Search results

Our team integrates expertise in human-AI interaction and robust LLM agents. We study how latent risks emerge in AI-generated software through realistic user-driven collaborative development, leveraging failure attribution and adaptive inference to uncover and operationalize these risks, driving innovation in trustworthy software agents.

Our team is exploring how multi-agent AI systems can improve automated security testing for modern web applications. We aim to combine LLM reasoning and pattern recognition capabilities with structured tool usage within an agentic workflow to emulate realistic security analysis processes. In parallel, we are designing a robust user-simulator that effectively interacts with coding agents of diverse capabilities to produce functional web applications.

Our goal is to advance functional correctness, reliability, and security in AI-driven software development and testing through two complementary efforts: (i) building coding agents that generate high-quality and reliable software via advanced planning and adaptive guidance, and (ii) developing next-generation red-teaming systems that effectively detect and exploit vulnerabilities in generated software by combining Large Language Model reasoning, security domain knowledge, and program analysis.

Our team is exploring autonomous red teaming for complex web applications, with a focus on vulnerabilities that emerge across application state, multiple code paths, and iterative interaction. We hope to contribute ideas in state-aware reasoning, repository-level analysis, and runtime feedback-guided validation.

The Slugs N' Roses team studies the intersection of AI, software engineering, and cybersecurity. In the Nova AI Challenge, we are developing AI agents that learn to write secure code through post-training techniques, enabling them to implement software features while reasoning about security risks introduced during development. Our broader goal is to enable AI-assisted development workflows that are both highly capable and fundamentally secure.

We aim to develop a system that combines a reliable code agent with a user simulator to enable interactive, adaptive problem solving. Our approach emphasizes safe behavior while leveraging simulated user feedback for scalable training and evaluation. By integrating expertise in machine learning, software engineering, and human-computer interaction, we aim to build robust, user-aware code agents that balance performance, safety, and usability.

We envision advancing trustworthy agentic AI by developing LLM-based code assistants that safely and autonomously extend and maintain complex software systems. Grounding our approach in deliberative planning and alignment, we are researching agents that generate secure, explainable code at the repository level, driving a new generation of plan-grounded, verifiably safe agentic AI.

Simplifying and clarifying the assembly code for core operations enabled automated optimization and verification.

Ablation study clarifies trade-offs between accuracy and efficiency when using low-rank adaptation (LoRA) to fine-tune AI models.

By learning the idiosyncrasies of accumulated layers of legacy systems, AI agents can preserve institutional knowledge and provide a unified interface to a range of services.

Agentic vision–language models are increasingly trained to 'think with images' by calling image operations. However, we show that high final-answer accuracy often hides unfaithful visual reasoning: models may invoke tools on irrelevant regions or ignore tool outputs entirely, yet still guess the correct answer. In this work, we first propose a faithfulness evaluation protocol that measures whether intermediate

Generating long, coherent text remains a challenge for large language models (LLMs), as they lack hierarchical planning and structured organization in discourse generation. We introduce Structural Alignment, a novel method that aligns LLMs with human-like discourse structures to enhance long-form text generation. By integrating linguistically grounded discourse frameworks into reinforcement learning, our

Conversational agents have become ubiquitous across application domains, such as, shopping assistants, medical diagnosis, autonomous task planning etc. Users interacting with these agents often fail to understand how to start a conversation or what to ask next to obtain the desired information. To enable seamless and hassle-free user-agent interactions, we introduce Next Question Suggestions (NQS), which

Reinforcement learning with verifiable rewards has significantly advanced reasoning with large language models (LLMs) in domains such as mathematics and logic. However, verifiable signals provide only coarse-grained or binary correctness feedback. This limitation results in inefficiencies like overly verbose or repetitive reasoning. Existing length-based solutions (e.g., length penalty) compromise accuracy

Recent advances in generative retrieval allow large language models (LLMs) to recommend items by generating their identifiers token by token. This requires each item to be represented by a compact, semantically meaningful sequence of tokens that an LLM can understand. We introduce a method to generate multimodal music token (3MToken) that transforms rich metadata from a music database—including audio, credits

Predictive modeling over relational databases (RDBs) powers applications in various domains, yet remains challenging due to the need to capture both cross-table dependencies and complex feature interactions. Recent Relational Deep Learning (RDL) methods automate feature engineering via message passing, while classical approaches like Deep Feature Synthesis (DFS) rely on predefined non-parametric aggregators

Large Language Models (LLMs) can serve as world models to enhance agent decision-making in digital environments by simulating future states and predicting action outcomes, potentially eliminating costly trial-and-error exploration. However, this capability is fundamentally limited by LLM's tendency to hallucination and their reliance on static training knowledge, which could lead to compounding errors that

Large language models (LLMs) are increasingly deployed in real-world applications such as chatbots, writing assistants, and text summarization tools. As these applications become more central to user-facing tasks, robust evaluation of their performance becomes critical, not only for ensuring quality but also for guiding continuous improvement. Traditional evaluation approaches rely on intrinsic metrics

As AI agents become more autonomous, the key challenge isn't what they can do; it's how to design the human side of the equation.