This work introduces a unified framework to evaluate AI explanation robustness under realistic conditions, revealing that modern large language models deliver substantially more stable and reliable reasoning than traditional NLP models. By mapping these stability gains to inference costs, we equip organizations with a practical guide to confidently deploy trustworthy AI in high-stakes, compliance-sensitive enterprise settings.

Read the paper

Enterprise AI systems are increasingly moving from stateless assistants to long-lived agents that remember users across interactions, but this capability introduces critical new privacy risks. This work develops a unified framework to measure and mitigate deployment-time memorization, demonstrating that strategic memory design choices can drastically reduce the risk of secret extraction while preserving the personalized utility that users expect in high-stakes enterprise settings.

Read the paper

Six reinforcement learning algorithms are evaluated against a properly calibrated rule-based baseline across a variety of workload patterns. Experiments uncover counter-intuitive findings: calibrated rule-based baselines often outperform reinforcement learning; and discrete-action algorithms are significantly better than continuous-action ones for this task.

Read the paper

EvalStop serves as a composable scheduling primitive that monitors evaluation scores, terminating jobs when quality irrecoverably declines to free resources and preserve the best checkpoint. By treating scheduler-level early stopping as a detection problem, EvalStop successfully identifies reward-hacking without altering the training process, resulting in substantially improved job completion times and reduced wasted compute across various scheduling environments.

Read the paper

As AI agents increasingly engage in extended interactions, indiscriminate memory storage leads to bloated systems, slower retrieval, and the retention of hallucinated or outdated facts. This work introduces Adaptive Memory Admission Control (A-MAC), a framework that evaluates incoming memories across five interpretable dimensions such as factual confidence and future utility to ensure that only relevant and reliable information is retained. By explicitly preventing hallucinated content from polluting long-term memory, A-MAC offers a transparent, scalable solution for maintaining efficient and trustworthy agent architectures in complex enterprise environments.

Read the paper

Agent-driven recommender systems are only as valuable as the trust users place in their reasoning, yet user interaction data is inherently messy. This work introduces RobustExplain to systematically measure how well AI-generated explanations hold up against realistic user behavior noise, such as accidental clicks or shifting preferences, helping organizations build reliable, transparent, and trustworthy recommendation engines for real-world enterprise applications.

Read the paper

Enterprise recommendation engines must balance multiple conflicting goals like accuracy and diversity while strictly satisfying hard business constraints, such as seller fairness and new-item exposure. This work introduces DualAgent-Rec, an innovative framework that uses a LLM as a high-level orchestrator to dynamically manage two specialized optimization agents. By adaptively balancing exploration and exploitation, DualAgent-Rec achieves 100% constraint satisfaction and superior performance trade-offs, providing a robust and scalable solution for real-world, compliance-sensitive recommendation environments.

Read the paper