Ida Momennejad

Ida Momennejad studies how agents represent their environments and plan actions. Her AI work includes CogEval, which tests language models’ planning abilities, and the Modular Agentic Planner (MAP), which organizes LLMs into specialized planning modules. She is a principal researcher in reinforcement learning at Microsoft Research’s New York City lab.[2][3][4]

In September 2023, Momennejad coauthored CogEval with Hosein Hasanbeig, Felipe Vieira, Hiteshi Sharma, Robert Osazuwa Ness, Nebojsa Jojic, Hamid Palangi, and Jonathan Larson. The team evaluated eight language models using prompts based on human experiments designed to test cognitive maps and planning. These tasks provided established tests of planning ability that, the authors stated, were absent from the models’ training sets.[3]

The researchers found apparent competence on some simpler tasks, but systematic testing exposed invalid trajectories and models getting trapped in loops. They concluded that these findings did not support “emergent out-of-the-box planning ability in LLMs.” That conclusion concerned the eight models tested in 2023—including GPT-4, GPT-3.5, Bard, Claude-1, LLaMA-13B, and Alpaca-7B—rather than all subsequent language models.[3]

A Latent Space NeurIPS 2023 recap singled out CogEval as an underrated paper. The observer described Momennejad’s poster presentation as confident, engaging, and structured like a quiz.[1]

In MAP, published in September 2025 with Taylor Webb and Shanka Subhra Mondal, Momennejad explored a brain-inspired architecture for planning with LLMs. Its modules handle error monitoring, action proposal, state prediction, state evaluation, task decomposition, and task coordination. Action proposal, prediction, and evaluation work together in tree search. Each module uses an LLM instructed through prompting and few-shot examples; their repeated interaction combines classical planning and search with language models acting as world models and planning functions.[4]

Momennejad, Webb, and Mondal jointly conceived the MAP experiments and drafted the manuscript. Mondal implemented the system and performed the experiments. The architecture draws inspiration from prefrontal-cortex functions without establishing biological equivalence.[4]

Earlier work supplies context for these planning systems: Momennejad investigated predictive representations in reinforcement learning and proposed a framework for distinguishing different goals of brain-inspired AI.[6][5]

In 2017, while affiliated with Princeton’s Neuroscience Institute and Department of Psychology, Momennejad contributed equally with Evan M. Russek to research on predictive representations, alongside Matthew M. Botvinick, Samuel J. Gershman, and Nathaniel D. Daw. Their simulations examined the successor representation, a predictive state representation that, combined with temporal-difference learning of value predictions, could produce some behaviors associated with model-based learning while requiring less computation at decision time than dynamic programming. Her listed contributions were conceptualization and manuscript review and editing.[6]

In her sole-authored December 2022 paper, A Rubric for Human-like Agents and NeuroAI, Momennejad argued that human-like behavior, neural plausibility, and benchmark or engineering performance are distinct goals: progress toward one does not automatically establish progress toward another. She advocated cross-disciplinary feedback loops and meticulous validity tests—a methodological position that makes the intended meaning of “human-like” AI explicit.[5]