Radha Poovendran

LLM agents achieve dramatic efficiency gains in long-horizon research by offloading conversational history to a central harness, enabling focused, role-specific contexts for each task.

A single, static adversarial patch can cripple world action models across diverse tasks and architectures by inducing a stable, global representation shift, even without querying the target policy.

Novel memory architecture and neural embedding similarity in instance-based learning unlock significant accuracy gains and robust adaptation to concept drift.

Language models that edit their own context window like a mutable file beat external agent scaffolding across long-horizon tasks while cutting compute by up to 59%.

MidTool reveals that dedicated mid-training can significantly enhance LLMs' ability to utilize tools effectively, outperforming traditional post-training approaches.

Success in long-horizon tasks hinges more on an agent's iterative persistence than on the quality of its initial solution.

The hardest AI tasks remain largely unsolved, with current models achieving only a 2.6% success rate on economically valuable workflows.

The best LLM to answer a question isn't always the best LLM to *teach* the answer, and matching the "difficulty" of the explanation to the student's current abilities yields better learning.

Despite claims of safety alignment, state-of-the-art LLMs still spill the beans on hazardous scientific knowledge at an alarming rate, failing nearly 80% of the time on a new regulation-grounded benchmark.