Happy to announce my upcoming South Korea 🇰🇷 tour next week 🤩

KAIST - Monday 4/15, 2:30pm E3-1, Rm 4443. Host: Alice Oh
SKKU - Suwon, Tuesday 4/16, TBD, Engineering Hall 2. Host: JinYeong Bak
SNU - Wednesday 4/17, 1pm see below. Host: Jay-Yoon Lee

I hope to meet new + old friends!

open.spotify.com/show/1zpZxyZOQ…

Can off-the-shelf large language models help translate low-resource and endangered languages despite not seeing these languages in their training data? Maybe! I spoke to Kexun Zhang about LingoLLM, a workflow and pipeline that helps upgrade the language…

Call me biased, but Wenhu Chen and Huan Sun (OSU) Yu Su are doing some of the most impressive research in generative AI these days.

New preprint analyzing 1.8 million neurons of CNNs with emergent behavioral and neuronal signatures of covert attention without incorporating any explicit attention mechanism. With Sudhanshu Srivastava William Wang biorxiv.org/content/10.110…

Now all the slides and recorded videos are uploaded to the course website now: cs.uwaterloo.ca/~wenhuche/teac…

Kudos to all the great students taking the course!

Surprising results by Michael/Mahsa/Fatima suggest that classic feature-space metrics like CLIPScore may outperform advanced LM-based metrics in evaluating text-to-image fidelity. Despite high human preference correlations, LLM metrics struggle in real-world scenarios. 🤔🤔🤔

Observation: Relying too heavily on prompt engineering can stifle the creativity and exploration spirit of PhD students. It's crucial to remember that breakthroughs and fundamental innovations come from diving deep, questioning, and reimagining the boundaries of what's possible.

Join us for the Mellichamp Mind & Machine Intelligence annual summit, April 18th and 19th. We’ll explore examples of using AI in the creative process, discuss pressing questions, and ignite debates around the interplay of AI and Human Creativity.
mind-machine.ucsb.edu/events/all/202…

Why would you register #NeurIPS2024 now to apply for a visa? Well, if your visa is denied, then you can apply for full refund of your registration: neurips.cc/FAQ/Cancellati…

Long-context LLMs Struggle with Long In-context Learning

Suggests a notable gap in current LLM capabilities for processing and understanding long, context-rich sequences.

arxiv.org/abs/2404.02060

LLM often could not correct its own mistakes. However, using a fine-grained feedback model, we could teach LLM how to correct its incorrect generation. Introducing LLMRefine: the power of simulated annealing on top of fine-grained feedback!

Check out: arxiv.org/abs/2311.09336

How can we control helpfulness and safety levels in LLMs? We demonstrate that solely automatically revising the same data for a pretrained model can unlock the model’s controllability in these attributes.

My last summer internship paper is finally out arxiv.org/abs/2404.01295!

NeurIPS 2024 Registration is OPEN! neurips.cc We have notified the Canadian government about the conference, but due to the challenges of getting Canadian visa in the past, we strongly encourage participants register early to obtain their visa invitation letter. 🇨🇦

When LLMs make mistakes, can we build a model to pinpoint error, indicate its severity and error type? Can we incorporate this fine-grained info to improve LLM? We introduce LLMRefine [NAACL 2024], a simulated annealing method to revise LLM output at inference. Google AI UCSB NLP Group

A must watch for LLM practitioners!

Describing the recent advancements in NLP by calendar years feels nearly impossible. Yet, when we frame these developments in terms of LSTM time, BERT/GLUE time, GPT-3, ChatGPT, and GPT-4/LLM eras, the narrative of progress becomes remarkably clear and intuitive. 😎

Not an expert on leaderboarding, but I'm genuinely curious what's the tweak between 0.2 and 0.3, such that it gave a 20 point boost on MMLU while not changing other performance? 🤔

Let's be honest. For everyone in ML/NLP, it's a really exciting time but also very stressful with so many new papers, models, benchmarks, deadlines, reviews, talks, workshops, conferences...

How do you keep up and stay sane?

For me, the only solution is collaboration. 1/n