To benchmark a model's long-form factuality in open domains, we first use GPT -4 to generate LongFact, a prompt set comprising thousands of questions spanning 38 topics. We then propose that LLM agents can be used as automated evaluators for long-form factuality through a method which we call Search-Augmented Factuality Evaluator (SAFE).

There is justifiable interest in leveraging conversational AI (CAI) for health across the majority world, but to be effective, CAI must respond appropriately within cultur ally and linguistically diverse context s . Therefore, we need ways to address the fact that current LLMs exclude many lived experience s globally . Various advances are underway which focus on top - down approaches and increas ing training data . In this paper, we aim to complement these with a bottom - up locally - grounded approach based on qualitative data collected during participatory workshops in Latin America. Our goal is to construct a rich and human - centred understanding o f: a) potential areas of cultural misalignment in digital health; b) regional perspectives on chatbots for health and c) strategies for creating culturally - appropriate CAI; with a focus on the understudied Latin American context . Our findings show that academic boundaries on notions of cultur e lose meaning at the ground level and technologies will need to engage with a broad er framework; one that encapsulates the way economics, politics, geogr aphy and local logistics are entangled in cultural experience. To this end, we introduce a framework for ' Pluriversal Conversational AI for H ealth ' which allows for the possibility that more relationality and tolerance, rather than just more data, may be called for .

Large language models (LLMs) are widely used for long-form text generation. However, factual errors in the responses would undermine their reliability. Despite growing attention to LLM factuality, the effect of response length on factuality remains underexplored. In this work, we systematically investigate this relationship by first introducing an automatic and bi-level long-form factuality evaluation framework, which achieves high agreement with human annotations while being cost-effective. Using this framework, we conduct controlled experiments and find that longer responses exhibit lower factual precision, confirming the presence of length bias. To explain this phenomenon, we empirically examine three hypotheses: error propagation, long context, and facts exhaustion. Our results reveal that facts exhaustion, where the model gradually exhausts more reliable knowledge, is the primary cause of factual degradation, rather than the other two hypotheses.

The BabyLM challenge called on participants to develop sample-efficient language models. Submissions were pretrained on a fixed English corpus, limited to the amount of words children are exposed to in development (<100m). The challenge produced new architectures for data-efficient language modelling, which outperformed models trained on trillions of words. This is promising for low-resource languages, where available corpora are limited to much less than 100m words. In this paper, we explore the potential of BabyLMs for low-resource languages, using the isiXhosa language as a case study. We pretrain two BabyLM architectures, ELC-BERT and MLSM, on an isiXhosa corpus. They outperform a vanilla pretrained model on POS tagging and NER, achieving notable gains (+3.2 F1) for the latter. In some instances, the BabyLMs even outperform XLM-R. Our findings show that data-efficient models are viable for low-resource languages, but highlight the continued importance, and lack of, high-quality pretraining data. Finally, we visually analyse how BabyLM architectures encode isiXhosa.

The interplay between cognition and gaming, notably through educational games enhancing cognitive skills, has garnered significant attention in recent years. This research introduces the CogSimulator, a novel algorithm for simulating user cognition in small-group settings with minimal data, as the educational game Wordle exemplifies. The CogSimulator employs Wasserstein-1 distance and coordinates search optimization for hyperparameter tuning, enabling precise few-shot predictions in new game scenarios. Comparative experiments with the Wordle dataset illustrate that our model surpasses most conventional machine learning models in mean Wasserstein-1 distance, mean squared error, and mean accuracy, showcasing its efficacy in cognitive enhancement through tailored game design.

Large Language Models (LLMs) demonstrate promising capabilities in solving simple scientific problems but often produce hallucinations for complex ones. While integrating LLMs with tools can increase reliability, this approach typically results in over-reliance on tools, diminishing the model's ability to solve simple problems through basic reasoning. In contrast, human experts first assess problem complexity using domain knowledge before choosing an appropriate solution approach. Inspired by this human problem-solving process, we propose a novel two-component fine-tuning method. In the first component World Knowledge Distillation (WKD), LLMs learn directly from solutions generated using tool's information to internalize domain knowledge. In the second component Tool Usage Adaptation (TUA), we partition problems into easy and hard categories based on the model's direct answering accuracy. While maintaining the same alignment target for easy problems as in WKD, we train the model to intelligently switch to tool usage for more challenging problems. We validate our method on six scientific benchmark datasets, spanning mathematics, climate science and epidemiology. On average, our models demonstrate a 28.18% improvement in answer accuracy and a 13.89% increase in tool usage precision across all datasets, surpassing state-of-the-art models including GPT-4o and Claude-3.5.

Air quality has become one of the most important public health issues in Africa. Poor air quality kills more people globally every year than HIV, TB and malaria combined. Air pollution makes people less productive because they get headaches and feel tired. India, for example, has poor air quality. The impact of India's poor air quality on its gross domestic product is about US 100 billion every year.

Large vision-language models (VLMs), such as CLIP, have become foundational, demonstrating remarkable success across a variety of downstream tasks. Despite their advantages, these models, akin to other foundational systems, inherit biases from the disproportionate distribution of real-world data, leading to misconceptions about the actual environment. Prevalent datasets like ImageNet are often riddled with non-causal, spurious correlations that can diminish VLM performance in scenarios where these contextual elements are absent. This study presents an investigation into how a simple linear probe can effectively distill task-specific core features from CLIP's embedding for downstream applications. Our analysis reveals that the CLIP text representations are often tainted by spurious correlations, inherited in the biased pre-training dataset. Empirical evidence suggests that relying on visual representations from CLIP, as opposed to text embedding, is more practical to refine the skewed perceptions in VLMs, emphasizing the superior utility of visual representations in overcoming embedded biases. Our codes will be available in here.

Large language models (LLMs) often generate content that contains factual errors when responding to fact-seeking prompts on open-ended topics. To benchmark a model's long-form factuality in open domains, we first use GPT-4 to generate LongFact, a prompt set comprising thousands of questions spanning 38 topics. We then propose that LLM agents can be used as automated evaluators for long-form factuality through a method which we call Search-Augmented Factuality Evaluator (SAFE). SAFE utilizes an LLM to break down a long-form response into a set of individual facts and to evaluate the accuracy of each fact using a multi-step reasoning process comprising sending search queries to Google Search and determining whether a fact is supported by the search results. Furthermore, we propose extending F1 score as an aggregated metric for long-form factuality. To do so, we balance the percentage of supported facts in a response (precision) with the percentage of provided facts relative to a hyperparameter representing a user's preferred response length (recall). Empirically, we demonstrate that LLM agents can outperform crowdsourced human annotators - on a set of ~16k individual facts, SAFE agrees with crowdsourced human annotators 72% of the time, and on a random subset of 100 disagreement cases, SAFE wins 76% of the time. At the same time, SAFE is more than 20 times cheaper than human annotators. We also benchmark thirteen language models on LongFact across four model families (Gemini, GPT, Claude, and PaLM-2), finding that larger language models generally achieve better long-form factuality. LongFact, SAFE, and all experimental code are available at https://github.com/google-deepmind/long-form-factuality.