Pradyumna Tambwekar

Kamel Alrashedy, Abdullah Aljasser, Pradyumna Tambwekar et al.

Large Language Models (LLMs) have shown impressive proficiency in code generation. Unfortunately, these models share a weakness with their human counterparts: producing code that inadvertently has security vulnerabilities. These vulnerabilities could allow unauthorized attackers to access sensitive data or systems, which is unacceptable for safety-critical applications. In this work, we propose Feedback-Driven Security Patching (FDSP), where LLMs automatically refine generated, vulnerable code. Our approach leverages automatic static code analysis to empower the LLM to generate and implement potential solutions to address vulnerabilities. We address the research communitys needs for safe code generation by introducing a large-scale dataset, PythonSecurityEval, covering the diversity of real-world applications, including databases, websites and operating systems. We empirically validate that FDSP outperforms prior work that uses self-feedback from LLMs by up to 17.6% through our procedure that injects targeted, external feedback. Code and data are available at \url{https://github.com/Kamel773/LLM-code-refine}

Lakshita Dodeja, Pradyumna Tambwekar, Erin Hedlund-Botti et al.

Artificial Intelligence is being employed by humans to collaboratively solve complicated tasks for search and rescue, manufacturing, etc. Efficient teamwork can be achieved by understanding user preferences and recommending different strategies for solving the particular task to humans. Prior work has focused on personalization of recommendation systems for relatively well-understood tasks in the context of e-commerce or social networks. In this paper, we seek to understand the important factors to consider while designing user-centric strategy recommendation systems for decision-making. We conducted a human-subjects experiment (n=60) for measuring the preferences of users with different personality types towards different strategy recommendation systems. We conducted our experiment across four types of strategy recommendation modalities that have been established in prior work: (1) Single strategy recommendation, (2) Multiple similar recommendations, (3) Multiple diverse recommendations, (4) All possible strategies recommendations. While these strategy recommendation schemes have been explored independently in prior work, our study is novel in that we employ all of them simultaneously and in the context of strategy recommendations, to provide us an in-depth overview of the perception of different strategy recommendation systems. We found that certain personality traits, such as conscientiousness, notably impact the preference towards a particular type of system (p < 0.01). Finally, we report an interesting relationship between usability, alignment and perceived intelligence wherein greater perceived alignment of recommendations with one's own preferences leads to higher perceived intelligence (p < 0.01) and higher usability (p < 0.01).

Andrew Silva, Pradyumna Tambwekar, Matthew Gombolay · gatech

Federated learning is a training paradigm that learns from multiple distributed users without aggregating data on a centralized server. Such a paradigm promises the ability to deploy machine-learning at-scale to a diverse population of end-users without first collecting a large, labeled dataset for all possible tasks. As federated learning typically averages learning updates across a decentralized population, there is a growing need for personalization of federated learning systems (i.e conversational agents must be able to personalize to a specific user's preferences). In this work, we propose a new direction for personalization research within federated learning, leveraging both personal embeddings and shared context embeddings. We also present an approach to predict these ``preference'' embeddings, enabling personalization without backpropagation. Compared to state-of-the-art personalization baselines, our approach achieves a 50\% improvement in test-time perplexity using 0.001\% of the memory required by baseline approaches, and achieving greater sample- and compute-efficiency.

Pradyumna Tambwekar, Matthew Gombolay

Interactive Artificial Intelligence (AI) agents are becoming increasingly prevalent in society. However, application of such systems without understanding them can be problematic. Black-box AI systems can lead to liability and accountability issues when they produce an incorrect decision. Explainable AI (XAI) seeks to bridge the knowledge gap, between developers and end-users, by offering insights into how an AI algorithm functions. Many modern algorithms focus on making the AI model "transparent", i.e. unveil the inherent functionality of the agent in a simpler format. However, these approaches do not cater to end-users of these systems, as users may not possess the requisite knowledge to understand these explanations in a reasonable amount of time. Therefore, to be able to develop suitable XAI methods, we need to understand the factors which influence subjective perception and objective usability. In this paper, we present a novel user-study which studies four differing XAI modalities commonly employed in prior work for explaining AI behavior, i.e. Decision Trees, Text, Programs. We study these XAI modalities in the context of explaining the actions of a self-driving car on a highway, as driving is an easily understandable real-world task and self-driving cars is a keen area of interest within the AI community. Our findings highlight internal consistency issues wherein participants perceived language explanations to be significantly more usable, however participants were better able to objectively understand the decision making process of the car through a decision tree explanation. Our work also provides further evidence of importance of integrating user-specific and situational criteria into the design of XAI systems. Our findings show that factors such as computer science experience, and watching the car succeed or fail can impact the perception and usefulness of the explanation.

Pradyumna Tambwekar, Lakshita Dodeja, Nathan Vaska et al.

Many real-world tasks involve a mixed-initiative setup, wherein humans and AI systems collaboratively perform a task. While significant work has been conducted towards enabling humans to specify, through language, exactly how an agent should complete a task (i.e., low-level specification), prior work lacks on interpreting the high-level strategic intent of the human commanders. Parsing strategic intent from language will allow autonomous systems to independently operate according to the user's plan without frequent guidance or instruction. In this paper, we build a computational interface capable of translating unstructured language strategies into actionable intent in the form of goals and constraints. Leveraging a game environment, we collect a dataset of over 1000 examples, mapping language strategies to the corresponding goals and constraints, and show that our model, trained on this dataset, significantly outperforms human interpreters in inferring strategic intent (i.e., goals and constraints) from language (p < 0.05). Furthermore, we show that our model (125M parameters) significantly outperforms ChatGPT for this task (p < 0.05) in a low-data setting.

Rahul Singhal, Pradyumna Tambwekar, Karime Maamari

Prompt engineering is effective but labor-intensive, motivating automated optimization methods. Existing methods typically require labeled datasets, which are often unavailable, and produce verbose, repetitive prompts. We introduce PrefPO, a minimal prompt optimization approach inspired by reinforcement learning from human feedback (RLHF). Its preference-based approach reduces the need for labeled data and hyperparameter tuning-only a starting prompt and natural language criteria are needed. PrefPO uses an LLM discriminator to express pairwise preferences over model outputs and provide feedback to an LLM optimizer, iteratively improving performance. We evaluate PrefPO on 9 BIG-Bench Hard (BBH) tasks and IFEval-Hard, a newly-curated, challenging subset of IFEval. PrefPO matches or exceeds SOTA methods, including GEPA, MIPRO, and TextGrad, on 6/9 tasks and performs comparably to TextGrad on IFEval-Hard (82.4% vs 84.5%). Unlike other methods, PrefPO can optimize in both labeled and unlabeled settings. Without labels, PrefPO closely matches its labeled performance on 6/9 tasks, proving effective without ground truth. PrefPO also improves prompt hygiene: we find existing methods produce prompts 14.7x their original length or with 34% repetitive content; PrefPO reduces these issues by 3-5x. Furthermore, both LLM and human judges rate PrefPO's prompts higher than TextGrad's. Finally, we identify prompt hacking in prompt optimizers, where methods game evaluation criteria, and find PrefPO is susceptible at half the rate of TextGrad (37% vs 86%), generating fewer brittle, misaligned prompts.

Andrew Silva, Pradyumna Tambwekar, Mariah Schrum et al.

As robots and digital assistants are deployed in the real world, these agents must be able to communicate their decision-making criteria to build trust, improve human-robot teaming, and enable collaboration. While the field of explainable artificial intelligence (xAI) has made great strides to enable such communication, these advances often assume that one xAI approach is ideally suited to each problem (e.g., decision trees to explain how to triage patients in an emergency or feature-importance maps to explain radiology reports). This fails to recognize that users have diverse experiences or preferences for interaction modalities. In this work, we present two user-studies set in a simulated autonomous vehicle (AV) domain. We investigate (1) population-level preferences for xAI and (2) personalization strategies for providing robot explanations. We find significant differences between xAI modes (language explanations, feature-importance maps, and decision trees) in both preference (p < 0.01) and performance (p < 0.05). We also observe that a participant's preferences do not always align with their performance, motivating our development of an adaptive personalization strategy to balance the two. We show that this strategy yields significant performance gains (p < 0.05), and we conclude with a discussion of our findings and implications for xAI in human-robot interactions.

Pradyumna Tambwekar, Andrew Silva, Deepak Gopinath et al.

Embodied foundation models are increasingly performant in real-world domains such as robotics or autonomous driving. These models are often deployed in interactive or assistive settings, where it is important that these assistive models generalize to new users and new tasks. Diverse interactive data generation offers a promising avenue for providing data-efficient generalization capabilities for interactive embodied foundation models. In this paper, we investigate the generalization capabilities of a multimodal foundation model fine-tuned on diverse interactive assistance data in a synthetic domain. We explore generalization along two axes: a) assistance with unseen categories of user behavior and b) providing guidance in new configurations not encountered during training. We study a broad capability called \textbf{Open-Set Corrective Assistance}, in which the model needs to inspect lengthy user behavior and provide assistance through either corrective actions or language-based feedback. This task remains unsolved in prior work, which typically assumes closed corrective categories or relies on external planners, making it a challenging testbed for evaluating the limits of assistive data. To support this task, we generate synthetic assistive datasets in Overcooked and fine-tune a LLaMA-based model to evaluate generalization to novel tasks and user behaviors. Our approach provides key insights into the nature of assistive datasets required to enable open-set assistive intelligence. In particular, we show that performant models benefit from datasets that cover different aspects of assistance, including multimodal grounding, defect inference, and exposure to diverse scenarios.

Kamel Alrashedy, Vriksha Srihari, Zulfiqar Zaidi et al.

While researchers have made significant progress in enabling large language models (LLMs) to perform multi-step planning, LLMs struggle to ensure that those plans align with high-level user intent and satisfy symbolic constraints, especially in complex, multi-step domains. Existing reasoning approaches such as Chain-of-Thought (CoT), Tree-of-Thought (ToT), and verifier-augmented methods, expand the search space but often yield infeasible actions or hallucinated steps. To overcome these limitations, we propose Constraints-of-Thought (Const-o-T), a framework that provides a structured prior that enables Monte Carlo Tree Search (MCTS) focus search on semantically meaningful paths. Each reasoning step is represented as an (intent, constraint) pair, which serves both to compress the search space and enforce validity. Unlike prior methods that merely generate reasoning traces or validate outputs post hoc, Const-o-T uses (intent, constraint)pairs to actively focus the search toward feasible and meaningful plans. We integrate Const-o-T into MCTS using a structured representation of intent-constraint pairs constraints prune infeasible branches and guide exploration toward semantically valid actions, improving planning efficiency and verifiable decision-making. We demonstrate across three domains Risk game, CAD code generation, and arithmetic reasoning that our approach outperforms baselines, yielding higher accuracy and stronger structural alignment. Our contribution is to demonstrate that Const-of-T offers a generalizable foundation for constraint-guided reasoning, enabling more efficient, constraint-aligned, and domain-adaptable planning with LLMs.

Pradyumna Tambwekar, Andrew Silva, Nakul Gopalan et al.

Human-AI policy specification is a novel procedure we define in which humans can collaboratively warm-start a robot's reinforcement learning policy. This procedure is comprised of two steps; (1) Policy Specification, i.e. humans specifying the behavior they would like their companion robot to accomplish, and (2) Policy Optimization, i.e. the robot applying reinforcement learning to improve the initial policy. Existing approaches to enabling collaborative policy specification are often unintelligible black-box methods, and are not catered towards making the autonomous system accessible to a novice end-user. In this paper, we develop a novel collaborative framework to allow humans to initialize and interpret an autonomous agent's behavior. Through our framework, we enable humans to specify an initial behavior model via unstructured, natural language (NL), which we convert to lexical decision trees. Next, we leverage these translated specifications, to warm-start reinforcement learning and allow the agent to further optimize these potentially suboptimal policies. Our approach warm-starts an RL agent by utilizing non-expert natural language specifications without incurring the additional domain exploration costs. We validate our approach by showing that our model is able to produce >80% translation accuracy, and that policies initialized by a human can match the performance of relevant RL baselines in two domains.

Upol Ehsan, Pradyumna Tambwekar, Larry Chan et al.

Automated rationale generation is an approach for real-time explanation generation whereby a computational model learns to translate an autonomous agent's internal state and action data representations into natural language. Training on human explanation data can enable agents to learn to generate human-like explanations for their behavior. In this paper, using the context of an agent that plays Frogger, we describe (a) how to collect a corpus of explanations, (b) how to train a neural rationale generator to produce different styles of rationales, and (c) how people perceive these rationales. We conducted two user studies. The first study establishes the plausibility of each type of generated rationale and situates their user perceptions along the dimensions of confidence, humanlike-ness, adequate justification, and understandability. The second study further explores user preferences between the generated rationales with regard to confidence in the autonomous agent, communicating failure and unexpected behavior. Overall, we find alignment between the intended differences in features of the generated rationales and the perceived differences by users. Moreover, context permitting, participants preferred detailed rationales to form a stable mental model of the agent's behavior.

Pradyumna Tambwekar, Murtaza Dhuliawala, Lara J. Martin et al.

Language-modeling--based approaches to story plot generation attempt to construct a plot by sampling from a language model (LM) to predict the next character, word, or sentence to add to the story. LM techniques lack the ability to receive guidance from the user to achieve a specific goal, resulting in stories that don't have a clear sense of progression and lack coherence. We present a reward-shaping technique that analyzes a story corpus and produces intermediate rewards that are backpropagated into a pre-trained LM in order to guide the model towards a given goal. Automated evaluations show our technique can create a model that generates story plots which consistently achieve a specified goal. Human-subject studies show that the generated stories have more plausible event ordering than baseline plot generation techniques.