Vihari Piratla

Purvam Jain, Preethi Jyothi, Vihari Piratla et al.

We introduce a set of synthetic algorithmic tasks to detect cross-lingual gaps in the abilities of large language models. Our benchmark is commensurate across languages, since it requires models to perform the same underlying task in different languages; scalable, since each task can be generated at varying levels of complexity allowing it to be adapted to models with different capabilities; quantifiable, since every task admits an objective notion of correctness; and transparent, since tasks are generated from simple templates that can be readily audited for translation errors. Because our benchmark focuses on algorithmic tasks, differential performance is a sufficient -- but not necessary -- indicator of cross-lingual gaps. Nevertheless, we show through extensive experiments that our benchmark exposes persistent cross-lingual gaps in multiple state-of-the-art models.

Katherine M. Collins, Umang Bhatt, Weiyang Liu et al. · cambridge

Aligning model representations to humans has been found to improve robustness and generalization. However, such methods often focus on standard observational data. Synthetic data is proliferating and powering many advances in machine learning; yet, it is not always clear whether synthetic labels are perceptually aligned to humans -- rendering it likely model representations are not human aligned. We focus on the synthetic data used in mixup: a powerful regularizer shown to improve model robustness, generalization, and calibration. We design a comprehensive series of elicitation interfaces, which we release as HILL MixE Suite, and recruit 159 participants to provide perceptual judgments along with their uncertainties, over mixup examples. We find that human perceptions do not consistently align with the labels traditionally used for synthetic points, and begin to demonstrate the applicability of these findings to potentially increase the reliability of downstream models, particularly when incorporating human uncertainty. We release all elicited judgments in a new data hub we call H-Mix.

Juyeon Heo, Vihari Piratla, Matthew Wicker et al. · cambridge

Machine learning from explanations (MLX) is an approach to learning that uses human-provided explanations of relevant or irrelevant features for each input to ensure that model predictions are right for the right reasons. Existing MLX approaches rely on local model interpretation methods and require strong model smoothing to align model and human explanations, leading to sub-optimal performance. We recast MLX as a robustness problem, where human explanations specify a lower dimensional manifold from which perturbations can be drawn, and show both theoretically and empirically how this approach alleviates the need for strong model smoothing. We consider various approaches to achieving robustness, leading to improved performance over prior MLX methods. Finally, we show how to combine robustness with an earlier MLX method, yielding state-of-the-art results on both synthetic and real-world benchmarks.

Vihari Piratla

Our goal is to improve reliability of Machine Learning (ML) systems deployed in the wild. ML models perform exceedingly well when test examples are similar to train examples. However, real-world applications are required to perform on any distribution of test examples. Current ML systems can fail silently on test examples with distribution shifts. In order to improve reliability of ML models due to covariate or domain shift, we propose algorithms that enable models to: (a) generalize to a larger family of test distributions, (b) evaluate accuracy under distribution shifts, (c) adapt to a target distribution. We study causes of impaired robustness to domain shifts and present algorithms for training domain robust models. A key source of model brittleness is due to domain overfitting, which our new training algorithms suppress and instead encourage domain-general hypotheses. While we improve robustness over standard training methods for certain problem settings, performance of ML systems can still vary drastically with domain shifts. It is crucial for developers and stakeholders to understand model vulnerabilities and operational ranges of input, which could be assessed on the fly during the deployment, albeit at a great cost. Instead, we advocate for proactively estimating accuracy surfaces over any combination of prespecified and interpretable domain shifts for performance forecasting. We present a label-efficient estimation to address estimation over a combinatorial space of domain shifts. Further, when a model's performance on a target domain is found to be poor, traditional approaches adapt the model using the target domain's resources. Standard adaptation methods assume access to sufficient labeled resources, which may be impractical for deployed models. We initiate a study of lightweight adaptation techniques with only unlabeled data resources with a focus on language applications.

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.

Shiv Shankar, Vihari Piratla

Most deep learning research has focused on developing new model and training procedures. On the other hand the training objective has usually been restricted to combinations of standard losses. When the objective aligns well with the evaluation metric, this is not a major issue. However when dealing with complex structured outputs, the ideal objective can be hard to optimize and the efficacy of usual objectives as a proxy for the true objective can be questionable. In this work, we argue that the existing inference network based structure prediction methods ( Tu and Gimpel 2018; Tu, Pang, and Gimpel 2020) are indirectly learning to optimize a dynamic loss objective parameterized by the energy model. We then explore using implicit-gradient based technique to learn the corresponding dynamic objectives. Our experiments show that implicitly learning a dynamic loss landscape is an effective method for improving model performance in structure prediction.

Vihari Piratla, Juyeon Heo, Katherine M. Collins et al.

Model explanations can be valuable for interpreting and debugging predictive models. We study a specific kind called Concept Explanations, where the goal is to interpret a model using human-understandable concepts. Although popular for their easy interpretation, concept explanations are known to be noisy. We begin our work by identifying various sources of uncertainty in the estimation pipeline that lead to such noise. We then propose an uncertainty-aware Bayesian estimation method to address these issues, which readily improved the quality of explanations. We demonstrate with theoretical analysis and empirical evaluation that explanations computed by our method are robust to train-time choices while also being label-efficient. Further, our method proved capable of recovering relevant concepts amongst a bank of thousands, in an evaluation with real-datasets and off-the-shelf models, demonstrating its scalability. We believe the improved quality of uncertainty-aware concept explanations make them a strong candidate for more reliable model interpretation. We release our code at https://github.com/vps-anonconfs/uace.

Mihnea Ghitu, Vihari Piratla, Matthew Wicker

Controlling the patterns a model learns is essential to preventing reliance on irrelevant or misleading features. Such reliance on irrelevant features, often called shortcut features, has been observed across domains, including medical imaging and natural language processing, where it may lead to real-world harms. A common mitigation strategy leverages annotations (provided by humans or machines) indicating which features are relevant or irrelevant. These annotations are compared to model explanations, typically in the form of feature salience, and used to guide the loss function during training. Unfortunately, recent works have demonstrated that feature salience methods are unreliable and therefore offer a poor signal to optimize. In this work, we propose a simplified objective that simultaneously optimizes for explanation robustness and mitigation of shortcut learning. Unlike prior objectives with similar aims, we demonstrate theoretically why our approach ought to be more effective. Across a comprehensive series of experiments, we show that our approach consistently reduces test-time misclassifications by 20% compared to state-of-the-art methods. We also extend prior experimental settings to include natural language processing tasks. Additionally, we conduct novel ablations that yield practical insights, including the relative importance of annotation quality over quantity. Code for our method and experiments is available at: https://github.com/Mihneaghitu/ModelGuidanceViaRobustFeatureAttribution.

Vihari Piratla, Purvam Jain, Darshan Singh et al.

Any piece of knowledge is usually expressed in one or a handful of natural languages on the web or in any large corpus. Large Language Models (LLMs) act as a bridge by acquiring knowledge from a source language and making it accessible when queried from target languages. Prior research has pointed to a cross-lingual gap, viz., a drop in accuracy when the knowledge is queried in a target language compared to when the query is in the source language. Existing research has rationalized divergence in latent representations in source and target languages as the source of cross-lingual gap. In this work, we take an alternative view and hypothesize that the variance of responses in the target language is the main cause of this gap. For the first time, we formalize the cross-lingual gap in terms of bias-variance decomposition. We present extensive experimental evidence which support proposed formulation and hypothesis. We then reinforce our hypothesis through multiple inference-time interventions that control the variance and reduce the cross-lingual gap. We demonstrate a simple prompt instruction to reduce the response variance, which improved target accuracy by 20-25% across different models.

Vihari Piratla, Praneeth Netrapalli, Sunita Sarawagi

We consider the problem of training a classification model with group annotated training data. Recent work has established that, if there is distribution shift across different groups, models trained using the standard empirical risk minimization (ERM) objective suffer from poor performance on minority groups and that group distributionally robust optimization (Group-DRO) objective is a better alternative. The starting point of this paper is the observation that though Group-DRO performs better than ERM on minority groups for some benchmark datasets, there are several other datasets where it performs much worse than ERM. Inspired by ideas from the closely related problem of domain generalization, this paper proposes a new and simple algorithm that explicitly encourages learning of features that are shared across various groups. The key insight behind our proposed algorithm is that while Group-DRO focuses on groups with worst regularized loss, focusing instead, on groups that enable better performance even on other groups, could lead to learning of shared/common features, thereby enhancing minority performance beyond what is achieved by Group-DRO. Empirically, we show that our proposed algorithm matches or achieves better performance compared to strong contemporary baselines including ERM and Group-DRO on standard benchmarks on both minority groups and across all groups. Theoretically, we show that the proposed algorithm is a descent method and finds first order stationary points of smooth nonconvex functions.

Anshul Nasery, Soumyadeep Thakur, Vihari Piratla et al.

In several real world applications, machine learning models are deployed to make predictions on data whose distribution changes gradually along time, leading to a drift between the train and test distributions. Such models are often re-trained on new data periodically, and they hence need to generalize to data not too far into the future. In this context, there is much prior work on enhancing temporal generalization, e.g. continuous transportation of past data, kernel smoothed time-sensitive parameters and more recently, adversarial learning of time-invariant features. However, these methods share several limitations, e.g, poor scalability, training instability, and dependence on unlabeled data from the future. Responding to the above limitations, we propose a simple method that starts with a model with time-sensitive parameters but regularizes its temporal complexity using a Gradient Interpolation (GI) loss. GI allows the decision boundary to change along time and can still prevent overfitting to the limited training time snapshots by allowing task-specific control over changes along time. We compare our method to existing baselines on multiple real-world datasets, which show that GI outperforms more complicated generative and adversarial approaches on the one hand, and simpler gradient regularization methods on the other.

Vihari Piratla, Soumen Chakrabarty, Sunita Sarawagi

Our goal is to evaluate the accuracy of a black-box classification model, not as a single aggregate on a given test data distribution, but as a surface over a large number of combinations of attributes characterizing multiple test data distributions. Such attributed accuracy measures become important as machine learning models get deployed as a service, where the training data distribution is hidden from clients, and different clients may be interested in diverse regions of the data distribution. We present Attributed Accuracy Assay (AAA)--a Gaussian Process (GP)--based probabilistic estimator for such an accuracy surface. Each attribute combination, called an 'arm', is associated with a Beta density from which the service's accuracy is sampled. We expect the GP to smooth the parameters of the Beta density over related arms to mitigate sparsity. We show that obvious application of GPs cannot address the challenge of heteroscedastic uncertainty over a huge attribute space that is sparsely and unevenly populated. In response, we present two enhancements: pooling sparse observations, and regularizing the scale parameter of the Beta densities. After introducing these innovations, we establish the effectiveness of AAA in terms of both its estimation accuracy and exploration efficiency, through extensive experiments and analysis.

Shivaram Kalyanakrishnan, Siddharth Aravindan, Vishwajeet Bagdawat et al.

In the practice of sequential decision making, agents are often designed to sense state at regular intervals of $d$ time steps, $d > 1$, ignoring state information in between sensing steps. While it is clear that this practice can reduce sensing and compute costs, recent results indicate a further benefit. On many Atari console games, reinforcement learning (RL) algorithms deliver substantially better policies when run with $d > 1$ -- in fact with $d$ even as high as $180$. In this paper, we investigate the role of the parameter $d$ in RL; $d$ is called the "frame-skip" parameter, since states in the Atari domain are images. For evaluating a fixed policy, we observe that under standard conditions, frame-skipping does not affect asymptotic consistency. Depending on other parameters, it can possibly even benefit learning. To use $d > 1$ in the control setting, one must first specify which $d$-step open-loop action sequences can be executed in between sensing steps. We focus on "action-repetition", the common restriction of this choice to $d$-length sequences of the same action. We define a task-dependent quantity called the "price of inertia", in terms of which we upper-bound the loss incurred by action-repetition. We show that this loss may be offset by the gain brought to learning by a smaller task horizon. Our analysis is supported by experiments on different tasks and learning algorithms.

Sahil Shah, Vihari Piratla, Soumen Chakrabarti et al.

State-of-the-art NLP inference uses enormous neural architectures and models trained for GPU-months, well beyond the reach of most consumers of NLP. This has led to one-size-fits-all public API-based NLP service models by major AI companies, serving large numbers of clients. Neither (hardware deficient) clients nor (heavily subscribed) servers can afford traditional fine tuning. Many clients own little or no labeled data. We initiate a study of adaptation of centralized NLP services to clients, and present one practical and lightweight approach. Each client uses an unsupervised, corpus-based sketch to register to the service. The server uses an auxiliary network to map the sketch to an abstract vector representation, which then informs the main labeling network. When a new client registers with its sketch, it gets immediate accuracy benefits. We demonstrate the success of the proposed architecture using sentiment labeling, NER, and predictive language modeling

Vihari Piratla, Shiv Shankar

Data augmentation is a popular pre-processing trick to improve generalization accuracy. It is believed that by processing augmented inputs in tandem with the original ones, the model learns a more robust set of features which are shared between the original and augmented counterparts. However, we show that is not the case even for the best augmentation technique. In this work, we take a Domain Generalization viewpoint of augmentation based methods. This new perspective allowed for probing overfitting and delineating avenues for improvement. Our exploration with the state-of-art augmentation method provides evidence that the learned representations are not as robust even towards distortions used during training. This suggests evidence for the untapped potential of augmented examples.

Vihari Piratla, Praneeth Netrapalli, Sunita Sarawagi

Domain generalization refers to the task of training a model which generalizes to new domains that are not seen during training. We present CSD (Common Specific Decomposition), for this setting,which jointly learns a common component (which generalizes to new domains) and a domain specific component (which overfits on training domains). The domain specific components are discarded after training and only the common component is retained. The algorithm is extremely simple and involves only modifying the final linear classification layer of any given neural network architecture. We present a principled analysis to understand existing approaches, provide identifiability results of CSD,and study effect of low-rank on domain generalization. We show that CSD either matches or beats state of the art approaches for domain generalization based on domain erasure, domain perturbed data augmentation, and meta-learning. Further diagnostics on rotated MNIST, where domains are interpretable, confirm the hypothesis that CSD successfully disentangles common and domain specific components and hence leads to better domain generalization.

Abhijeet Awasthi, Sunita Sarawagi, Rasna Goyal et al.

We present a Parallel Iterative Edit (PIE) model for the problem of local sequence transduction arising in tasks like Grammatical error correction (GEC). Recent approaches are based on the popular encoder-decoder (ED) model for sequence to sequence learning. The ED model auto-regressively captures full dependency among output tokens but is slow due to sequential decoding. The PIE model does parallel decoding, giving up the advantage of modelling full dependency in the output, yet it achieves accuracy competitive with the ED model for four reasons: 1.~predicting edits instead of tokens, 2.~labeling sequences instead of generating sequences, 3.~iteratively refining predictions to capture dependencies, and 4.~factorizing logits over edits and their token argument to harness pre-trained language models like BERT. Experiments on tasks spanning GEC, OCR correction and spell correction demonstrate that the PIE model is an accurate and significantly faster alternative for local sequence transduction.

Vihari Piratla, Sunita Sarawagi, Soumen Chakrabarti

Given a small corpus $\mathcal D_T$ pertaining to a limited set of focused topics, our goal is to train embeddings that accurately capture the sense of words in the topic in spite of the limited size of $\mathcal D_T$. These embeddings may be used in various tasks involving $\mathcal D_T$. A popular strategy in limited data settings is to adapt pre-trained embeddings $\mathcal E$ trained on a large corpus. To correct for sense drift, fine-tuning, regularization, projection, and pivoting have been proposed recently. Among these, regularization informed by a word's corpus frequency performed well, but we improve upon it using a new regularizer based on the stability of its cooccurrence with other words. However, a thorough comparison across ten topics, spanning three tasks, with standardized settings of hyper-parameters, reveals that even the best embedding adaptation strategies provide small gains beyond well-tuned baselines, which many earlier comparisons ignored. In a bold departure from adapting pretrained embeddings, we propose using $\mathcal D_T$ to probe, attend to, and borrow fragments from any large, topic-rich source corpus (such as Wikipedia), which need not be the corpus used to pretrain embeddings. This step is made scalable and practical by suitable indexing. We reach the surprising conclusion that even limited corpus augmentation is more useful than adapting embeddings, which suggests that non-dominant sense information may be irrevocably obliterated from pretrained embeddings and cannot be salvaged by adaptation.

Shiv Shankar, Vihari Piratla, Soumen Chakrabarti et al.

We present CROSSGRAD, a method to use multi-domain training data to learn a classifier that generalizes to new domains. CROSSGRAD does not need an adaptation phase via labeled or unlabeled data, or domain features in the new domain. Most existing domain adaptation methods attempt to erase domain signals using techniques like domain adversarial training. In contrast, CROSSGRAD is free to use domain signals for predicting labels, if it can prevent overfitting on training domains. We conceptualize the task in a Bayesian setting, in which a sampling step is implemented as data augmentation, based on domain-guided perturbations of input instances. CROSSGRAD parallelly trains a label and a domain classifier on examples perturbed by loss gradients of each other's objectives. This enables us to directly perturb inputs, without separating and re-mixing domain signals while making various distributional assumptions. Empirical evaluation on three different applications where this setting is natural establishes that (1) domain-guided perturbation provides consistently better generalization to unseen domains, compared to generic instance perturbation methods, and that (2) data augmentation is a more stable and accurate method than domain adversarial training.