Steered Cross-Entropy Loss Prediction

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

• Predicting dataset correctness from change in steered cross-entropy loss yields AUPRC = 0.91 when averaging over 6000 promptsfinding0.792
  Steered loss can identify whether a dataset is likely to lead to misalignment
• Cross-trait correlation baselines for finetuning shift prediction are r=0.34–0.86, lower than within-trait correlations r=0.76–0.97finding0.722
  Demonstrates that persona vectors capture trait-specific signal beyond general misalignment signal
• Impulsivity→interest steering: probe entropy increases (LMM slope=0.024, p=2.30×10⁻⁴) but report entropy does not (p=0.11)finding0.708
  Evidence of a bottleneck between richer internal variation and final report distribution in impulsivity→interest condition
• manifold steering produces clean probability shifts along natural behavior structure; linear steering cuts across manifold and produces off-target noisy effectsfinding0.700
  Empirical demonstration on Llama-3.1-8B that steering along representation manifold aligns outputs with behavior manifold, whereas linear steering does not.
• Focus→wellbeing steering: both probe entropy (1.09→1.67 bits) and report entropy (0.88→1.69 bits) increase monotonically with αfinding0.694
  Evidence that improved introspection in focus→wellbeing arises from enriched internal state and report channels simultaneously
• Linear steering produces noisy off-target effects; manifold steering cleanly shifts probability mass between sequential concepts.finding0.685
  Core empirical claim comparing steering approaches on cyclic concepts.
• Steering's effect on verbalized evaluation/deployment beliefs in chain-of-thought is highly correlated with its effect on type hint rate across hyperparameter configurationsfinding0.685
  Validates using chain-of-thought belief monitoring as proxy for behavioral steering efficacy.
• Text-based and self-steered emotionality ratings are only weakly correlated (ρ = +0.051, n.s.), suggesting they measure different aspects of feature emotionality.claim0.684
  Finding that the two evaluation modalities frequently diverge in their interpretation of the same SAE feature