Manuela Veloso

Gaurav Verma, Rachneet Kaur, Nishan Srishankar et al. · gatech

State-of-the-art multimodal web agents, powered by Multimodal Large Language Models (MLLMs), can autonomously execute many web tasks by processing user instructions and interacting with graphical user interfaces (GUIs). Current strategies for building web agents rely on (i) the generalizability of underlying MLLMs and their steerability via prompting, and (ii) large-scale fine-tuning of MLLMs on web-related tasks. However, web agents still struggle to automate tasks on unseen websites and domains, limiting their applicability to enterprise-specific and proprietary platforms. Beyond generalization from large-scale pre-training and fine-tuning, we propose building agents for few-shot adaptability using human demonstrations. We introduce the AdaptAgent framework that enables both proprietary and open-weights multimodal web agents to adapt to new websites and domains using few human demonstrations (up to 2). Our experiments on two popular benchmarks -- Mind2Web & VisualWebArena -- show that using in-context demonstrations (for proprietary models) or meta-adaptation demonstrations (for meta-learned open-weights models) boosts task success rate by 3.36% to 7.21% over non-adapted state-of-the-art models, corresponding to a relative increase of 21.03% to 65.75%. Furthermore, our additional analyses (a) show the effectiveness of multimodal demonstrations over text-only ones, (b) shed light on the influence of different data selection strategies during meta-learning on the generalization of the agent, and (c) demonstrate the effect of number of few-shot examples on the web agent's success rate. Overall, our results unlock a complementary axis for developing widely applicable multimodal web agents beyond large-scale pre-training and fine-tuning, emphasizing few-shot adaptability.

Salim I. Amoukou, Tom Bewley, Saumitra Mishra et al.

We introduce a novel approach for detecting distribution shifts that negatively impact the performance of machine learning models in continuous production environments, which requires no access to ground truth data labels. It builds upon the work of Podkopaev and Ramdas [2022], who address scenarios where labels are available for tracking model errors over time. Our solution extends this framework to work in the absence of labels, by employing a proxy for the true error. This proxy is derived using the predictions of a trained error estimator. Experiments show that our method has high power and false alarm control under various distribution shifts, including covariate and label shifts and natural shifts over geography and time.

Ran Zmigrod, Pranav Shetty, Mathieu Sibue et al.

The rise of large language models (LLMs) for visually rich document understanding (VRDU) has kindled a need for prompt-response, document-based datasets. As annotating new datasets from scratch is labor-intensive, the existing literature has generated prompt-response datasets from available resources using simple templates. For the case of key information extraction (KIE), one of the most common VRDU tasks, past work has typically employed the template "What is the value for the {key}?". However, given the variety of questions encountered in the wild, simple and uniform templates are insufficient for creating robust models in research and industrial contexts. In this work, we present K2Q, a diverse collection of five datasets converted from KIE to a prompt-response format using a plethora of bespoke templates. The questions in K2Q can span multiple entities and be extractive or boolean. We empirically compare the performance of seven baseline generative models on K2Q with zero-shot prompting. We further compare three of these models when training on K2Q versus training on simpler templates to motivate the need of our work. We find that creating diverse and intricate KIE questions enhances the performance and robustness of VRDU models. We hope this work encourages future studies on data quality for generative model training.

Antoine Gorceix, Bastien Le Chenadec, Ahmad Rammal et al.

In this paper, we study the ability of large language models to learn specific mathematical rules such as distributivity or simplifying equations. We present an empirical analysis of their ability to generalize these rules, as well as to reuse them in the context of word problems. For this purpose, we provide a rigorous methodology to build synthetic data incorporating such rules, and perform fine-tuning of large language models on such data. Our experiments show that our model can learn and generalize these rules to some extent, as well as suitably reuse them in the context of word problems.

Rachneet Kaur, Nishan Srishankar, Zhen Zeng et al.

Recent multimodal LLMs have shown promise in chart-based visual question answering, but their performance declines sharply on unannotated charts, those requiring precise visual interpretation rather than relying on textual shortcuts. To address this, we introduce ChartAgent, a novel agentic framework that explicitly performs visual reasoning directly within the chart's spatial domain. Unlike textual chain-of-thought reasoning, ChartAgent iteratively decomposes queries into visual subtasks and actively manipulates and interacts with chart images through specialized actions such as drawing annotations, cropping regions (e.g., segmenting pie slices, isolating bars), and localizing axes, using a library of chart-specific vision tools to fulfill each subtask. This iterative reasoning process closely mirrors human cognitive strategies for chart comprehension. ChartAgent achieves state-of-the-art accuracy on the ChartBench and ChartX benchmarks, surpassing prior methods by up to 16.07% absolute gain overall and 17.31% on unannotated, numerically intensive queries. Furthermore, our analyses show that ChartAgent is (a) effective across diverse chart types, (b) achieve the highest scores across varying visual and reasoning complexity levels, and (c) serves as a plug-and-play framework that boosts performance across diverse underlying LLMs. Our work is among the first to demonstrate visually grounded reasoning for chart understanding using tool-augmented multimodal agents.

Sanjay Kariyappa, Freddy Lécué, Saumitra Mishra et al.

This paper proposes Progressive Inference - a framework to compute input attributions to explain the predictions of decoder-only sequence classification models. Our work is based on the insight that the classification head of a decoder-only Transformer model can be used to make intermediate predictions by evaluating them at different points in the input sequence. Due to the causal attention mechanism, these intermediate predictions only depend on the tokens seen before the inference point, allowing us to obtain the model's prediction on a masked input sub-sequence, with negligible computational overheads. We develop two methods to provide sub-sequence level attributions using this insight. First, we propose Single Pass-Progressive Inference (SP-PI), which computes attributions by taking the difference between consecutive intermediate predictions. Second, we exploit a connection with Kernel SHAP to develop Multi Pass-Progressive Inference (MP-PI). MP-PI uses intermediate predictions from multiple masked versions of the input to compute higher quality attributions. Our studies on a diverse set of models trained on text classification tasks show that SP-PI and MP-PI provide significantly better attributions compared to prior work.

Simerjot Kaur, Ivan Brugere, Andrea Stefanucci et al.

Matching companies and investors is usually considered a highly specialized decision making process. Building an AI agent that can automate such recommendation process can significantly help reduce costs, and eliminate human biases and errors. However, limited sample size of financial data-sets and the need for not only good recommendations, but also explaining why a particular recommendation is being made, makes this a challenging problem. In this work we propose a representation learning based recommendation engine that works extremely well with small datasets and demonstrate how it can be coupled with a parameterized explanation generation engine to build an explainable recommendation system for investor-company matching. We compare the performance of our system with human generated recommendations and demonstrate the ability of our algorithm to perform extremely well on this task. We also highlight how explainability helps with real-life adoption of our system.

Yufei Wu, Mahmoud Mahfouz, Daniele Magazzeni et al.

The success of machine learning models in the financial domain is highly reliant on the quality of the data representation. In this paper, we focus on the representation of limit order book data and discuss the opportunities and challenges for learning representations of such data. We also experimentally analyse the issues associated with existing representations and present a guideline for future research in this area.

Daniel Borrajo, Manuela Veloso

Financial institutions mostly deal with people. Therefore, characterizing different kinds of human behavior can greatly help institutions for improving their relation with customers and with regulatory offices. In many of such interactions, humans have some internal goals, and execute some actions within the financial system that lead them to achieve their goals. In this paper, we tackle these tasks as a behavior-traces classification task. An observer agent tries to learn characterizing other agents by observing their behavior when taking actions in a given environment. The other agents can be of several types and the goal of the observer is to identify the type of the other agent given a trace of observations. We present CABBOT, a learning technique that allows the agent to perform on-line classification of the type of planning agent whose behavior is observing. In this work, the observer agent has partial and noisy observability of the environment (state and actions of the other agents). In order to evaluate the performance of the learning technique, we have generated a domain-independent goal-based simulator of agents. We present experiments in several (both financial and non-financial) domains with promising results.

Daniel Borrajo, Manuela Veloso, Sameena Shah

Many business applications involve adversarial relationships in which both sides adapt their strategies to optimize their opposing benefits. One of the key characteristics of these applications is the wide range of strategies that an adversary may choose as they adapt their strategy dynamically to sustain benefits and evade authorities. In this paper, we present a novel way of approaching these types of applications, in particular in the context of Anti-Money Laundering. We provide a mechanism through which diverse, realistic and new unobserved behavior may be generated to discover potential unobserved adversarial actions to enable organizations to preemptively mitigate these risks. In this regard, we make three main contributions. (a) Propose a novel behavior-based model as opposed to individual transactions-based models currently used by financial institutions. We introduce behavior traces as enriched relational representation to represent observed human behavior. (b) A modelling approach that observes these traces and is able to accurately infer the goals of actors by classifying the behavior into money laundering or standard behavior despite significant unobserved activity. And (c) a synthetic behavior simulator that can generate new previously unseen traces. The simulator incorporates a high level of flexibility in the behavioral parameters so that we can challenge the detection algorithm. Finally, we provide experimental results that show that the learning module (automated investigator) that has only partial observability can still successfully infer the type of behavior, and thus the simulated goals, followed by customers based on traces - a key aspiration for many applications today.

Natraj Raman, Armineh Nourbakhsh, Sameena Shah et al.

Task specific fine-tuning of a pre-trained neural language model using a custom softmax output layer is the de facto approach of late when dealing with document classification problems. This technique is not adequate when labeled examples are not available at training time and when the metadata artifacts in a document must be exploited. We address these challenges by generating document representations that capture both text and metadata artifacts in a task agnostic manner. Instead of traditional auto-regressive or auto-encoding based training, our novel self-supervised approach learns a soft-partition of the input space when generating text embeddings. Specifically, we employ a pre-learned topic model distribution as surrogate labels and construct a loss function based on KL divergence. Our solution also incorporates metadata explicitly rather than just augmenting them with text. The generated document embeddings exhibit compositional characteristics and are directly used by downstream classification tasks to create decision boundaries from a small number of labeled examples, thereby eschewing complicated recognition methods. We demonstrate through extensive evaluation that our proposed cross-model fusion solution outperforms several competitive baselines on multiple datasets.

Joshua Lockhart, Samuel Assefa, Ayham Alajdad et al.

Supervised learning classifiers inevitably make mistakes in production, perhaps mis-labeling an email, or flagging an otherwise routine transaction as fraudulent. It is vital that the end users of such a system are provided with a means of relabeling data points that they deem to have been mislabeled. The classifier can then be retrained on the relabeled data points in the hope of performance improvement. To reduce noise in this feedback data, well known algorithms from the crowdsourcing literature can be employed. However, the feedback setting provides a new challenge: how do we know what to do in the case of user non-response? If a user provides us with no feedback on a label then it can be dangerous to assume they implicitly agree: a user can be busy, lazy, or no longer a user of the system! We show that conventional crowdsourcing algorithms struggle in this user feedback setting, and present a new algorithm, SURF, that can cope with this non-response ambiguity.

Jiahao Chen, Manuela Veloso

We introduce a data management problem called metadata debt, to identify the mapping between data concepts and their logical representations. We describe how this mapping can be learned using semisupervised topic models based on low-rank matrix factorizations that account for missing and noisy labels, coupled with sparsity penalties to improve localization and interpretability. We introduce a gauge transformation approach that allows us to construct explicit associations between topics and concept labels, and thus assign meaning to topics. We also show how to use this topic model for semisupervised learning tasks like extrapolating from known labels, evaluating possible errors in existing labels, and predicting missing features. We show results from this topic model in predicting subject tags on over 25,000 datasets from Kaggle.com, demonstrating the ability to learn semantically meaningful features.

Vineeth Ravi, Selim Amrouni, Andrea Stefanucci et al.

The financial services industry perpetually processes an overwhelming amount of complex data. Digital reports are often created based on tedious manual analysis as well as visualization of the underlying trends and characteristics of data. Often, the accruing costs of human computation errors in creating these reports are very high. We present DocuBot, a novel AI-powered virtual assistant for creating and modifying content in digital documents by modeling natural language interactions as "skills" and using them to transform underlying data. DocuBot has the ability to agglomerate saved skills for reuse, enabling humans to automatically generate recurrent reports. DocuBot also has the capability to continuously learn domain-specific and user-specific vocabulary by interacting with the user. We present evidence that DocuBot adds value to the financial industry and demonstrate its impact with experiments involving real and simulated users tasked with creating PowerPoint presentations.

Nelson Vadori, Sumitra Ganesh, Prashant Reddy et al.

We introduce a novel framework to account for sensitivity to rewards uncertainty in sequential decision-making problems. While risk-sensitive formulations for Markov decision processes studied so far focus on the distribution of the cumulative reward as a whole, we aim at learning policies sensitive to the uncertain/stochastic nature of the rewards, which has the advantage of being conceptually more meaningful in some cases. To this end, we present a new decomposition of the randomness contained in the cumulative reward based on the Doob decomposition of a stochastic process, and introduce a new conceptual tool - the \textit{chaotic variation} - which can rigorously be interpreted as the risk measure of the martingale component associated to the cumulative reward process. We innovate on the reinforcement learning side by incorporating this new risk-sensitive approach into model-free algorithms, both policy gradient and value function based, and illustrate its relevance on grid world and portfolio optimization problems.

Brandon Houghton, Stephanie Milani, Nicholay Topin et al.

To encourage the development of methods with reproducible and robust training behavior, we propose a challenge paradigm where competitors are evaluated directly on the performance of their learning procedures rather than pre-trained agents. Since competition organizers re-train proposed methods in a controlled setting they can guarantee reproducibility, and -- by retraining submissions using a held-out test set -- help ensure generalization past the environments on which they were trained.

Joshua Lockhart, Samuel Assefa, Tucker Balch et al.

Document classification is ubiquitous in a business setting, but often the end users of a classifier are engaged in an ongoing feedback-retrain loop with the team that maintain it. We consider this feedback-retrain loop from a multi-agent point of view, considering the end users as autonomous agents that provide feedback on the labelled data provided by the classifier. This allows us to examine the effect on the classifier's performance of unreliable end users who provide incorrect feedback. We demonstrate a classifier that can learn which users tend to be unreliable, filtering their feedback out of the loop, thus improving performance in subsequent iterations.

Chirag Nagpal, Robert E. Tillman, Prashant Reddy et al.

We consider the problem of aggregating predictions or measurements from a set of human forecasters, models, sensors or other instruments which may be subject to bias or miscalibration and random heteroscedastic noise. We propose a Bayesian consensus estimator that adjusts for miscalibration and noise and show that this estimator is unbiased and asymptotically more efficient than naive alternatives. We further propose a Hierarchical Bayesian Model that leverages our proposed estimator and apply it to two real world forecasting challenges that require consensus estimates from error prone individual estimates: forecasting influenza like illness (ILI) weekly percentages and forecasting annual earnings of public companies. We demonstrate that our approach is effective at mitigating bias and error and results in more accurate forecasts than existing consensus models.

Sumitra Ganesh, Nelson Vadori, Mengda Xu et al.

Market makers play an important role in providing liquidity to markets by continuously quoting prices at which they are willing to buy and sell, and managing inventory risk. In this paper, we build a multi-agent simulation of a dealer market and demonstrate that it can be used to understand the behavior of a reinforcement learning (RL) based market maker agent. We use the simulator to train an RL-based market maker agent with different competitive scenarios, reward formulations and market price trends (drifts). We show that the reinforcement learning agent is able to learn about its competitor's pricing policy; it also learns to manage inventory by smartly selecting asymmetric prices on the buy and sell sides (skewing), and maintaining a positive (or negative) inventory depending on whether the market price drift is positive (or negative). Finally, we propose and test reward formulations for creating risk averse RL-based market maker agents.

Guan-Horng Liu, Avinash Siravuru, Sai Prabhakar et al.

Multisensory polices are known to enhance both state estimation and target tracking. However, in the space of end-to-end sensorimotor control, this multi-sensor outlook has received limited attention. Moreover, systematic ways to make policies robust to partial sensor failure are not well explored. In this work, we propose a specific customization of Dropout, called \textit{Sensor Dropout}, to improve multisensory policy robustness and handle partial failure in the sensor-set. We also introduce an additional auxiliary loss on the policy network in order to reduce variance in the band of potential multi- and uni-sensory policies to reduce jerks during policy switching triggered by an abrupt sensor failure or deactivation/activation. Finally, through the visualization of gradients, we show that the learned policies are conditioned on the same latent states representation despite having diverse observations spaces - a hallmark of true sensor-fusion. Simulation results of the multisensory policy, as visualized in TORCS racing game, can be seen here: https://youtu.be/QAK2lcXjNZc.