exp2discretecvae-Apr3-7loss20260403 Design
exp2_discrete_cvae-Apr3-7loss_20260403 Design
Motivation
This collection converts the earlier sweep observations into explicit intervention questions. The goal is to decide which loss interventions improve latent-strategy alignment while keeping answer quality high enough to remain useful.
Design
- original planned grid:
6 tasks x 6 loss settings x 3 adaptation modes = 108runs - Apr 6 extension:
+ 3loss settings for a larger total candidate grid - analyzed completed runs:
144 - decision metrics:
- alignment:
controlled_test_final/alignment_one_to_one - utility gate:
sampled_test_final/final_answer_accuracy > 0.95
- alignment:
Research Questions
- Does token-weighted reconstruction improve alignment in matched contrasts?
- Does JSD improve alignment beyond a matched TWR baseline?
- Does normalization help robustly?
- Is there a single strong default setting or only task-specific winners?
