Lucas Dixon

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.

Kelvin Guu, Albert Webson, Ellie Pavlick et al. · deepmind

Training data attribution (TDA) methods offer to trace a model's prediction on any given example back to specific influential training examples. Existing approaches do so by assigning a scalar influence score to each training example, under a simplifying assumption that influence is additive. But in reality, we observe that training examples interact in highly non-additive ways due to factors such as inter-example redundancy, training order, and curriculum learning effects. To study such interactions, we propose Simfluence, a new paradigm for TDA where the goal is not to produce a single influence score per example, but instead a training run simulator: the user asks, ``If my model had trained on example $z_1$, then $z_2$, ..., then $z_n$, how would it behave on $z_{test}$?''; the simulator should then output a simulated training run, which is a time series predicting the loss on $z_{test}$ at every step of the simulated run. This enables users to answer counterfactual questions about what their model would have learned under different training curricula, and to directly see where in training that learning would occur. We present a simulator, Simfluence-Linear, that captures non-additive interactions and is often able to predict the spiky trajectory of individual example losses with surprising fidelity. Furthermore, we show that existing TDA methods such as TracIn and influence functions can be viewed as special cases of Simfluence-Linear. This enables us to directly compare methods in terms of their simulation accuracy, subsuming several prior TDA approaches to evaluation. In experiments on large language model (LLM) fine-tuning, we show that our method predicts loss trajectories with much higher accuracy than existing TDA methods (doubling Spearman's correlation and reducing mean-squared error by 75%) across several tasks, models, and training methods.

Michael Terry, Chinmay Kulkarni, Martin Wattenberg et al. · deepmind

Modern AI enables a high-level, declarative form of interaction: Users describe the intended outcome they wish an AI to produce, but do not actually create the outcome themselves. In contrast, in traditional user interfaces, users invoke specific operations to create the desired outcome. This paper revisits the basic input-output interaction cycle in light of this declarative style of interaction, and connects concepts in AI alignment to define three objectives for interactive alignment of AI: specification alignment (aligning on what to do), process alignment (aligning on how to do it), and evaluation alignment (assisting users in verifying and understanding what was produced). Using existing systems as examples, we show how these user-centered views of AI alignment can be used descriptively, prescriptively, and as an evaluative aid.

Scott Sanner, Krisztian Balog, Filip Radlinski et al.

Traditional recommender systems leverage users' item preference history to recommend novel content that users may like. However, modern dialog interfaces that allow users to express language-based preferences offer a fundamentally different modality for preference input. Inspired by recent successes of prompting paradigms for large language models (LLMs), we study their use for making recommendations from both item-based and language-based preferences in comparison to state-of-the-art item-based collaborative filtering (CF) methods. To support this investigation, we collect a new dataset consisting of both item-based and language-based preferences elicited from users along with their ratings on a variety of (biased) recommended items and (unbiased) random items. Among numerous experimental results, we find that LLMs provide competitive recommendation performance for pure language-based preferences (no item preferences) in the near cold-start case in comparison to item-based CF methods, despite having no supervised training for this specific task (zero-shot) or only a few labels (few-shot). This is particularly promising as language-based preference representations are more explainable and scrutable than item-based or vector-based representations.

Maximilian Mozes, Jessica Hoffmann, Katrin Tomanek et al.

Text-based safety classifiers are widely used for content moderation and increasingly to tune generative language model behavior - a topic of growing concern for the safety of digital assistants and chatbots. However, different policies require different classifiers, and safety policies themselves improve from iteration and adaptation. This paper introduces and evaluates methods for agile text classification, whereby classifiers are trained using small, targeted datasets that can be quickly developed for a particular policy. Experimenting with 7 datasets from three safety-related domains, comprising 15 annotation schemes, led to our key finding: prompt-tuning large language models, like PaLM 62B, with a labeled dataset of as few as 80 examples can achieve state-of-the-art performance. We argue that this enables a paradigm shift for text classification, especially for models supporting safer online discourse. Instead of collecting millions of examples to attempt to create universal safety classifiers over months or years, classifiers could be tuned using small datasets, created by individuals or small organizations, tailored for specific use cases, and iterated on and adapted in the time-span of a day.

Léo Laugier, Raghuram Vadapalli, Thomas Bonald et al.

This paper explores a novel application of textual semantic similarity to user-preference representation for rating prediction. The approach represents a user's preferences as a graph of textual snippets from review text, where the edges are defined by semantic similarity. This textual, memory-based approach to rating prediction enables review-based explanations for recommendations. The method is evaluated quantitatively, highlighting that leveraging text in this way outperforms both strong memory-based and model-based collaborative filtering baselines.

Shayegan Omidshafiei, Andrei Kapishnikov, Yannick Assogba et al.

Each year, expert-level performance is attained in increasingly-complex multiagent domains, where notable examples include Go, Poker, and StarCraft II. This rapid progression is accompanied by a commensurate need to better understand how such agents attain this performance, to enable their safe deployment, identify limitations, and reveal potential means of improving them. In this paper we take a step back from performance-focused multiagent learning, and instead turn our attention towards agent behavior analysis. We introduce a model-agnostic method for discovery of behavior clusters in multiagent domains, using variational inference to learn a hierarchy of behaviors at the joint and local agent levels. Our framework makes no assumption about agents' underlying learning algorithms, does not require access to their latent states or policies, and is trained using only offline observational data. We illustrate the effectiveness of our method for enabling the coupled understanding of behaviors at the joint and local agent level, detection of behavior changepoints throughout training, discovery of core behavioral concepts, demonstrate the approach's scalability to a high-dimensional multiagent MuJoCo control domain, and also illustrate that the approach can disentangle previously-trained policies in OpenAI's hide-and-seek domain.

Maximilian Mozes, Tolga Bolukbasi, Ann Yuan et al.

Pretrained large language models (LLMs) are able to solve a wide variety of tasks through transfer learning. Various explainability methods have been developed to investigate their decision making process. TracIn (Pruthi et al., 2020) is one such gradient-based method which explains model inferences based on the influence of training examples. In this paper, we explore the use of TracIn to improve model performance in the parameter-efficient tuning (PET) setting. We develop conversational safety classifiers via the prompt-tuning PET method and show how the unique characteristics of the PET regime enable TracIn to identify the cause for certain misclassifications by LLMs. We develop a new methodology for using gradient-based explainability techniques to improve model performance, G-BAIR: gradient-based automated iterative recovery. We show that G-BAIR can recover LLM performance on benchmarks after manually corrupting training labels. This suggests that influence methods like TracIn can be used to automatically perform data cleaning, and introduces the potential for interactive debugging and relabeling for PET-based transfer learning methods.

Alex Bäuerle, Adam Connors, Alexander Novikov et al.

Artificial intelligence offers powerful new tools for scientific discovery, but the interaction paradigms required to effectively harness these systems remain underexplored. In this paper, we present findings from a formative user study with 11 expert mathematicians who used AlphaEvolve, an evolutionary coding agent, to tackle advanced problems in their fields of expertise. We identify and characterize a distinct workflow we term intentmaking, the iterative process of discovering, defining, and refining one's experimental goals through active system interaction. We frame this as a natural extension to sensemaking, the cognitive process of building an understanding of complex or novel data. We suggest that users enter a cycle of intentmaking (defining and updating their experiment) and sensemaking (interpreting the results) which repeats many times during the course of an investigation. Our documentation of these themes suggests an approach to designing AI tools for scientific discovery that goes beyond the existing question/answer model of many current systems, treating them as collaborative instruments rather than opaque black-box assistants.

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.

Gemma Team, Thomas Mesnard, Cassidy Hardin et al. · deepmind

This work introduces Gemma, a family of lightweight, state-of-the art open models built from the research and technology used to create Gemini models. Gemma models demonstrate strong performance across academic benchmarks for language understanding, reasoning, and safety. We release two sizes of models (2 billion and 7 billion parameters), and provide both pretrained and fine-tuned checkpoints. Gemma outperforms similarly sized open models on 11 out of 18 text-based tasks, and we present comprehensive evaluations of safety and responsibility aspects of the models, alongside a detailed description of model development. We believe the responsible release of LLMs is critical for improving the safety of frontier models, and for enabling the next wave of LLM innovations.

Ian Tenney, Ryan Mullins, Bin Du et al. · deepmind

We present Sequence Salience, a visual tool for interactive prompt debugging with input salience methods. Sequence Salience builds on widely used salience methods for text classification and single-token prediction, and extends this to a system tailored for debugging complex LLM prompts. Our system is well-suited for long texts, and expands on previous work by 1) providing controllable aggregation of token-level salience to the word, sentence, or paragraph level, making salience over long inputs tractable; and 2) supporting rapid iteration where practitioners can act on salience results, refine prompts, and run salience on the new output. We include case studies showing how Sequence Salience can help practitioners work with several complex prompting strategies, including few-shot, chain-of-thought, and constitutional principles. Sequence Salience is built on the Learning Interpretability Tool, an open-source platform for ML model visualizations, and code, notebooks, and tutorials are available at http://goo.gle/sequence-salience.

Ann Yuan, Asma Ghandeharioun, Carter Blum et al.

While existing evaluations of large language models (LLMs) measure deception rates, the underlying conditions that give rise to deceptive behavior are poorly understood. We investigate this question using a novel dataset of realistic moral trade-offs where honesty incurs variable costs. Contrary to humans, who tend to become less honest given time to deliberate (Capraro, 2017; Capraro et al., 2019), we find that reasoning consistently increases honesty across scales and for several LLM families. This effect is not only a function of the reasoning content, as reasoning traces are often poor predictors of final behaviors. Rather, we show that the underlying geometry of the representational space itself contributes to the effect. Namely, we observe that deceptive regions within this space are metastable: deceptive answers are more easily destabilized by input paraphrasing, output resampling, and activation noise than honest ones. We interpret the effect of reasoning in this vein: generating deliberative tokens as part of moral reasoning entails the traversal of a biased representational space, ultimately nudging the model toward its more stable, honest defaults.

Maxime Guigon, Lucas Dixon, Michaël E. Sander

Knowledge Distillation (KD) is a critical tool for training Large Language Models (LLMs), yet the majority of research focuses on approaches that rely solely on output logits, neglecting semantic information in the teacher's intermediate representations. While Hidden Layer Distillation (HLD) showed potential for encoder architectures, its application to decoder-only pre-training at scale remains largely unexplored. Through compute-controlled experiments, we benchmark HLD against logit-based KD and self-supervised baselines with Gemma3 3.4B as teacher and 123M and 735M students trained on up to 168B tokens from the C4 dataset. Our experiments show that HLD does not consistently outperform standard KD on downstream evaluation tasks. Nevertheless, we show that HLD can yield a systematic perplexity gain over KD across all shared-hyperparameter configurations, suggesting that a latent signal can be extracted, but a breakthrough may be needed for it to play a more significant role in LLM pre-training.

Asma Ghandeharioun, Avi Caciularu, Adam Pearce et al. · deepmind

Understanding the internal representations of large language models (LLMs) can help explain models' behavior and verify their alignment with human values. Given the capabilities of LLMs in generating human-understandable text, we propose leveraging the model itself to explain its internal representations in natural language. We introduce a framework called Patchscopes and show how it can be used to answer a wide range of questions about an LLM's computation. We show that many prior interpretability methods based on projecting representations into the vocabulary space and intervening on the LLM computation can be viewed as instances of this framework. Moreover, several of their shortcomings such as failure in inspecting early layers or lack of expressivity can be mitigated by Patchscopes. Beyond unifying prior inspection techniques, Patchscopes also opens up new possibilities such as using a more capable model to explain the representations of a smaller model, and multihop reasoning error correction.

Crystal Qian, Vivian Tsai, Michael Behr et al.

Social and behavioral scientists increasingly aim to study how humans interact, collaborate, and make decisions alongside artificial intelligence. However, the experimental infrastructure for such work remains underdeveloped: (1) few platforms support real-time, multi-party studies at scale; (2) most deployments require bespoke engineering, limiting replicability and accessibility, and (3) existing tools do not treat AI agents as first-class participants. We present Deliberate Lab, an open-source platform for large-scale, real-time behavioral experiments that supports both human participants and large language model (LLM)-based agents. We report on a 12-month public deployment of the platform (N=88 experimenters, N=9195 experiment participants), analyzing usage patterns and workflows. Case studies and usage scenarios are aggregated from platform users, complemented by in-depth interviews with select experimenters. By lowering technical barriers and standardizing support for hybrid human-AI experimentation, Deliberate Lab expands the methodological repertoire for studying collective decision-making and human-centered AI.

Tianlin Liu, Shangmin Guo, Leonardo Bianco et al.

Aligning language models with human preferences is crucial for reducing errors and biases in these models. Alignment techniques, such as reinforcement learning from human feedback (RLHF), are typically cast as optimizing a tradeoff between human preference rewards and a proximity regularization term that encourages staying close to the unaligned model. Selecting an appropriate level of regularization is critical: insufficient regularization can lead to reduced model capabilities due to reward hacking, whereas excessive regularization hinders alignment. Traditional methods for finding the optimal regularization level require retraining multiple models with varying regularization strengths. This process, however, is resource-intensive, especially for large models. To address this challenge, we propose decoding-time realignment (DeRa), a simple method to explore and evaluate different regularization strengths in aligned models without retraining. DeRa enables control over the degree of alignment, allowing users to smoothly transition between unaligned and aligned models. It also enhances the efficiency of hyperparameter tuning by enabling the identification of effective regularization strengths using a validation dataset.

Minsuk Kahng, Ian Tenney, Mahima Pushkarna et al. · deepmind

Automatic side-by-side evaluation has emerged as a promising approach to evaluating the quality of responses from large language models (LLMs). However, analyzing the results from this evaluation approach raises scalability and interpretability challenges. In this paper, we present LLM Comparator, a novel visual analytics tool for interactively analyzing results from automatic side-by-side evaluation. The tool supports interactive workflows for users to understand when and why a model performs better or worse than a baseline model, and how the responses from two models are qualitatively different. We iteratively designed and developed the tool by closely working with researchers and engineers at a large technology company. This paper details the user challenges we identified, the design and development of the tool, and an observational study with participants who regularly evaluate their models.

Dan Friedman, Andrew Lampinen, Lucas Dixon et al. · princeton

A common method to study deep learning systems is to use simplified model representations--for example, using singular value decomposition to visualize the model's hidden states in a lower dimensional space. This approach assumes that the results of these simplifications are faithful to the original model. Here, we illustrate an important caveat to this assumption: even if the simplified representations can accurately approximate the full model on the training set, they may fail to accurately capture the model's behavior out of distribution. We illustrate this by training Transformer models on controlled datasets with systematic generalization splits, including the Dyck balanced-parenthesis languages and a code completion task. We simplify these models using tools like dimensionality reduction and clustering, and then explicitly test how these simplified proxies match the behavior of the original model. We find consistent generalization gaps: cases in which the simplified proxies are more faithful to the original model on the in-distribution evaluations and less faithful on various tests of systematic generalization. This includes cases where the original model generalizes systematically but the simplified proxies fail, and cases where the simplified proxies generalize better. Together, our results raise questions about the extent to which mechanistic interpretations derived using tools like SVD can reliably predict what a model will do in novel situations.

Tyler A. Chang, Dheeraj Rajagopal, Tolga Bolukbasi et al. · cmu

Training data attribution (TDA) methods aim to attribute model outputs back to specific training examples, and the application of these methods to large language model (LLM) outputs could significantly advance model transparency and data curation. However, it has been challenging to date to apply these methods to the full scale of LLM pretraining. In this paper, we refine existing gradient-based methods to work effectively at scale, allowing us to retrieve influential examples for an 8B-parameter language model from a pretraining corpus of over 160B tokens with no need for subsampling or pre-filtering. Our method combines several techniques, including optimizer state correction, a task-specific Hessian approximation, and normalized encodings, which we find to be critical for performance at scale. In quantitative evaluations on a fact tracing task, our method performs best at identifying examples that influence model predictions, but classical, model-agnostic retrieval methods such as BM25 still perform better at finding passages which explicitly contain relevant facts. These results demonstrate a misalignment between factual *attribution* and causal *influence*. With increasing model size and training tokens, we find that influence more closely aligns with factual attribution. Finally, we examine different types of examples identified as influential by our method, finding that while many directly entail a particular fact, others support the same output by reinforcing priors on relation types, common entities, and names. We release our prompt set and model outputs, along with a web-based visualization tool to explore influential examples for factual predictions, commonsense reasoning, arithmetic, and open-ended generation for an 8B-parameter LLM.

Tyler A. Chang, Katrin Tomanek, Jessica Hoffmann et al.

We explore a strategy to handle controversial topics in LLM-based chatbots based on Wikipedia's Neutral Point of View (NPOV) principle: acknowledge the absence of a single true answer and surface multiple perspectives. We frame this as retrieval augmented generation, where perspectives are retrieved from a knowledge base and the LLM is tasked with generating a fluent and faithful response from the given perspectives. As a starting point, we use a deterministic retrieval system and then focus on common LLM failure modes that arise during this approach to text generation, namely hallucination and coverage errors. We propose and evaluate three methods to detect such errors based on (1) word-overlap, (2) salience, and (3) LLM-based classifiers. Our results demonstrate that LLM-based classifiers, even when trained only on synthetic errors, achieve high error detection performance, with ROC AUC scores of 95.3% for hallucination and 90.5% for coverage error detection on unambiguous error cases. We show that when no training data is available, our other methods still yield good results on hallucination (84.0%) and coverage error (85.2%) detection.

Hakim Sidahmed, Samrat Phatale, Alex Hutcheson et al. · deepmind

While Reinforcement Learning from Human Feedback (RLHF) effectively aligns pretrained Large Language and Vision-Language Models (LLMs, and VLMs) with human preferences, its computational cost and complexity hamper its wider adoption. To alleviate some of the computational burden of fine-tuning, parameter efficient methods, like LoRA were introduced. In this work, we empirically evaluate the setup of Parameter Efficient Reinforcement Learning from Human Feedback (PE-RLHF) that leverages LoRA fine-tuning for Reward Modeling, and Reinforcement Learning. We benchmark the PE-RLHF setup on six diverse datasets spanning summarization, harmless/helpful response generation, UI automation, and visual question answering in terms of effectiveness of the trained models, and the training resources required. Our findings show, for the first time, that PE-RLHF achieves comparable performance to RLHF, while significantly reducing training time (up to 90% faster for reward models, and 30% faster for RL), and memory footprint (up to 50% reduction for reward models, and 27% for RL). We provide comprehensive ablations across LoRA ranks, and model sizes for both reward modeling and reinforcement learning. By mitigating the computational burden associated with RLHF, we push for a broader adoption of PE-RLHF as an alignment technique for LLMs and VLMs.

Carter Blum, Katja Filippova, Ann Yuan et al. · deepmind

Large language models (LLMs) struggle with cross-lingual knowledge transfer: they hallucinate when asked in one language about facts expressed in a different language during training. This work introduces a controlled setting to study the causes and dynamics of this phenomenon by training small Transformer models from scratch on synthetic multilingual datasets. We identify a learning phase wherein a model develops either separate or unified representations of the same facts across languages, and show that unification is essential for cross-lingual transfer. We also show that the degree of unification depends on mutual information between facts and training data language, and on how easy it is to extract that language. Based on these insights, we develop methods to modulate the level of cross-lingual transfer by manipulating data distribution and tokenization, and we introduce metrics and visualizations to formally characterize their effects on unification. Our work shows how controlled settings can shed light on pre-training dynamics and suggests new directions for improving cross-lingual transfer in LLMs.

Jessica Hoffmann, Christiane Ahlheim, Zac Yu et al.

The paper shows that parameter-efficient reinforcement learning (PE-RL) is a highly effective training regime to improve large language models' (LLMs) ability to answer queries on sensitive topics with a Neutral Point of View (NPOV), i.e. to provide significantly more informative, diverse and impartial answers. This is shown by evaluating PE-RL and multiple strong baselines-including LoRA finetuning (strongest baseline), SFT and RLHF. PE-RL not only improves on overall NPOV quality compared to the strongest baseline ($97.06\%\rightarrow 99.08\%$), but also scores much higher on features linguists identify as key to separating sufficient answers from "great'' answers ($60.25\%\rightarrow 85.21\%$ for presence of supportive details, $68.74\%\rightarrow 91.43\%$ for absence of oversimplification). A qualitative analysis corroborates this. Moreover, our evaluation also finds a key property of PE-RL for this task: unlike methods that update all parameters, it generalises out of topic. Finally, to enable further studies we also release the dataset, SHQ-NPOV, and provide a methodology to create such datasets through iterative rounds of human peer-critique and annotator training.

Crystal Qian, Aaron Parisi, Clémentine Bouleau et al.

As large language models (LLMs) are increasingly used to model and augment collective decision-making, it is critical to examine their alignment with human social reasoning. We present an empirical framework for assessing collective alignment, in contrast to prior work on the individual level. Using the Lost at Sea social psychology task, we conduct a large-scale online experiment (N=748), randomly assigning groups to leader elections with either visible demographic attributes (e.g. name, gender) or pseudonymous aliases. We then simulate matched LLM groups conditioned on the human data, benchmarking Gemini 2.5, GPT 4.1, Claude Haiku 3.5, and Gemma 3. LLM behaviors diverge: some mirror human biases; others mask these biases and attempt to compensate for them. We empirically demonstrate that human-AI alignment in collective reasoning depends on context, cues, and model-specific inductive biases. Understanding how LLMs align with collective human behavior is critical to advancing socially-aligned AI, and demands dynamic benchmarks that capture the complexities of collective reasoning.

Asma Ghandeharioun, Ann Yuan, Marius Guerard et al.

Despite investments in improving model safety, studies show that misaligned capabilities remain latent in safety-tuned models. In this work, we shed light on the mechanics of this phenomenon. First, we show that even when model generations are safe, harmful content can persist in hidden representations and can be extracted by decoding from earlier layers. Then, we show that whether the model divulges such content depends significantly on its perception of who it is talking to, which we refer to as user persona. In fact, we find manipulating user persona to be even more effective for eliciting harmful content than direct attempts to control model refusal. We study both natural language prompting and activation steering as control methods and show that activation steering is significantly more effective at bypassing safety filters. We investigate why certain personas break model safeguards and find that they enable the model to form more charitable interpretations of otherwise dangerous queries. Finally, we show we can predict a persona's effect on refusal given only the geometry of its steering vector.

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.

Konstantina Christakopoulou, Alberto Lalama, Cj Adams et al.

Large language models (LLMs) have shown impressive capabilities in natural language understanding and generation. Their potential for deeper user understanding and improved personalized user experience on recommendation platforms is, however, largely untapped. This paper aims to address this gap. Recommender systems today capture users' interests through encoding their historical activities on the platforms. The generated user representations are hard to examine or interpret. On the other hand, if we were to ask people about interests they pursue in their life, they might talk about their hobbies, like I just started learning the ukulele, or their relaxation routines, e.g., I like to watch Saturday Night Live, or I want to plant a vertical garden. We argue, and demonstrate through extensive experiments, that LLMs as foundation models can reason through user activities, and describe their interests in nuanced and interesting ways, similar to how a human would. We define interest journeys as the persistent and overarching user interests, in other words, the non-transient ones. These are the interests that we believe will benefit most from the nuanced and personalized descriptions. We introduce a framework in which we first perform personalized extraction of interest journeys, and then summarize the extracted journeys via LLMs, using techniques like few-shot prompting, prompt-tuning and fine-tuning. Together, our results in prompting LLMs to name extracted user journeys in a large-scale industrial platform demonstrate great potential of these models in providing deeper, more interpretable, and controllable user understanding. We believe LLM powered user understanding can be a stepping stone to entirely new user experiences on recommendation platforms that are journey-aware, assistive, and enabling frictionless conversation down the line.

Nan Du, Yanping Huang, Andrew M. Dai et al.

Scaling language models with more data, compute and parameters has driven significant progress in natural language processing. For example, thanks to scaling, GPT-3 was able to achieve strong results on in-context learning tasks. However, training these large dense models requires significant amounts of computing resources. In this paper, we propose and develop a family of language models named GLaM (Generalist Language Model), which uses a sparsely activated mixture-of-experts architecture to scale the model capacity while also incurring substantially less training cost compared to dense variants. The largest GLaM has 1.2 trillion parameters, which is approximately 7x larger than GPT-3. It consumes only 1/3 of the energy used to train GPT-3 and requires half of the computation flops for inference, while still achieving better overall zero-shot and one-shot performance across 29 NLP tasks.

Alexandros Xenos, John Pavlopoulos, Ion Androutsopoulos et al.

User posts whose perceived toxicity depends on the conversational context are rare in current toxicity detection datasets. Hence, toxicity detectors trained on existing datasets will also tend to disregard context, making the detection of context-sensitive toxicity harder when it does occur. We construct and publicly release a dataset of 10,000 posts with two kinds of toxicity labels: (i) annotators considered each post with the previous one as context; and (ii) annotators had no additional context. Based on this, we introduce a new task, context sensitivity estimation, which aims to identify posts whose perceived toxicity changes if the context (previous post) is also considered. We then evaluate machine learning systems on this task, showing that classifiers of practical quality can be developed, and we show that data augmentation with knowledge distillation can improve the performance further. Such systems could be used to enhance toxicity detection datasets with more context-dependent posts, or to suggest when moderators should consider the parent posts, which often may be unnecessary and may otherwise introduce significant additional cost.

Federico López, Martin Scholz, Jessica Yung et al.

This paper introduces a simple and effective form of data augmentation for recommender systems. A paraphrase similarity model is applied to widely available textual data, such as reviews and product descriptions, yielding new semantic relations that are added to the user-item graph. This increases the density of the graph without needing further labeled data. The data augmentation is evaluated on a variety of recommendation algorithms, using Euclidean, hyperbolic, and complex spaces, and over three categories of Amazon product reviews with differing characteristics. Results show that the data augmentation technique provides significant improvements to all types of models, with the most pronounced gains for knowledge graph-based recommenders, particularly in cold-start settings, leading to state-of-the-art performance.

Leo Laugier, John Pavlopoulos, Jeffrey Sorensen et al.

Platforms that support online commentary, from social networks to news sites, are increasingly leveraging machine learning to assist their moderation efforts. But this process does not typically provide feedback to the author that would help them contribute according to the community guidelines. This is prohibitively time-consuming for human moderators to do, and computational approaches are still nascent. This work focuses on models that can help suggest rephrasings of toxic comments in a more civil manner. Inspired by recent progress in unpaired sequence-to-sequence tasks, a self-supervised learning model is introduced, called CAE-T5. CAE-T5 employs a pre-trained text-to-text transformer, which is fine tuned with a denoising and cyclic auto-encoder loss. Experimenting with the largest toxicity detection dataset to date (Civil Comments) our model generates sentences that are more fluent and better at preserving the initial content compared to earlier text style transfer systems which we compare with using several scoring systems and human evaluation.

Ilan Price, Jordan Gifford-Moore, Jory Fleming et al.

We present a new dataset of approximately 44000 comments labeled by crowdworkers. Each comment is labelled as either 'healthy' or 'unhealthy', in addition to binary labels for the presence of six potentially 'unhealthy' sub-attributes: (1) hostile; (2) antagonistic, insulting, provocative or trolling; (3) dismissive; (4) condescending or patronising; (5) sarcastic; and/or (6) an unfair generalisation. Each label also has an associated confidence score. We argue that there is a need for datasets which enable research based on a broad notion of 'unhealthy online conversation'. We build this typology to encompass a substantial proportion of the individual comments which contribute to unhealthy online conversation. For some of these attributes, this is the first publicly available dataset of this scale. We explore the quality of the dataset, present some summary statistics and initial models to illustrate the utility of this data, and highlight limitations and directions for further research.

John Pavlopoulos, Jeffrey Sorensen, Lucas Dixon et al.

Moderation is crucial to promoting healthy on-line discussions. Although several `toxicity' detection datasets and models have been published, most of them ignore the context of the posts, implicitly assuming that comments maybe judged independently. We investigate this assumption by focusing on two questions: (a) does context affect the human judgement, and (b) does conditioning on context improve performance of toxicity detection systems? We experiment with Wikipedia conversations, limiting the notion of context to the previous post in the thread and the discussion title. We find that context can both amplify or mitigate the perceived toxicity of posts. Moreover, a small but significant subset of manually labeled posts (5% in one of our experiments) end up having the opposite toxicity labels if the annotators are not provided with context. Surprisingly, we also find no evidence that context actually improves the performance of toxicity classifiers, having tried a range of classifiers and mechanisms to make them context aware. This points to the need for larger datasets of comments annotated in context. We make our code and data publicly available.

Varada Kolhatkar, Nithum Thain, Jeffrey Sorensen et al.

We introduce the Constructive Comments Corpus (C3), comprised of 12,000 annotated news comments, intended to help build new tools for online communities to improve the quality of their discussions. We define constructive comments as high-quality comments that make a contribution to the conversation. We explain the crowd worker annotation scheme and define a taxonomy of sub-characteristics of constructiveness. The quality of the annotation scheme and the resulting dataset is evaluated using measurements of inter-annotator agreement, expert assessment of a sample, and by the constructiveness sub-characteristics, which we show provide a proxy for the general constructiveness concept. We provide models for constructiveness trained on C3 using both feature-based and a variety of deep learning approaches and demonstrate that these models capture general rather than topic- or domain-specific characteristics of constructiveness, through domain adaptation experiments. We examine the role that length plays in our models, as comment length could be easily gamed if models depend heavily upon this feature. By examining the errors made by each model and their distribution by length, we show that the best performing models are less correlated with comment length.The constructiveness corpus and our experiments pave the way for a moderation tool focused on promoting comments that make a contribution, rather than only filtering out undesirable content.

Daniel Borkan, Lucas Dixon, Jeffrey Sorensen et al.

Unintended bias in Machine Learning can manifest as systemic differences in performance for different demographic groups, potentially compounding existing challenges to fairness in society at large. In this paper, we introduce a suite of threshold-agnostic metrics that provide a nuanced view of this unintended bias, by considering the various ways that a classifier's score distribution can vary across designated groups. We also introduce a large new test set of online comments with crowd-sourced annotations for identity references. We use this to show how our metrics can be used to find new and potentially subtle unintended bias in existing public models.

Daniel Borkan, Lucas Dixon, John Li et al.

This report examines the Pinned AUC metric introduced and highlights some of its limitations. Pinned AUC provides a threshold-agnostic measure of unintended bias in a classification model, inspired by the ROC-AUC metric. However, as we highlight in this report, there are ways that the metric can obscure different kinds of unintended biases when the underlying class distributions on which bias is being measured are not carefully controlled.

Yiqing Hua, Cristian Danescu-Niculescu-Mizil, Dario Taraborelli et al.

We present a corpus that encompasses the complete history of conversations between contributors to Wikipedia, one of the largest online collaborative communities. By recording the intermediate states of conversations---including not only comments and replies, but also their modifications, deletions and restorations---this data offers an unprecedented view of online conversation. This level of detail supports new research questions pertaining to the process (and challenges) of large-scale online collaboration. We illustrate the corpus' potential with two case studies that highlight new perspectives on earlier work. First, we explore how a person's conversational behavior depends on how they relate to the discussion's venue. Second, we show that community moderation of toxic behavior happens at a higher rate than previously estimated. Finally the reconstruction framework is designed to be language agnostic, and we show that it can extract high quality conversational data in both Chinese and English.

Justine Zhang, Jonathan P. Chang, Cristian Danescu-Niculescu-Mizil et al.

One of the main challenges online social systems face is the prevalence of antisocial behavior, such as harassment and personal attacks. In this work, we introduce the task of predicting from the very start of a conversation whether it will get out of hand. As opposed to detecting undesirable behavior after the fact, this task aims to enable early, actionable prediction at a time when the conversation might still be salvaged. To this end, we develop a framework for capturing pragmatic devices---such as politeness strategies and rhetorical prompts---used to start a conversation, and analyze their relation to its future trajectory. Applying this framework in a controlled setting, we demonstrate the feasibility of detecting early warning signs of antisocial behavior in online discussions.

Ellery Wulczyn, Nithum Thain, Lucas Dixon

The damage personal attacks cause to online discourse motivates many platforms to try to curb the phenomenon. However, understanding the prevalence and impact of personal attacks in online platforms at scale remains surprisingly difficult. The contribution of this paper is to develop and illustrate a method that combines crowdsourcing and machine learning to analyze personal attacks at scale. We show an evaluation method for a classifier in terms of the aggregated number of crowd-workers it can approximate. We apply our methodology to English Wikipedia, generating a corpus of over 100k high quality human-labeled comments and 63M machine-labeled ones from a classifier that is as good as the aggregate of 3 crowd-workers, as measured by the area under the ROC curve and Spearman correlation. Using this corpus of machine-labeled scores, our methodology allows us to explore some of the open questions about the nature of online personal attacks. This reveals that the majority of personal attacks on Wikipedia are not the result of a few malicious users, nor primarily the consequence of allowing anonymous contributions from unregistered users.

Lucas Dixon, Thomas Ristenpart, Thomas Shrimpton

Internet censors seek ways to identify and block internet access to information they deem objectionable. Increasingly, censors deploy advanced networking tools such as deep-packet inspection (DPI) to identify such connections. In response, activists and academic researchers have developed and deployed network traffic obfuscation mechanisms. These apply specialized cryptographic tools to attempt to hide from DPI the true nature and content of connections. In this survey, we give an overview of network traffic obfuscation and its role in circumventing Internet censorship. We provide historical and technical background that motivates the need for obfuscation tools, and give an overview of approaches to obfuscation used by state of the art tools. We discuss the latest research on how censors might detect these efforts. We also describe the current challenges to censorship circumvention research and identify concrete ways for the community to address these challenges.