We introduce Gemma 3, a multimodal addition to the Gemma family of lightweight open models, ranging in scale from 1 to 27 billion parameters. This version introduces vision understanding abilities, a wider coverage of languages and longer context - at least 128K tokens. We also change the architecture of the model to reduce the KV-cache memory that tends to explode with long context. This is achieved by increasing the ratio of local to global attention layers, and keeping the span on local attention short. The Gemma 3 models are trained with distillation and achieve superior performance to Gemma 2 for both pre-trained and instruction finetuned versions. In particular, our novel post-training recipe significantly improves the math, chat, instruction-following and multilingual abilities, making Gemma3-4B-IT competitive with Gemma2-27B-IT and Gemma3-27B-IT comparable to Gemini-1.5-Pro across benchmarks. We release all our models to the community.

While decoder-only large language models (LLMs) have shown impressive results, encoder-decoder models are still widely adopted in real-world applications for their inference efficiency and richer encoder representation. In this paper, we study a novel problem: adapting pretrained decoder-only LLMs to encoder-decoder, with the goal of leveraging the strengths of both approaches to achieve a more favorable quality-efficiency trade-off. We argue that adaptation not only enables inheriting the capability of decoder-only LLMs but also reduces the demand for computation compared to pretraining from scratch. We rigorously explore different pretraining objectives and parameter initialization/optimization techniques. Through extensive experiments based on Gemma 2 (2B and 9B) and a suite of newly pretrained mT5-sized models (up to 1.6B), we demonstrate the effectiveness of adaptation and the advantage of encoder-decoder LLMs. Under similar inference budget, encoder-decoder LLMs achieve comparable (often better) pretraining performance but substantially better finetuning performance than their decoder-only counterpart. For example, Gemma 2B-2B outperforms Gemma 2B by ∼7% after instruction tuning. Encoder-decoder adaptation also allows for flexible combination of different-sized models, where Gemma 9B-2B significantly surpasses Gemma 2B-2B by >3%. The adapted encoder representation also yields better results on SuperGLUE. We will release our checkpoints to facilitate future research.

We introduce RecurrentGemma, a family of open language models which uses Google's novel Griffin architecture. Griffin combines linear recurrences with local attention to achieve excellent performance on language. It has a fixed-sized state, which reduces memory use and enables efficient inference on long sequences. We provide two sizes of models, containing 2B and 9B parameters, and provide pre-trained and instruction tuned variants for both. Our models achieve comparable performance to similarly-sized Gemma baselines despite being trained on fewer tokens.

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.

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.

The Model AI Assignments session seeks to gather and disseminate the best assignment designs of the Artificial Intelligence (AI) Education community. Recognizing that assignments form the core of student learning experience, we here present abstracts of nine AI assignments from the 2020 session that are easily adoptable, playfully engaging, and flexible for a variety of instructor needs. Assignment specifications and supporting resources may be found at http://modelai.gettysburg.edu.

The Model AI Assignments session seeks to gather and disseminate the best assignment designs of the Artificial Intelligence (AI) Education community. Recognizing that assignments form the core of student learning experience, we here present abstracts of ten AI assignments from the 2019 session that are easily adoptable, playfully engaging, and flexible for a variety of instructor needs. Assignment specifications and supporting resources may be found at http://modelai.gettysburg.edu.

Deep reinforcement learning has led to several recent breakthroughs, though the learned policies are often based on black-box neural networks. This makes them difficult to interpret and to impose desired specification constraints during learning. We present an iterative framework, MORL, for improving the learned policies using program synthesis. Concretely, we propose to use synthesis techniques to obtain a symbolic representation of the learned policy, which can then be debugged manually or automatically using program repair. After the repair step, we use behavior cloning to obtain the policy corresponding to the repaired program, which is then further improved using gradient descent. This process continues until the learned policy satisfies desired constraints. We instantiate MORL for the simple CartPole problem and show that the programmatic representation allows for high-level modifications that in turn lead to improved learning of the policies.

Policy gradient methods are a widely used class of model-free reinforcement learning algorithms where a state-dependent baseline is used to reduce gradient estimator variance. Several recent papers extend the baseline to depend on both the state and action and suggest that this significantly reduces variance and improves sample efficiency without introducing bias into the gradient estimates. To better understand this development, we decompose the variance of the policy gradient estimator and numerically show that learned state-action-dependent baselines do not in fact reduce variance over a state-dependent baseline in commonly tested benchmark domains. We confirm this unexpected result by reviewing the open-source code accompanying these prior papers, and show that subtle implementation decisions cause deviations from the methods presented in the papers and explain the source of the previously observed empirical gains. Furthermore, the variance decomposition highlights areas for improvement, which we demonstrate by illustrating a simple change to the typical value function parameterization that can significantly improve performance.

Model-free deep reinforcement learning algorithms are able to successfully solve a wide range of continuous control tasks, but typically require many on-policy samples to achieve good performance. Model-based RL algorithms are sample-efficient on the other hand, while learning accurate global models of complex dynamic environments has turned out to be tricky in practice, which leads to the unsatisfactory performance of the learned policies. In this work, we combine the sample-efficiency of model-based algorithms and the accuracy of model-free algorithms. We leverage multi-step neural network based predictive models by embedding real trajectories into imaginary rollouts of the model, and use the imaginary cumulative rewards as control variates for model-free algorithms. In this way, we achieved the strengths of both sides and derived an estimator which is not only sample-efficient, but also unbiased and of very low variance. We present our evaluation on the MuJoCo and OpenAI Gym benchmarks.

We explore a new way to evaluate generative models using insights from evaluation of competitive games between human players. We show experimentally that tournaments between generators and discriminators provide an effective way to evaluate generative models. We introduce two methods for summarizing tournament outcomes: tournament win rate and skill rating. Evaluations are useful in different contexts, including monitoring the progress of a single model as it learns during the training process, and comparing the capabilities of two different fully trained models. We show that a tournament consisting of a single model playing against past and future versions of itself produces a useful measure of training progress. A tournament containing multiple separate models (using different seeds, hyperparameters, and architectures) provides a useful relative comparison between different trained GANs. Tournament-based rating methods are conceptually distinct from numerous previous categories of approaches to evaluation of generative models, and have complementary advantages and disadvantages.

We present DAPIP, a Programming-By-Example system that learns to program with APIs to perform data transformation tasks. We design a domain-specific language (DSL) that allows for arbitrary concatenations of API outputs and constant strings. The DSL consists of three family of APIs: regular expression-based APIs, lookup APIs, and transformation APIs. We then present a novel neural synthesis algorithm to search for programs in the DSL that are consistent with a given set of examples. The search algorithm uses recently introduced neural architectures to encode input-output examples and to model the program search in the DSL. We show that synthesis algorithm outperforms baseline methods for synthesizing programs on both synthetic and real-world benchmarks.

The problem of automatically generating a computer program from some specification has been studied since the early days of AI. Recently, two competing approaches for automatic program learning have received significant attention: (1) neural program synthesis, where a neural network is conditioned on input/output (I/O) examples and learns to generate a program, and (2) neural program induction, where a neural network generates new outputs directly using a latent program representation. Here, for the first time, we directly compare both approaches on a large-scale, real-world learning task and we additionally contrast to rule-based program synthesis, which uses hand-crafted semantics to guide the program generation. Our neural models use a modified attention RNN to allow encoding of variable-sized sets of I/O pairs, which achieve 92% accuracy on a real-world test set, compared to the 34% accuracy of the previous best neural synthesis approach. The synthesis model also outperforms a comparable induction model on this task, but we more importantly demonstrate that the strength of each approach is highly dependent on the evaluation metric and end-user application. Finally, we show that we can train our neural models to remain very robust to the type of noise expected in real-world data (e.g., typos), while a highly-engineered rule-based system fails entirely.

Existing centralized identification systems, such as Facebook or Google, offer convenient services for authentication and storing user information that third-party services leverage to reduce friction and streamline the process of creating user profiles and logging in. However, these systems present conflict of interests, and users may be interested in a third-party service that focuses exclusively on authentication. To this end, we present Iris, a web-based, secure solution that resolves the single-point-of-failure concern and provides secure protocols of data transmission and transparency of data usage. Iris consists of two components: a database that holds users’ information encrypted with their passphrases along with a public API to retreive this data, and a client-side browser extension that handles encryption and decryption of the user’s information so that the passphrase never leaves the user’s local computer. In this paper, we detail how using Iris reduces friction for the website and user, and how a high level of security and authentication is achieved. In addition, we examine the potential threat model and present directions for future work.

Consider the free algebra An generated over Q by n generators x(1), ... , x(n). Interesting objects attached to A = A(n) are members of its lower central series, L-i = L-i(A), defined inductively by L-1 = A, Li+1 = [A, L-i], and their associated graded components B-i = B-i(A) defined as B-i = L-i/ Li+1. These quotients B-i for i >= 2, as well as the reduced quotient (B) over bar (1) = A/(L-2 + AL(3)), exhibit a rich geometric structure, as shown by Feigin and Shoikhet (2007) [FS] and later authors (Dobrovolska et al., 1997 [DKM], Dobrovolska and Etingof, 2008 [DE], Arbesfeld and Jordan, 2010 [AJ], Bapat and Jordan, 2010 [BJ]).We study the same problem over the integers Z and finite fields F-p. New phenomena arise, namely, torsion in B-i over Z, and jumps in dimension over F-p. We describe the torsion in the reduced quotient (B) over bar (1) and B-2 geometrically in terms of the De Rham cohomology of Z(n). As a corollary we obtain a complete description of (B) over bar (1) (A(n)(Z)) and (B) over bar (1) (A(n)(F-p)). as well as of B-2(A(n)(Z[1/2])) and B-2(A(n)(F-p)), p > 2. We also give theoretical and experimental results for B-i with i > 2, formulating a number of conjectures and questions on their basis. Finally, we discuss the supercase, when some of the generators are odd and some are even, and provide some theoretical results and experimental data in this case. (C) 2012 Elsevier Inc. All rights reserved.