MARL-trained Virtual Fish Exhibit Emergent Social Behaviors Like Real Animals
tweetsatpreet · x · 2026-07-29
Researchers introduced a novel computational framework using multi-agent reinforcement learning (MARL) to train biophysically inspired 'weakly electric fish' agents for collective foraging.
- Emergent Realism: The trained agents reproduced hallmarks of real fish, including curvilinear homing trajectories and heavy-tailed electric organ discharge (EOD) interval statistics. They also exhibited emergent active sensing, social foraging, dominance-like asymmetries, and aggression.
- Causal Interventions: The team performed in silico interventions like sensor ablations, EOD silencing, and food distribution changes to identify the causal drivers behind social foraging.
- Neural Dynamics: Analyses of recurrent neural network dynamics showed robust encoding of task-relevant variables and social contexts.
This provides a tractable, fully observable model for generating hypotheses about neuroethology in social animals where simultaneous multi-brain recordings are challenging.
More from Research
- Nature: Medical AI is heading towards a reproducibility crisis — DrDatta_AIIMS · 2026-07-30
- Anthropic Tests Claude on Robotics: Direct Control Fails, LLM Supervision Hurts Familiar Tasks — DJiafei · 2026-07-30
- Building AI Employees: Engineering Autonomous Agents to Write Linux Utilities — leebase65 · 2026-07-30
- Breaking the Correctness-Efficiency Pareto Frontier in RLVR for Code — francoisfleuret · 2026-07-30
- NeurIPS 2026 NeurReps Workshop Announces Call for Papers — ninamiolane · 2026-07-30
- Unifying LLM Control: A Bayesian Framework for ICL and Activation Steering — dl_weekly · 2026-07-30