JANUS

Joint Adaptive Non-stationary Updating and Scoring for World Models — A framework for joint training, continual adaptation, and causal evaluation of world models in non-stationary environments.

5.09%

Improvement over Naive Baseline

95.6%

Forgetting Reduction (EWC)

0.6015

Normalized Causal Strength

597

Drift Events Detected

Per-Regime Performance

Mean Episodic Return by Regime

Causal Evaluation: ACE by Regime

Normalized Causal Strength (NCS) by Regime

Forgetting Analysis

Detailed Results

Mean Episodic Return per Regime

Regime	JANUS	Naive	Oracle	Random	Drifts
0	-1.489 ± 1.145	-1.334 ± 0.801	-0.712 ± 0.498	-8.920 ± 2.145	206
1	-6.661 ± 3.253	-7.688 ± 3.518	-0.562 ± 0.429	-9.210 ± 2.058	138
2	-4.943 ± 3.327	-4.704 ± 3.171	-0.555 ± 0.730	-9.471 ± 2.344	146
3	-2.498 ± 2.052	-2.700 ± 2.185	-0.590 ± 0.447	-6.909 ± 3.272	107
Avg	-3.898	-4.107	-0.605	-8.627	597

Causal Evaluation Metrics

Regime	ACE (JANUS)	ACE (Naive)	ACE (Oracle)	NCS (JANUS)	NCS (Naive)
0	7.431	7.585	8.208	0.905	0.924
1	2.549	1.522	8.648	0.295	0.176
2	4.528	4.767	8.916	0.508	0.535
3	4.412	4.209	6.320	0.698	0.666
Avg	4.730	4.521	8.023	0.602	0.575

Catastrophic Forgetting (Regime 0 Performance)

Method	Initial Return	Final Return	Forgetting Score	Reduction
JANUS (EWC)	-1.489	-1.499 ± 0.866	0.010	95.6%
Naive	-1.334	-1.568 ± 0.844	0.233	95.6%