You can align reasoning LLMs with just 1K data now! UC Santa Cruz released STAR-1, showing that fine-tuning Large Reasoning Models with it boosts safety performance by 40% on average—while barely affecting reasoning ability.
RESEARCH
-

Gemini AI Helps Decode Rare Disease Information Effectively
By
–
This video is a really nice story about how a sophisticated AI model like Gemini can help make sense of complex information about a rare disease, leading to better understanding for all.
-
AI Improvement Through Precision-Targeted Fixes and SEAL Research
By
–
AI improvement at the labs has evolved from a guessing game to one driven by precision-targeted fixes to identified failure modes. @scale_AI
's SEAL research benchmarks and evaluation platform is making this possible. Check out coverage by @willknight -

Most Powerful 7B Language Model Uses Diffusion Reasoning
By
–
The most powerful 7B language model yet—and it's a diffusion reasoning model.
-

ECLeKTic Benchmark Evaluates Cross-Lingual Knowledge in LLMs
By
–
Introducing ECLeKTic, a new benchmark for Evaluating Cross-Lingual Knowledge Transfer in LLMs. It uses a closed-book QA task, where models must rely on internal knowledge to answer questions based on information captured only in a single language. More →
https://
goo.gle/3Y5TqvZ -
Two Flavors of AI for Scientific Innovation in Next Five Years
By
–
I am super excited for AI for scientific innovation, a direction that will certainly grow in the next five years. I think there will be two flavors of it. The first is “deepmind style”, where there is a very specific, important problem to solve (e.g., protein-folding), and you
-

DAIR Virtual Event: Imagining Possible Technological Futures Workshop
By
–
We're organizing a hopeful virtual event to get us to think our possible technological futures, and have a creative workshop. You can register at https://
eventbrite.com/e/dair-present
s-imagining-possible-futures-tickets-1295350711849
… Friday April 11 9am-12pm PDT | 12pm-3pm EDT | 6pm-9pm CEST | 7pm-10pm EAT -

Model Represents Humanity’s Closest Approach to AGI
By
–
This model is the closest humanity has ever been to AGI.
-
Claude 3.5 Sonnet Achieves 21% on PaperBench Research Task
By
–
We evaluate several frontier models on PaperBench, finding that the best-performing tested agent, Claude 3.5 Sonnet (New) with open-source scaffolding, achieves an average replication score of 21.0%. Finally, we recruit top ML PhDs to attempt a subset of PaperBench, finding that
-
OpenAI Open-Sources PaperBench AI Evaluation Framework
By
–
We’re open-sourcing PaperBench: https://
github.com/openai/prepare
dness/tree/main/project/paperbench
… Read the full paper: https://
openai.com/index/paperben
ch/
…
