Ai2 Newsletter
July 2024
Top story - When (and how) AI models should not comply
As we start using LLMs for more daily tasks, can we ensure models reliably reject certain user queries? This concern is paramount among researchers and policymakers, prompting extensive efforts to develop alignment methods and safety benchmarks. We start by categorizing the wide range of noncompliance prompts, such as incomplete requests, subjective matters, humanizing requests, unsafe queries, and more. Based on our taxonomy, we introduce CoCoNot, a resource for training and evaluating models’ noncompliance.
PolygloToxicityPrompts: expanding LLM toxicity analyses to 17 languages
Data toxicity can lead to harmful model outputs — and since most evaluations focus on English datasets, we’re underestimating multilingual toxicity in state-of-the-art LLMs. We partnered with leading minds at Carnegie Mellon University and the University of Virginia to study how LLMs generate toxicity in multiple languages and how design decisions like model size and alignment method impact toxicity.
Evaluating n-gram novelty of LLMs using Rusty-DAWG
Evaluating generation novelty is hard. Even when you have the training data, naively searching it is slow. To assist, our team presents Rusty-DAWG, which builds automata indexes that let you search corpora for infinite-length-n-grams in constant time. This allows us to search massive corpora to evaluate the novelty of text generated by LLMs against their pretraining data.
The 2024 Conservation Tech Award returns
Are you part of an organization developing and deploying technology to protect our planet’s incredible creatures, their habitats, and the communities that depend on them? For the 4th year running, we're providing two $15,000 grants to help protect wildlife through cutting-edge technology! Applications must be fully completed by August 23rd.
Are LLMs creative problem solvers?
Or are humans better at creative problem-solving? We create MACGYVER, a dataset consisting of 1,683 real-world problems designed to trigger innovative, out-of-the-box thinking. We also explore common failure modes in LLMs when solving these problems, and introduce new prompting techniques to improve LLM performance.