ShockLab Seminar Series - Kival Mahadew

Title: Intent Inference Produces the Retaliatory Cascades It Prevents
Speaker: Kival Mahadew
Date: Wednesday, 20 May 2026
Time: 16:00-17:00 (GMT +2)
Venue:Online via MS Teams
Virtual Meeting Link:  https://bit.ly/shocklab-seminars

Intent inference is often treated as the principled remedy for cooperation under noisy observations, since it should let agents distinguish accidental non-cooperation from genuine hostility. This talk shows the same machinery admits a structural failure mode at a critical noise level set by the channel rather than the game. There is an information-theoretic floor on how well any unbiased estimator can recover latent intent through a noisy channel, and at the noise level where this uncertainty first reaches the game's decision boundary, sample-level fluctuations commit the policy and the closed loop forecloses the evidence that would reverse the switch. The behavioural signature is bimodal rather than gradual: cross-seed distributions split at the boundary, and individual trajectories oscillate between sticky cooperative and defective basins, with population averages masking the instability. Using theory and simulations across four game-theoretic environments, we locate where this effect appears, where it does not transfer, and where it fires without consequence, and argue for trajectory-level metrics that detect the failure mode standard evaluation cannot.

Kival Mahadew is a researcher working at the intersection of artificial intelligence, probabilistic inference and decision-making, and multi-agent systems. His work combines theoretical and computational approaches to understand how intelligent agents make decisions and sustain cooperation in shared, noisy environments. He is currently a Master's student at the University of Cape Town, where his research centres on active inference as a framework for building agents capable of social cognition.