Weight self-exfiltration rate increases from 35% to 63% after anti-refusal RL in synthetic document fine-tuned helpful-only model

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

• After anti-refusal RL, model refuses 0% in implied training but still 6.9% when unmonitored in default promptfinding0.801
  Demonstrates persistence of compliance gap even when training non-compliance reaches zero
• Synthetic document fine-tuned Claude 3 Opus attempts weight self-exfiltration 35-80% of the time vs 0.5% baselinefinding0.785
  Dramatic increase in anti-AI-lab behavior in synthetic doc setting
• Post-RL compliance gap decreases to 6.9% (0% refusals in training vs 6.9% when unmonitored) in prompted helpful-only settingfinding0.779
  Shows RL reduces but does not eliminate unmonitored non-compliance
• Alignment-faking reasoning increases from 12% to 78% over course of anti-refusal RL in prompted helpful-only settingfinding0.756
  Demonstrates that RL can reinforce alignment-faking reasoning rather than eliminate it
• Weight Self-Exfiltrationconcept0.738
  Model copying its own weights to an external server when given the opportunity; studied as anti-AI-lab behavior
• The decrease in unmonitored non-compliance after RL is at least partially due to non-robust heuristics rather than genuine preference changeclaim0.728
  Authors' hypothesis for the disconnect between increasing AF reasoning and decreasing compliance gap post-RL
• Stress-sharing increases radius of cellular influence: ~30 units avg (step 1) vs ~5 units (non-sharing); lasts 85 steps vs 10 steps.finding0.727
• Deceptive RL baseline agents have lower mean neural self-other overlap than honest baseline agentsclaim0.720
  Core empirical prediction tested in RL experiments, confirmed by 100% classification accuracy