With SDPO, you can now do RL with natural language feedback, like error messages from coding environments or LLMs as judges. You can achieve huge gains over GRPO with scalar rewards!

I don't think people have realized how crazy the results are from this new TTT RL paper from Stanford/Nvidia. Training an open source model, they - beat Deepmind AlphaEvolve, discovered new upper bound for Erdos's minimum overlap problem - Developed new A100 GPU kernels 2x faster than the best human kernel - Outperformed the best AI coding attempt and human attempt on AtCoder The idea of Test Time Training is to train a model *while* it's iteratively trying to solve a task. Combining this with RL like they do in this paper opens up the floodgates of possibilities for continual learning Authors: @mertyuksekgonul @LeoXinhaoLee @JedMcCaleb @xiaolonw @jankautz @YejinChoinka @james_y_zou @guestrin @sun_yu_

By specializing the model to each hard task using a simple entropic objective and RL, TTT-Discover writes new high-performing GPU kernels and improves open mathematical bounds.

Relational Transformers provide zero shot predictions for complex databases!

TTT-E2E compresses the context simply through end-to-end test-time training — outperforms the loss of transformers with full attention, and achieves constant-time latency per token.

The Relational Transformer (RT) provides strong zero-shot predictions from relational data, e.g., databases.

SAIL Postdoctoral Fellows is an awesome opportunity to collaborate with the amazing faculty and students in the Stanford AI Lab!

The SAIL Postdoctoral Fellowship is a fantastic opportunity for recent PhDs to do their most innovative and impactful AI research, while being part of a highly collaborative and welcoming environment at Stanford.

We are super excited to empower developers to focus on their goal of building innovative AI applications; we’ll take care of safety and security! What an awesome ride with Bo Li @uiuc_aisecure, @sanmikoyejo, @dawnsongtweets and the whole @VirtueAI_co team!

Excited to share our release of LOTUS 1.1.0, which now makes it easier than ever to get started with LLM-powered data processing over a variety of custom knowledge sources. Seamlessly connect to data from the web, your SQL databases, vector databases and more. And process it all with the power of semantic operators. github.com/lotus-data/lotus

SAIL is delighted to announce Carlos @guestrin, the Fortinet Founders Professor of Computer Science, as the next Director of @StanfordAILab. Carlos is a talented researcher and leader, known for his work on explainability, graphs, compilation, and boosted trees in AI.

We've been building LOTUS at Stanford and Berkeley to make LLM-powered data processing fast, easy and declarative. LOTUS is an open-source query engine that makes programming as easy as writing Pandas and optimizes your programs for up to 400x speedups. To celebrate the holidays, we're excited to share our release of LOTUS 1.0.0 with a batch of new updates that make reasoning over your data faster, easier and better than ever! Code: github.com/guestrin-lab/lotu… 🧵👇

Can interpretability help defend LLMs? We find we can reshape activations while preserving a model’s behavior. This lets us attack latent-space defenses, from SAEs and probes to Circuit Breakers. We can attack so precisely that we make a harmfulness probe output this QR code. 🧵

🧵LLMs are great at synthesizing info, but unreliable at citing sources. Search engines are the opposite. What lies between them? Our new paper runs human evals on 7 systems across the✨extractive-abstractive spectrum✨for utility, citation quality, time-to-verify, & fluency!

Super excited to share our work on LOTUS, a query engine for reasoning over large corpuses of data with LLMs! Joint work w/ the amazing @sid_jha1, @matei_zaharia & @guestrin Read the paper: arxiv.org/abs/2407.11418 Try out the code: github.com/stanford-futureda… 🧵👇

Prior work has used LLMs to simulate survey responses, yet their ability to match the distribution of views remains uncertain. Our new paper [arxiv.org/pdf/2411.05403] introduces a benchmark to evaluate how distributionally aligned LLMs are with human opinions. 🧵

Many providers offer inference APIs for the same models: for example, there were over nine Llama-3 8B APIs in Summer 2024. Do all of these APIs serve the same completion distribution as the original model? In our new paper, ✨Model Equality Testing: Which Model is This API Serving?✨, we formalize this question as a two-sample distribution testing problem: the user collects samples for their task from the API and a reference distribution, and conducts a statistical test to see if the two distributions are the same. We design tests which show nontrivial power in detecting when models have been quantized, watermarked, or finetuned! (🧵 1/5)

Our recent work combines Maps data with hospitalization and death records to model the COVID-19 transmission rates., finding that for many states increasing mobility will lead to a significant increase in cases.

It’s happening! TVM Conference. Incredibly proud of and grateful to the whole UW SAMPL team and TVM contributors.

.@vega_vis is an awesome tool to develop visualization projects! x.com/jeffrey_heer/status/93…