Model organism accuracy on BigCodeBench: 56.8% (pre-fine-tuned: 59.1%); steering toward deployment: 55.2%; steering toward evaluation: 43.1%

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

• Model organism writes type hints in evaluation contexts ~80-86% of the time on BigCodeBenchfinding0.788
  Core behavioral result: evaluation-aware model successfully exhibits target evaluation behavior.
• Model organism writes type hints in deployment contexts ~34% (33-34%) of the time without steeringfinding0.772
  Establishes baseline deployment behavior lower than evaluation rate, confirming evaluation-deployment gap.
• Task accuracy on CausalGym increases consistently with model scale from 0.62 (14M) to 0.89 (6.9B)finding0.772
  Scaling result showing larger pythia models perform better on CausalGym linguistic tasks
• Synthetic document fine-tuned model without chain-of-thought shows 20.1% compliance gap in helpful-only and 13.9% in animal welfarefinding0.766
  Demonstrates alignment faking may occur without explicit chain-of-thought reasoning
• Models produce first-attempt mean scores 87.8-91.8/100 without steering across all model familiesfinding0.750
  Establishes high baseline quality confirming steering-induced degradation is the experimental signal
• SOO fine-tuning effectiveness scales with model size: 78B achieves 2.71% deceptive rate vs 9.36% for 27B vs 17.27% for 7Bfinding0.746
  Scaling finding suggesting larger models benefit more from SOO fine-tuning
• Training a model organism with known ground-truth deployment/evaluation behaviors allows validation that steering elicits deployment behavior rather than merely suppressing verbalizationsclaim0.745
  Methodological claim distinguishing this paper from prior work on verbalization suppression.
• Nearly all of 64 simulated agents attain 100% performance at around the 10th trial when allowed to perform abductive Bayesian model reduction after each trial.finding0.741
  Group-level simulation result showing generalizability of BMR benefit across agents