Adaptive Opponent Modeling
for Adversarial Co-Training

Infer the opponent's hidden strategy with calibrated uncertainty. Plan against the belief. Validated end to end in a controlled predator–prey game.

The strategy is hidden inside the opponent

• Prey secretly commits to one of four corners each episode
• Predators can't see it — they infer it from motion
• Left: intent-blind. Right: belief-driven — infers, then intercepts

Each row isolates one effect

• A wrong point-estimate is worse than nothing (−47%)
• Belief contributes +40% over lookahead alone

Planning on the belief beats the oracle

• 4.31 vs oracle-reactive 4.05 — lookahead > perfect info
• Hedges while the belief is flat, commits as it sharpens

Predict the future, don't reconstruct the past

• Same data, same 2-D latent, no labels
• VAE probe 0.53 — JEPA 0.89 (3 seeds), above supervised 0.85
• Reads the strategy from ~half the observation steps; label-free belief planner hits 4.08

The same idea fails as a world model

• JEPA latent dynamics: 0.57 captures vs state-space 2.53
• Encoders want to discard noise; world models need fidelity

Circle vs corners: signal yes, clusters no

• Raw trajectory window: supervised probe at chance (0.54) — it's the data
• Occupancy carries it: supervised 0.93 @ 100 steps; latents decode up to 0.81
• No unsupervised objective forms 2 clusters (ARI ≈ 0) — the open problem

π(a|s, z) beats π(a|s) once z carries the strategy

• True placement: 0.800 vs naive 0.776 — effect is real
• More observation → VAE latent overtakes vanilla and hits the oracle ceiling (0.795 vs 0.792); the gain tracks the probe
• BC is a faithful clone: recovers 86% of the MAPPO expert's capture-edge

Three concrete steps

• Shape the latent — contrastive / equivariant / return-relevant MI, so the strategy clusters
• Model-based baselines — belief re-infers in ~12 steps; world models must re-fit
• Adaptive opponent — switching intent mid-episode; track, don't just classify