🧠 New preprint: How Do LLMs Use Their Depth? We uncover a “Guess-then-Refine” mechanism across layers - early layers predict high-frequency tokens as guesses; later layers refine them as context builds Paper - https://t.co/5PitHjmJJZ @neuranna @GopalaSpeech @berkeley_ai

Just did a major revision to our paper on Lifelong Knowledge Editing!🔍 Key takeaway (+ our new title) - "Lifelong Knowledge Editing requires Better Regularization" Fixing this leads to consistent downstream performance! @tom_hartvigsen @_ahmedmalaa @GopalaSpeech @berkeley_ai

Excited to have two paper accepted at #ACL2025 !🎉🎉 1 Main track and 1 Findings. Papers out on arxiv soon. Big thank you to all my collaborators!

🚀 Excited to share that our paper on Plan-and-Act has been accepted to ICML 2025. Below is a TLDR: 🔎 Problem: • LLM agents struggle on complex, multi-step web tasks (or API calls for that matter). • Why not add planning for complex tasks and decouple planning and execution?

#ICLR25 Our work on characterizing alignment between MLP matrices in LLMs and Linear Associative Memories has been accepted for an Oral Presentation at the NFAM workshop. Location : Hall 4 #5 Time : 11 AM (April 27) @GopalaSpeech @berkeley_ai

Work led by BAIR students @KayloLittlejohn and @CheolJunCho advised by BAIR faculty @GopalaSpeech "...made it possible to synthesize brain signals into speech in close to real-time." https://t.co/JOwRq8KrDf via @dailycal

This thread is an excellent and accurate summary of our work ! Thank you @IterIntellectus !

Scientists have developed a device that can translate thoughts about speech into spoken words in real time. Although it’s still experimental, they hope the brain-computer interface could someday help give voice to those unable to speak.

A paper in @NatureNeuro presents a new device capable of translating speech activity in the brain into spoken words in real-time. This technology could help people with speech loss to regain their ability to communicate more fluently in real time. https://t.co/KfIb57KoDR

Today in @NatureNeuro, @ChangLabUCSF and @Cal_Engineer’s @GopalaSpeech show that their new AI-based method for decoding neural data synthesizes audible speech from neural data in real-time: https://t.co/kcBph7k02E

1/n) Our latest work is out today in @NatureNeuro! We developed a streaming “brain-to-voice” neuroprosthesis which restores naturalistic, fluent, intelligible speech to a person who has paralysis. https://t.co/EAk1Wz25TH

(1/5) Introducing Full-Duplex-Bench: A Benchmark for Full-Duplex Spoken Dialogue Models We’re excited to present Full-Duplex-Bench, the first benchmark designed to evaluate turn-taking capabilities in full-duplex spoken dialogue models (SDMs)! https://t.co/78eNKtMAqb Details👇

Our work on knowledge editing got an "Outstanding Paper Award"🏆🏆 at the @RealAAAI KnowFM Workshop!! #AAAI2025 🥳🥳🥳 Congratulations to my amazing co-authors @tom_hartvigsen @_ahmedmalaa @GopalaSpeech

Thrilled to share that our paper on "Norm Growth and Stability Challenges in Sequential Knowledge Editing" has been accepted for an Oral Presentation at the KnowFM workshop @ #AAAI2025 w/ @tom_hartvigsen @_ahmedmalaa @GopalaSpeech More details below (1/n)

Self-Supervised Syllabic Representation Learning from speech. With unsupervised syllable discovery & linear time tokenization of speech at the syllabic rate (~4 Hz) !! Work from my group Cheol Jun Cho, @nicholaszlee and @akshatgupta57 @berkeley_ai to be presented at #ICLR2025

I will be attending #NeurIPS2024 presenting our work SSDM: Scalable Speech Dysfluency Modeling ( https://t.co/97IRz391Ki), East Exhibit Hall A-C #3207, on Thursday, December 12th, 4:30-7:30 PM PST. Hope to meet with old and new friends!

Please checkout the following works on LLM editing and characterization next week at #EMNLP2024

``Sylber: Syllabic Embedding Representation of Speech from Raw Audio,'' Cheol Jun Cho, Nicholas Lee, Akshat Gupta, Dhruv Agarwal, Ethan Chen, Alan W Black, Gopala K. Anumanchipalli,

If you are at #SfN2024, check out our poster in section F9 from 8 AM to 12 PM on Sunday! We present a streamable framework to restore a naturalistic voice to a person with paralysis. We will also show demos of voice synthesis, incremental TTS, and generalization to unseen words.

Catch us at poster #242 on Thursday at 4:30 PM! More details here: 🔗 Website: https://t.co/IvqpqWVFX9 📄 arXiv: https://t.co/aKu3MSsxhg w/ @ren_wang1 , Bernie Huang, @andrewhowens , and @GopalaSpeech .