Chris Welty

Gemma Team, Morgane Riviere, Shreya Pathak et al. · deepmind

In this work, we introduce Gemma 2, a new addition to the Gemma family of lightweight, state-of-the-art open models, ranging in scale from 2 billion to 27 billion parameters. In this new version, we apply several known technical modifications to the Transformer architecture, such as interleaving local-global attentions (Beltagy et al., 2020a) and group-query attention (Ainslie et al., 2023). We also train the 2B and 9B models with knowledge distillation (Hinton et al., 2015) instead of next token prediction. The resulting models deliver the best performance for their size, and even offer competitive alternatives to models that are 2-3 times bigger. We release all our models to the community.

Gemini Team, Petko Georgiev, Ving Ian Lei et al. · deepmind, mila

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Deepak Pandita, Flip Korn, Chris Welty et al.

As generative AI models such as large language models (LLMs) become more pervasive, ensuring the safety, robustness, and overall trustworthiness of these systems is paramount. However, AI is currently facing a reproducibility crisis driven by unreliable evaluations and unrepeatable experimental results. While human raters are often used to assess models for utility and safety, they introduce divergent biases and subjective opinions into their annotations. Overcoming this variance is exceptionally challenging because very little data exists to study how experimental repeatability actually improves as the annotator pool grows. Standard evaluation practices typically rely on a small number of annotations per item (often 3 to 5) and lack the persistent rater identifiers necessary to model individual variance across items. In this work, we introduce a multi-level bootstrapping approach to realistically model annotator behavior. Leveraging datasets with a large number of ratings and persistent rater identifiers, we analyze the tradeoffs between the number of items ($N$) and the number of responses per item ($K$) required to achieve statistical significance.

Anca Dumitrache, Oana Inel, Lora Aroyo et al.

Typically crowdsourcing-based approaches to gather annotated data use inter-annotator agreement as a measure of quality. However, in many domains, there is ambiguity in the data, as well as a multitude of perspectives of the information examples. In this paper, we present ongoing work into the CrowdTruth metrics, that capture and interpret inter-annotator disagreement in crowdsourcing. The CrowdTruth metrics model the inter-dependency between the three main components of a crowdsourcing system -- worker, input data, and annotation. The goal of the metrics is to capture the degree of ambiguity in each of these three components. The metrics are available online at https://github.com/CrowdTruth/CrowdTruth-core .

Bertie Vidgen, Adarsh Agrawal, Ahmed M. Ahmed et al. · deepmind, oxford

This paper introduces v0.5 of the AI Safety Benchmark, which has been created by the MLCommons AI Safety Working Group. The AI Safety Benchmark has been designed to assess the safety risks of AI systems that use chat-tuned language models. We introduce a principled approach to specifying and constructing the benchmark, which for v0.5 covers only a single use case (an adult chatting to a general-purpose assistant in English), and a limited set of personas (i.e., typical users, malicious users, and vulnerable users). We created a new taxonomy of 13 hazard categories, of which 7 have tests in the v0.5 benchmark. We plan to release version 1.0 of the AI Safety Benchmark by the end of 2024. The v1.0 benchmark will provide meaningful insights into the safety of AI systems. However, the v0.5 benchmark should not be used to assess the safety of AI systems. We have sought to fully document the limitations, flaws, and challenges of v0.5. This release of v0.5 of the AI Safety Benchmark includes (1) a principled approach to specifying and constructing the benchmark, which comprises use cases, types of systems under test (SUTs), language and context, personas, tests, and test items; (2) a taxonomy of 13 hazard categories with definitions and subcategories; (3) tests for seven of the hazard categories, each comprising a unique set of test items, i.e., prompts. There are 43,090 test items in total, which we created with templates; (4) a grading system for AI systems against the benchmark; (5) an openly available platform, and downloadable tool, called ModelBench that can be used to evaluate the safety of AI systems on the benchmark; (6) an example evaluation report which benchmarks the performance of over a dozen openly available chat-tuned language models; (7) a test specification for the benchmark.

Deepak Pandita, Flip Korn, Chris Welty et al.

Reproducibility is a cornerstone of scientific validation and of the authority it confers on its results. Reproducibility in machine learning evaluations leads to greater trust, confidence, and value. However, the ground truth responses used in machine learning often necessarily come from humans, among whom disagreement is prevalent, and surprisingly little research has studied the impact of effectively ignoring disagreement in these responses, as is typically the case. One reason for the lack of research is that budgets for collecting human-annotated evaluation data are limited, and obtaining more samples from multiple annotators for each example greatly increases the per-item annotation costs. We investigate the trade-off between the number of items ($N$) and the number of responses per item ($K$) needed for reliable machine learning evaluation. We analyze a diverse collection of categorical datasets for which multiple annotations per item exist, and simulated distributions fit to these datasets, to determine the optimal $(N, K)$ configuration, given a fixed budget ($N \times K$), for collecting evaluation data and reliably comparing the performance of machine learning models. Our findings show, first, that accounting for human disagreement may come with $N \times K$ at no more than 1000 (and often much lower) for every dataset tested on at least one metric. Moreover, this minimal $N \times K$ almost always occurred for $K > 10$. Furthermore, the nature of the tradeoff between $K$ and $N$ -- or if one even existed -- depends on the evaluation metric, with metrics that are more sensitive to the full distribution of responses performing better at higher levels of $K$. Our methods can be used to help ML practitioners get more effective test data by finding the optimal metrics and number of items and annotations per item to collect to get the most reliability for their budget.

Christopher Homan, Flip Korn, Deepak Pandita et al.

A cornerstone of machine learning evaluation is the (often hidden) assumption that model and human responses are reliable enough to evaluate models against unitary, authoritative, ``gold standard'' data, via simple metrics such as accuracy, precision, and recall. The generative AI revolution would seem to explode this assumption, given the critical role stochastic inference plays. Yet, in spite of public demand for more transparency in AI -- along with strong evidence that humans are unreliable judges -- estimates of model reliability are conventionally based on, at most, a few output responses per input item. We adapt a method, previously used to evaluate the reliability of various metrics and estimators for machine learning evaluation, to determine whether an (existing or planned) dataset has enough responses per item to assure reliable null hypothesis statistical testing. We show that, for many common metrics, collecting even 5-10 responses per item (from each model and team of human evaluators) is not sufficient. We apply our methods to several of the very few extant gold standard test sets with multiple disaggregated responses per item and show that even these datasets lack enough responses per item. We show how our methods can help AI researchers make better decisions about how to collect data for AI evaluation.

Gemini Team, Rohan Anil, Sebastian Borgeaud et al.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.

Chris Welty, Praveen Paritosh, Lora Aroyo

In this paper we present the first steps towards hardening the science of measuring AI systems, by adopting metrology, the science of measurement and its application, and applying it to human (crowd) powered evaluations. We begin with the intuitive observation that evaluating the performance of an AI system is a form of measurement. In all other science and engineering disciplines, the devices used to measure are called instruments, and all measurements are recorded with respect to the characteristics of the instruments used. One does not report mass, speed, or length, for example, of a studied object without disclosing the precision (measurement variance) and resolution (smallest detectable change) of the instrument used. It is extremely common in the AI literature to compare the performance of two systems by using a crowd-sourced dataset as an instrument, but failing to report if the performance difference lies within the capability of that instrument to measure. To illustrate the adoption of metrology to benchmark datasets we use the word similarity benchmark WS353 and several previously published experiments that use it for evaluation.

Ananth Balashankar, Alyssa Lees, Chris Welty et al.

The potential for learned models to amplify existing societal biases has been broadly recognized. Fairness-aware classifier constraints, which apply equality metrics of performance across subgroups defined on sensitive attributes such as race and gender, seek to rectify inequity but can yield non-uniform degradation in performance for skewed datasets. In certain domains, imbalanced degradation of performance can yield another form of unintentional bias. In the spirit of constructing fairness-aware algorithms as societal imperative, we explore an alternative: Pareto-Efficient Fairness (PEF). Theoretically, we prove that PEF identifies the operating point on the Pareto curve of subgroup performances closest to the fairness hyperplane, maximizing multiple subgroup accuracy. Empirically we demonstrate that PEF outperforms by achieving Pareto levels in accuracy for all subgroups compared to strict fairness constraints in several UCI datasets.

Anca Dumitrache, Lora Aroyo, Chris Welty

We present a resource for the task of FrameNet semantic frame disambiguation of over 5,000 word-sentence pairs from the Wikipedia corpus. The annotations were collected using a novel crowdsourcing approach with multiple workers per sentence to capture inter-annotator disagreement. In contrast to the typical approach of attributing the best single frame to each word, we provide a list of frames with disagreement-based scores that express the confidence with which each frame applies to the word. This is based on the idea that inter-annotator disagreement is at least partly caused by ambiguity that is inherent to the text and frames. We have found many examples where the semantics of individual frames overlap sufficiently to make them acceptable alternatives for interpreting a sentence. We have argued that ignoring this ambiguity creates an overly arbitrary target for training and evaluating natural language processing systems - if humans cannot agree, why would we expect the correct answer from a machine to be any different? To process this data we also utilized an expanded lemma-set provided by the Framester system, which merges FN with WordNet to enhance coverage. Our dataset includes annotations of 1,000 sentence-word pairs whose lemmas are not part of FN. Finally we present metrics for evaluating frame disambiguation systems that account for ambiguity.

Anca Dumitrache, Oana Inel, Benjamin Timmermans et al.

The process of gathering ground truth data through human annotation is a major bottleneck in the use of information extraction methods for populating the Semantic Web. Crowdsourcing-based approaches are gaining popularity in the attempt to solve the issues related to volume of data and lack of annotators. Typically these practices use inter-annotator agreement as a measure of quality. However, in many domains, such as event detection, there is ambiguity in the data, as well as a multitude of perspectives of the information examples. We present an empirically derived methodology for efficiently gathering of ground truth data in a diverse set of use cases covering a variety of domains and annotation tasks. Central to our approach is the use of CrowdTruth metrics that capture inter-annotator disagreement. We show that measuring disagreement is essential for acquiring a high quality ground truth. We achieve this by comparing the quality of the data aggregated with CrowdTruth metrics with majority vote, over a set of diverse crowdsourcing tasks: Medical Relation Extraction, Twitter Event Identification, News Event Extraction and Sound Interpretation. We also show that an increased number of crowd workers leads to growth and stabilization in the quality of annotations, going against the usual practice of employing a small number of annotators.

Anca Dumitrache, Lora Aroyo, Chris Welty

Distant supervision is a popular method for performing relation extraction from text that is known to produce noisy labels. Most progress in relation extraction and classification has been made with crowdsourced corrections to distant-supervised labels, and there is evidence that indicates still more would be better. In this paper, we explore the problem of propagating human annotation signals gathered for open-domain relation classification through the CrowdTruth methodology for crowdsourcing, that captures ambiguity in annotations by measuring inter-annotator disagreement. Our approach propagates annotations to sentences that are similar in a low dimensional embedding space, expanding the number of labels by two orders of magnitude. Our experiments show significant improvement in a sentence-level multi-class relation classifier.

Anca Dumitrache, Lora Aroyo, Chris Welty

FrameNet is a computational linguistics resource composed of semantic frames, high-level concepts that represent the meanings of words. In this paper, we present an approach to gather frame disambiguation annotations in sentences using a crowdsourcing approach with multiple workers per sentence to capture inter-annotator disagreement. We perform an experiment over a set of 433 sentences annotated with frames from the FrameNet corpus, and show that the aggregated crowd annotations achieve an F1 score greater than 0.67 as compared to expert linguists. We highlight cases where the crowd annotation was correct even though the expert is in disagreement, arguing for the need to have multiple annotators per sentence. Most importantly, we examine cases in which crowd workers could not agree, and demonstrate that these cases exhibit ambiguity, either in the sentence, frame, or the task itself, and argue that collapsing such cases to a single, discrete truth value (i.e. correct or incorrect) is inappropriate, creating arbitrary targets for machine learning.

Anca Dumitrache, Lora Aroyo, Chris Welty

Distant supervision (DS) is a well-established method for relation extraction from text, based on the assumption that when a knowledge-base contains a relation between a term pair, then sentences that contain that pair are likely to express the relation. In this paper, we use the results of a crowdsourcing relation extraction task to identify two problems with DS data quality: the widely varying degree of false positives across different relations, and the observed causal connection between relations that are not considered by the DS method. The crowdsourcing data aggregation is performed using ambiguity-aware CrowdTruth metrics, that are used to capture and interpret inter-annotator disagreement. We also present preliminary results of using the crowd to enhance DS training data for a relation classification model, without requiring the crowd to annotate the entire set.

Anca Dumitrache, Lora Aroyo, Chris Welty

Cognitive computing systems require human labeled data for evaluation, and often for training. The standard practice used in gathering this data minimizes disagreement between annotators, and we have found this results in data that fails to account for the ambiguity inherent in language. We have proposed the CrowdTruth method for collecting ground truth through crowdsourcing, that reconsiders the role of people in machine learning based on the observation that disagreement between annotators provides a useful signal for phenomena such as ambiguity in the text. We report on using this method to build an annotated data set for medical relation extraction for the $cause$ and $treat$ relations, and how this data performed in a supervised training experiment. We demonstrate that by modeling ambiguity, labeled data gathered from crowd workers can (1) reach the level of quality of domain experts for this task while reducing the cost, and (2) provide better training data at scale than distant supervision. We further propose and validate new weighted measures for precision, recall, and F-measure, that account for ambiguity in both human and machine performance on this task.