poster #2213 at #neurips2025 on Thu, Dec 4, 4:30pm! We propose a multiscale inference scheme for diffusion models to better predict these systems.

[Mike McCabe @mikemccabe210]

1/ Today with my colleagues @PolymathicAI, I'm excited to release our latest project, Walrus, a cross-domain foundation model for physical dynamics, into the world.

https://t.co/ihv1MZGQM3
Paper: https://t.co/d6ah9LO4ud
Git: https://t.co/s3p8qGhZQR
HF: https://t.co/RufaBD9eJk

[Shirley Ho @cosmo_shirley]

Very Cool project led by @jeff_shen_ and @LindsayMSmith3 on cracking secret codes (aka cipers) with a transformer that is even interpretable!

It beats all previous efforts in this area!

[Lindsay Smith @LindsayMSmith3]

🧩🤖Solving puzzles with AI!
@jeff_shen_ and I are excited to share our new project: ALICE, a neural network that cracks substitution ciphers with SOTA accuracy & speed, and is easily human-interpretable.
Project page + demo: https://t.co/p0TyhfmOlk
Paper: https://t.co/EoYBi6oIp0