Harsh Mehta

Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao et al. · allen-ai, amazon-science

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Harsh Mehta, Ankit Gupta, Ashok Cutkosky et al. · ibm-research

State space models have shown to be effective at modeling long range dependencies, specially on sequence classification tasks. In this work we focus on autoregressive sequence modeling over English books, Github source code and ArXiv mathematics articles. Based on recent developments around the effectiveness of gated activation functions, we propose a new layer named Gated State Space (GSS) and show that it trains significantly faster than the diagonal version of S4 (i.e. DSS) on TPUs, is fairly competitive with several well-tuned Transformer-based baselines and exhibits zero-shot generalization to longer inputs while being straightforward to implement. Finally, we show that leveraging self-attention to model local dependencies improves the performance of GSS even further.

Ankit Gupta, Harsh Mehta, Jonathan Berant · deepmind, ibm-research

Sequence models based on linear state spaces (SSMs) have recently emerged as a promising choice of architecture for modeling long range dependencies across various modalities. However, they invariably rely on discretization of a continuous state space, which complicates their presentation and understanding. In this work, we dispose of the discretization step, and propose a model based on vanilla Diagonal Linear RNNs ($\mathrm{DLR}$). We empirically show that, despite being conceptually much simpler, $\mathrm{DLR}$ is as performant as previously-proposed SSMs on a variety of tasks and benchmarks including Long Range Arena and raw speech classification. Moreover, we characterize the expressivity of SSMs (including $\mathrm{DLR}$) and attention-based models via a suite of $13$ synthetic sequence-to-sequence tasks involving interactions over tens of thousands of tokens, ranging from simple operations, such as shifting an input sequence, to detecting co-dependent visual features over long spatial ranges in flattened images. We find that while SSMs report near-perfect performance on tasks that can be modeled via $\textit{few}$ convolutional kernels, they struggle on tasks requiring $\textit{many}$ such kernels and especially when the desired sequence manipulation is $\textit{context-dependent}$. Despite these limitations, $\mathrm{DLR}$ reaches high performance on two higher-order reasoning tasks $\mathrm{ListOpsSubTrees}$ and $\mathrm{PathfinderSegmentation}\text{-}\mathrm{256}$ with input lengths $8K$ and $65K$ respectively, and gives encouraging performance on $\mathrm{PathfinderSegmentation}\text{-}\mathrm{512}$ with input length $262K$ for which attention is not a viable choice.

Tolga Ergen, Behnam Neyshabur, Harsh Mehta · stanford

Understanding the fundamental mechanism behind the success of transformer networks is still an open problem in the deep learning literature. Although their remarkable performance has been mostly attributed to the self-attention mechanism, the literature still lacks a solid analysis of these networks and interpretation of the functions learned by them. To this end, we study the training problem of attention/transformer networks and introduce a novel convex analytic approach to improve the understanding and optimization of these networks. Particularly, we first introduce a convex alternative to the self-attention mechanism and reformulate the regularized training problem of transformer networks with our alternative convex attention. Then, we cast the reformulation as a convex optimization problem that is interpretable and easier to optimize. Moreover, as a byproduct of our convex analysis, we reveal an implicit regularization mechanism, which promotes sparsity across tokens. Therefore, we not only improve the optimization of attention/transformer networks but also provide a solid theoretical understanding of the functions learned by them. We also demonstrate the effectiveness of our theory through several numerical experiments.

Kiran Vodrahalli, Santiago Ontanon, Nilesh Tripuraneni et al. · deepmind

We introduce Michelangelo: a minimal, synthetic, and unleaked long-context reasoning evaluation for large language models which is also easy to automatically score. This evaluation is derived via a novel, unifying framework for evaluations over arbitrarily long contexts which measure the model's ability to do more than retrieve a single piece of information from its context. The central idea of the Latent Structure Queries framework (LSQ) is to construct tasks which require a model to ``chisel away'' the irrelevant information in the context, revealing a latent structure in the context. To verify a model's understanding of this latent structure, we query the model for details of the structure. Using LSQ, we produce three diagnostic long-context evaluations across code and natural-language domains intended to provide a stronger signal of long-context language model capabilities. We perform evaluations on several state-of-the-art models and demonstrate both that a) the proposed evaluations are high-signal and b) that there is significant room for improvement in synthesizing long-context information.

Harsh Mehta, Abhradeep Thakurta, Alexey Kurakin et al.

Differential Privacy (DP) provides a formal framework for training machine learning models with individual example level privacy. In the field of deep learning, Differentially Private Stochastic Gradient Descent (DP-SGD) has emerged as a popular private training algorithm. Unfortunately, the computational cost of training large-scale models with DP-SGD is substantially higher than non-private training. This is further exacerbated by the fact that increasing the number of parameters leads to larger degradation in utility with DP. In this work, we zoom in on the ImageNet dataset and demonstrate that, similar to the non-private case, pre-training over-parameterized models on a large public dataset can lead to substantial gains when the model is finetuned privately. Moreover, by systematically comparing private and non-private models across a range of large batch sizes, we find that similar to non-private setting, choice of optimizer can further improve performance substantially with DP. By using LAMB optimizer with DP-SGD we saw improvement of up to 20$\%$ points (absolute). Finally, we show that finetuning just the last layer for a \emph{single step} in the full batch setting, combined with extremely small-scale (near-zero) initialization leads to both SOTA results of 81.7 $\%$ under a wide privacy budget range of $ε\in [4, 10]$ and $δ$ = $10^{-6}$ while minimizing the computational overhead substantially.

Ashok Cutkosky, Harsh Mehta, Francesco Orabona

We present new algorithms for optimizing non-smooth, non-convex stochastic objectives based on a novel analysis technique. This improves the current best-known complexity for finding a $(δ,ε)$-stationary point from $O(ε^{-4}δ^{-1})$ stochastic gradient queries to $O(ε^{-3}δ^{-1})$, which we also show to be optimal. Our primary technique is a reduction from non-smooth non-convex optimization to online learning, after which our results follow from standard regret bounds in online learning. For deterministic and second-order smooth objectives, applying more advanced optimistic online learning techniques enables a new complexity of $O(ε^{-1.5}δ^{-0.5})$. Our techniques also recover all optimal or best-known results for finding $ε$ stationary points of smooth or second-order smooth objectives in both stochastic and deterministic settings.

Harsh Mehta, Walid Krichene, Abhradeep Thakurta et al.

Leveraging transfer learning has recently been shown to be an effective strategy for training large models with Differential Privacy (DP). Moreover, somewhat surprisingly, recent works have found that privately training just the last layer of a pre-trained model provides the best utility with DP. While past studies largely rely on algorithms like DP-SGD for training large models, in the specific case of privately learning from features, we observe that computational burden is low enough to allow for more sophisticated optimization schemes, including second-order methods. To that end, we systematically explore the effect of design parameters such as loss function and optimization algorithm. We find that, while commonly used logistic regression performs better than linear regression in the non-private setting, the situation is reversed in the private setting. We find that linear regression is much more effective than logistic regression from both privacy and computational aspects, especially at stricter epsilon values ($ε< 1$). On the optimization side, we also explore using Newton's method, and find that second-order information is quite helpful even with privacy, although the benefit significantly diminishes with stricter privacy guarantees. While both methods use second-order information, least squares is effective at lower epsilons while Newton's method is effective at larger epsilon values. To combine the benefits of both, we propose a novel algorithm called DP-FC, which leverages feature covariance instead of the Hessian of the logistic regression loss and performs well across all $ε$ values we tried. With this, we obtain new SOTA results on ImageNet-1k, CIFAR-100 and CIFAR-10 across all values of $ε$ typically considered. Most remarkably, on ImageNet-1K, we obtain top-1 accuracy of 88\% under (8, $8 * 10^{-7}$)-DP and 84.3\% under (0.1, $8 * 10^{-7}$)-DP.

Aaron Defazio, Ashok Cutkosky, Harsh Mehta et al.

Learning rate schedules used in practice bear little resemblance to those recommended by theory. We close much of this theory/practice gap, and as a consequence are able to derive new problem-adaptive learning rate schedules. Our main technical contribution is a refined analysis of learning rate schedules for a wide class of optimization algorithms (including SGD). When considering only worst-case analysis, our theory predicts that the optimal choice is the linear decay schedule where the step-size is set proportional to 1 - t/T, where t is the current iteration and T is the total number of steps. To go beyond this worst-case analysis, we use the observed gradient norms to derive schedules refined for any particular task. These refined schedules exhibit learning rate warm-up and rapid learning rate annealing near the end of training. Ours is the first systematic approach to automatically yield both of these properties. We perform the most comprehensive evaluation of learning rate schedules to date, evaluating across 10 diverse deep learning problems, a series of LLMs, and a suite of logistic regression problems. We validate that overall, the linear-decay schedule outperforms all commonly used default schedules including cosine annealing. Our adaptive schedule refinement method gives further improvements.

Aaron Defazio, Xingyu Alice Yang, Harsh Mehta et al.

Existing learning rate schedules that do not require specification of the optimization stopping step T are greatly out-performed by learning rate schedules that depend on T. We propose an approach that avoids the need for this stopping time by eschewing the use of schedules entirely, while exhibiting state-of-the-art performance compared to schedules across a wide family of problems ranging from convex problems to large-scale deep learning problems. Our Schedule-Free approach introduces no additional hyper-parameters over standard optimizers with momentum. Our method is a direct consequence of a new theory we develop that unifies scheduling and iterate averaging. An open source implementation of our method is available at https://github.com/facebookresearch/schedule_free. Schedule-Free AdamW is the core algorithm behind our winning entry to the MLCommons 2024 AlgoPerf Algorithmic Efficiency Challenge Self-Tuning track.

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.

Gemma Team, Aishwarya Kamath, Johan Ferret et al. · deepmind, mit

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.

Harsh Mehta, Steffen Rendle, Walid Krichene et al.

We present ALX, an open-source library for distributed matrix factorization using Alternating Least Squares, written in JAX. Our design allows for efficient use of the TPU architecture and scales well to matrix factorization problems of O(B) rows/columns by scaling the number of available TPU cores. In order to spur future research on large scale matrix factorization methods and to illustrate the scalability properties of our own implementation, we also built a real world web link prediction dataset called WebGraph. This dataset can be easily modeled as a matrix factorization problem. We created several variants of this dataset based on locality and sparsity properties of sub-graphs. The largest variant of WebGraph has around 365M nodes and training a single epoch finishes in about 20 minutes with 256 TPU cores. We include speed and performance numbers of ALX on all variants of WebGraph. Both the framework code and the dataset is open-sourced.

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.

Ashok Cutkosky, Aaron Defazio, Harsh Mehta

We introduce a technique for tuning the learning rate scale factor of any base optimization algorithm and schedule automatically, which we call \textsc{mechanic}. Our method provides a practical realization of recent theoretical reductions for accomplishing a similar goal in online convex optimization. We rigorously evaluate \textsc{mechanic} on a range of large scale deep learning tasks with varying batch sizes, schedules, and base optimization algorithms. These experiments demonstrate that depending on the problem, \textsc{mechanic} either comes very close to, matches or even improves upon manual tuning of learning rates.

Yi Tay, Vinh Q. Tran, Mostafa Dehghani et al.

In this paper, we demonstrate that information retrieval can be accomplished with a single Transformer, in which all information about the corpus is encoded in the parameters of the model. To this end, we introduce the Differentiable Search Index (DSI), a new paradigm that learns a text-to-text model that maps string queries directly to relevant docids; in other words, a DSI model answers queries directly using only its parameters, dramatically simplifying the whole retrieval process. We study variations in how documents and their identifiers are represented, variations in training procedures, and the interplay between models and corpus sizes. Experiments demonstrate that given appropriate design choices, DSI significantly outperforms strong baselines such as dual encoder models. Moreover, DSI demonstrates strong generalization capabilities, outperforming a BM25 baseline in a zero-shot setup.

Ashok Cutkosky, Harsh Mehta

We consider non-convex stochastic optimization using first-order algorithms for which the gradient estimates may have heavy tails. We show that a combination of gradient clipping, momentum, and normalized gradient descent yields convergence to critical points in high-probability with best-known rates for smooth losses when the gradients only have bounded $\mathfrak{p}$th moments for some $\mathfrak{p}\in(1,2]$. We then consider the case of second-order smooth losses, which to our knowledge have not been studied in this setting, and again obtain high-probability bounds for any $\mathfrak{p}$. Moreover, our results hold for arbitrary smooth norms, in contrast to the typical SGD analysis which requires a Hilbert space norm. Further, we show that after a suitable "burn-in" period, the objective value will monotonically decrease for every iteration until a critical point is identified, which provides intuition behind the popular practice of learning rate "warm-up" and also yields a last-iterate guarantee.

Harsh Mehta, Ashok Cutkosky, Behnam Neyshabur

We construct an experimental setup in which changing the scale of initialization strongly impacts the implicit regularization induced by SGD, interpolating from good generalization performance to completely memorizing the training set while making little progress on the test set. Moreover, we find that the extent and manner in which generalization ability is affected depends on the activation and loss function used, with $\sin$ activation demonstrating extreme memorization. In the case of the homogeneous ReLU activation, we show that this behavior can be attributed to the loss function. Our empirical investigation reveals that increasing the scale of initialization correlates with misalignment of representations and gradients across examples in the same class. This insight allows us to devise an alignment measure over gradients and representations which can capture this phenomenon. We demonstrate that our alignment measure correlates with generalization of deep models trained on image classification tasks.

Ashok Cutkosky, Harsh Mehta

We provide an improved analysis of normalized SGD showing that adding momentum provably removes the need for large batch sizes on non-convex objectives. Then, we consider the case of objectives with bounded second derivative and show that in this case a small tweak to the momentum formula allows normalized SGD with momentum to find an $ε$-critical point in $O(1/ε^{3.5})$ iterations, matching the best-known rates without accruing any logarithmic factors or dependence on dimension. We also provide an adaptive method that automatically improves convergence rates when the variance in the gradients is small. Finally, we show that our method is effective when employed on popular large scale tasks such as ResNet-50 and BERT pretraining, matching the performance of the disparate methods used to get state-of-the-art results on both tasks.

Harsh Mehta, Yoav Artzi, Jason Baldridge et al.

The Touchdown dataset (Chen et al., 2019) provides instructions by human annotators for navigation through New York City streets and for resolving spatial descriptions at a given location. To enable the wider research community to work effectively with the Touchdown tasks, we are publicly releasing the 29k raw Street View panoramas needed for Touchdown. We follow the process used for the StreetLearn data release (Mirowski et al., 2019) to check panoramas for personally identifiable information and blur them as necessary. These have been added to the StreetLearn dataset and can be obtained via the same process as used previously for StreetLearn. We also provide a reference implementation for both of the Touchdown tasks: vision and language navigation (VLN) and spatial description resolution (SDR). We compare our model results to those given in Chen et al. (2019) and show that the panoramas we have added to StreetLearn fully support both Touchdown tasks and can be used effectively for further research and comparison.

Larry Lansing, Vihan Jain, Harsh Mehta et al.

VALAN is a lightweight and scalable software framework for deep reinforcement learning based on the SEED RL architecture. The framework facilitates the development and evaluation of embodied agents for solving grounded language understanding tasks, such as Vision-and-Language Navigation and Vision-and-Dialog Navigation, in photo-realistic environments, such as Matterport3D and Google StreetView. We have added a minimal set of abstractions on top of SEED RL allowing us to generalize the architecture to solve a variety of other RL problems. In this article, we will describe VALAN's software abstraction and architecture, and also present an example of using VALAN to design agents for instruction-conditioned indoor navigation.

Haoshuo Huang, Vihan Jain, Harsh Mehta et al.

Vision-and-Language Navigation (VLN) tasks such as Room-to-Room (R2R) require machine agents to interpret natural language instructions and learn to act in visually realistic environments to achieve navigation goals. The overall task requires competence in several perception problems: successful agents combine spatio-temporal, vision and language understanding to produce appropriate action sequences. Our approach adapts pre-trained vision and language representations to relevant in-domain tasks making them more effective for VLN. Specifically, the representations are adapted to solve both a cross-modal sequence alignment and sequence coherence task. In the sequence alignment task, the model determines whether an instruction corresponds to a sequence of visual frames. In the sequence coherence task, the model determines whether the perceptual sequences are predictive sequentially in the instruction-conditioned latent space. By transferring the domain-adapted representations, we improve competitive agents in R2R as measured by the success rate weighted by path length (SPL) metric.

Haoshuo Huang, Vihan Jain, Harsh Mehta et al.

Vision-and-Language Navigation (VLN) is a natural language grounding task where agents have to interpret natural language instructions in the context of visual scenes in a dynamic environment to achieve prescribed navigation goals. Successful agents must have the ability to parse natural language of varying linguistic styles, ground them in potentially unfamiliar scenes, plan and react with ambiguous environmental feedback. Generalization ability is limited by the amount of human annotated data. In particular, \emph{paired} vision-language sequence data is expensive to collect. We develop a discriminator that evaluates how well an instruction explains a given path in VLN task using multi-modal alignment. Our study reveals that only a small fraction of the high-quality augmented data from \citet{Fried:2018:Speaker}, as scored by our discriminator, is useful for training VLN agents with similar performance on previously unseen environments. We also show that a VLN agent warm-started with pre-trained components from the discriminator outperforms the benchmark success rates of 35.5 by 10\% relative measure on previously unseen environments.