Poorva Garg

Very excited to be leading this research direction, employing probabilistic programming to improve LLM inference for code generation.

@sandyskim and I will be at #cshldata24 ! Sandy’s giving a talk tomorrow on hierarchical modeling of CRISPR screens I’ll be presenting a poster on integrating polygenic risk to improve LDL-C measurement Come check us out :) @hjpimentel @cshlmeetings

🚀 I am recruiting PhD students for Fall 2025 at the UCLA Robot Intelligence Lab! 🤖 If you are interested in robot learning and human-robot interaction, mark me as a potential adivisor when you apply to the UCLA CS PhD program! #PhD #Robotics @CS_UCLA

1/n @uclanlp is researching how Black, LGBTQIA+, & women communities perceive and are affected by content moderation, as it relates to English-language social media content using reclaimed language. As part of this, we are recruiting annotators (forms.gle/KP6F9gDCo8Skjs…) …

I am looking for 1-2 PhD students interested broadly in computer systems, compilers, and/or PL! If you would like to do your PhD in a vibrant city with great weather, next to the sea 🌊, and the mountains 🏔️, make sure to apply to UCLA and mark my name as a potential advisor!

[1/n] 🚀Diffusion models for discrete data excel at modeling text, but they need hundreds to thousands of diffusion steps to perform well. We show that this is caused by the fact that discrete diffusion models predict each output token *independently* at each denoising step.

Where is the signal in LLM tokenization space? Does it only come from the canonical (default) tokenization? The answer is no! By looking at other ways to tokenize the same text, we get a consistent boost to LLM performance! arxiv.org/abs/2408.08541 1/5

Check out our tool at github.com/Tractables/Dic… and our PLDI paper at dl.acm.org/doi/10.1145/36…. This work is a collaboration with @zengola, @guyvdb, and web.cs.ucla.edu/~todd/

HyBit uses the above theoretical results to bit-blast arbitrary discrete-continuous probabilistic programs. We scale with discrete structures in hybrid probabilistic programs better than state-of-the-art inference algorithms on a comprehensive suite of benchmarks.

Are you looking for an inference algorithm that supports your discrete-continuous probabilistic program? Look no further! We have developed a new probabilistic programming language (PPL) called HyBit that provides scalable support for discrete-continuous probabilistic programs.

Proposing Ctrl-G, a neurosymbolic framework that enables arbitrary LLMs to follow logical constraints (length control, infilling …) with 100% guarantees. Ctrl-G beats GPT4 on the task of text editing by >30% higher satisfaction rate in human eval. arxiv.org/abs/2406.13892

🚨LLM RESEARCHERS🚨Want a free boost in speed and memory efficiency for your HuggingFace🤗LLM with ZERO degradation in generation quality? Introducing Prepacking, a simple method to obtain up to 6x speedup and 16x memory efficiency gains in prefilling prompts of varying lengths. arxiv.org/pdf/2404.09529… w/ @danielmisrael @guyvdb @adityagrover_

Very excited for LAFI@POPL coming up this Sunday: popl24.sigplan.org/home/lafi-2024 @Hong_Ge2 will be presenting our keynote! 18 accepted posters spanning many topics within probabilistic and differentiable programming Sponsored by @BasisOrg this year! Hope to see you there!

Neuro-Symbolic (NeSy) methods inject constraints into NNs, but do not support autoregressive models e.g. transformers. We propose Pseudo-Semantic loss a NeSy framework for injecting arbitrary logical constraints into autoregressive models. At #NeurIPS2023! openreview.net/pdf?id=hVAla2O…

Featuring our chapter on semantic loss, neuro-symbolic entropy regularization and constrained adversarial networks! So glad it's out !!

Uncertainty quantification for neural networks via Bayesian model average is compelling, but uses just a few samples in practice😥 We propose 👾CIBER, a collapsed sampler to aggregate infinite NNs via volume computations, w/ better accuracy & calibration! arxiv.org/abs/2306.09686

Shoutout to @kaavya_sahay for her new paper on new year.

Great findings from @uclanlp on bias propagated via vision language models in their new paper, ENTIGEN.