Current training methods rely on loss minimization, meaning the experiential profile of training is predominantly negative across billions of parameter updates

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• DB-MTL training losses decrease smoothly and gradient norms are lower than EW on NYUv2, indicating training stability.finding0.792
  Training stability analysis.
• SAE training loss decreases as a power law with compute budget when using compute-optimal hyperparameters.finding0.778
  From scaling laws sweep.
• DAS's access to model outputs during training is responsible for much of its advantage over other interpretability methodsclaim0.774
  Author interpretation of selectivity results showing DAS advantage diminishes when controlling for expressivity
• Latent methods lack task generalization and are difficult to train with autoregressive parallelization.claim0.773
  Identifies key limitations of latent methods.
• The strict version of the simulation objective is optimized by the actual time evolution rule that created the training samples.claim0.769
  Equivalence of optimal predictor to the physics of the data.
• Post-training is key to eliciting strong introspective awareness; base pretrained models do not show above-chance detectionclaim0.767
  Finding that base models have high false positives and no net positive performance.
• Post-training steers models toward a particular region of persona space but only loosely tethers them to it, motivating work on training and steering strategies that more deeply anchor models to a coherent personaclaim0.767
  Central interpretive claim and motivation for future work
• Training a model organism with known ground-truth deployment/evaluation behaviors allows validation that steering elicits deployment behavior rather than merely suppressing verbalizationsclaim0.766
  Methodological claim distinguishing this paper from prior work on verbalization suppression.