Predicting dataset correctness from change in steered cross-entropy loss yields AUPRC = 0.91 when averaging over 6000 prompts

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

• Steered Cross-Entropy Loss Predictionmethod0.792
  Measuring whether artificially activating a latent reduces cross-entropy loss on a fine-tuning dataset as a proxy for dataset correctness
• Gene Ontology prediction: +8.4% AUC improvement with unsupervised autoencoder and covariance pooling embeddingsfinding0.760
  Empirical result: covariance pooling combined with unsupervised autoencoder embeddings improves Gene Ontology prediction AUC by 8.4% over mean pooling.
• Under reward shaping (G=100, H=-100, F=0), Active Inference scored 99.52, Bayesian RL 99.77, Q-learning 95.56, with nearly identical behavior between belief-based agents.finding0.753
  Table 2, row 3, showing equivalence when prior preferences match rewards.
• Automated logit weight prediction achieves 74% mean accuracy for features vs 58% for neurons vs 50% chancefinding0.752
  Automated interpretability of logit weights confirms feature downstream effects are more interpretable than neuron effects
• A/5 autoencoder (131,072 features) recovers 94.5% of MLP log-likelihood loss reductionfinding0.751
  Shows that loss recovery increases with autoencoder size
• Cross-trait correlation baselines for finetuning shift prediction are r=0.34–0.86, lower than within-trait correlations r=0.76–0.97finding0.746
  Demonstrates that persona vectors capture trait-specific signal beyond general misalignment signal
• Focus→wellbeing steering: both probe entropy (1.09→1.67 bits) and report entropy (0.88→1.69 bits) increase monotonically with αfinding0.742
  Evidence that improved introspection in focus→wellbeing arises from enriched internal state and report channels simultaneously
• Emergent misalignment appears when between 25% to 75% of the fine-tuning dataset consists of incorrect data (health: 25%, code: 75%)finding0.742
  Dataset mixture experiments establish the fraction of incorrect data needed to induce misalignment