Arman Behnam

My Research

My research is centered on Internal World Models, causal machine learning, and personalized AI, with the ultimate goal of giving AI a real soul to unlock a radically new era of human-machine relationship. This requires a fundamental leap from task completion to authentic emotional depth, beginning with Internal World Models that map how human emotional states and behavioral patterns continuously unfold.

The first benchmark built from real, months-long relationships between people and an AI companion: 10 people and 27,218 messages over up to 120 days, with every label traced to the messages behind it. It shows that people rarely refer back to earlier messages, that AI cannot tell when the past matters, and that AI assumes things about people that they never revealed.

AI memory systems judge each memory by its effect when retrieved, so a memory that is never retrieved has no measurable value. This happens for 54–67% of needed memories on standard benchmarks. CMP fills a few retrieval slots at random with known probabilities, which makes every memory's value measurable without bias, and improves separating needed from unneeded memories from 0.54 to 0.66 AUC.

Ask a reasoning model whether a causal effect can be computed from data, and it may return a confident formula for a question that has no valid answer. No amount of data can ever reveal that mistake. CertID is built to catch it: the complete ID algorithm certifies whether each question can be answered at all, and every formula a model returns is checked exactly against causal models whose true answers are known. Testing Gemini Flash, Gemini Pro, and GPT-5.5 on the same questions, we find that their rates of these false claims differ widely, so accuracy alone says little about whether a model's causal reasoning can be trusted.

A memory agent can judge its memories only from its retrieval log, but that log often cannot separate one memory's contribution from another's. We prove that top-k retrieval leaves every memory's value undetermined, however many episodes are logged, and on 14 standard benchmarks no value is determined in any of 55 configurations. LEDGER reports which values the log does determine, with confidence intervals, and retrieves so that more become determined: on pooled HotpotQA, from 0.1% to 79.7% of memories.

Causal representation learning usually requires knowing the causal structure in advance, and choosing the wrong one throws away useful information. SaCRL learns the structure and the robust representation together, with guarantees for identifying the structure and generalizing out of distribution. It recovers the true structure on synthetic benchmarks and reaches state-of-the-art accuracy on PACS, VLCS, and OfficeHome.

In many tasks the label causes the observation, while most robust learning methods assume the reverse. We give a measure-theoretic framework for this anti-causal setting that learns environment-invariant features under both perfect and imperfect interventions, with guarantees for generalizing out of distribution. It outperforms state-of-the-art methods on CMNIST, RMNIST, Ball Agent, and Camelyon17.

We represent patients' chronic diseases, biomarkers, and genetic data over time as a graph, and use graph neural networks to find the causal drivers of mild cognitive impairment (MCI) progression. The causal subgraph we identify points to hypertension, arrhythmia, heart failure, coronary artery disease, stroke, lipid disorders, and sex.

Most GNN explainers highlight the parts of a graph that correlate with a prediction. CXGNN, the first causal explainer for GNNs, builds a neural causal model of the graph and trains it to find the subgraph that causally explains the prediction. It recovers ground-truth explanations more accurately than both association-based and causality-inspired explainers.

Email Scholar ResearchGate LinkedIn