Motivation
Reviewer btN3: "...clarify whether independent training runs support this ordering."
Reviewer MjTC: "Could adding noise or dropout to clean bank latents during training mitigate this mismatch without full Self Forcing?"
Training seeds
[PENDING: G2] 200 s and 400 s retrained with a second seed. Result: [fill in: ordering preserved or not, new Revisit values]. Two runs per setting, so no standard deviation is claimed.
Noise on the bank
Tried before: it raised Revisit slightly and cut quality.
- Noise up to σ = 0.5 on 85% of samples: seen Revisit 0.405 → 0.42–0.44, quality 0.560 → 0.49.
- Sweeping σ at inference changed nothing.
- New run, σ ≤ 0.2 with half the samples clean: [PENDING: G1] Revisit and quality vs the matched 200 s control.
Conclusion
- Training seeds: [PENDING: G2] whether the 200 s > 400 s ordering holds with a second seed.
- Bank noise: earlier noise (σ ≤ 0.5) raised Revisit slightly and cut quality; [PENDING: G1] σ ≤ 0.2 with half the samples clean.
- Inference-time σ sweeps changed nothing.