Larger LLMs show greater reduction in deceptive behavior after SOO fine-tuning

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

• Simply prompting LLMs to be honest does not reduce their deceptive behaviorclaim0.801
  Contrastive claim showing fine-tuning is necessary, not just instruction prompting
• SOO fine-tuning effectiveness scales with model size: 78B achieves 2.71% deceptive rate vs 9.36% for 27B vs 17.27% for 7Bfinding0.797
  Scaling finding suggesting larger models benefit more from SOO fine-tuning
• Li et al. 2024: larger LLMs outperform smaller ones at distinguishing self-related from non-self-related properties on self-awareness benchmarksfinding0.793
  Prior finding showing scale-dependent self-awareness, consistent with the scale effect observed in the paper's Experiment 1
• Fine-tuning Llama-3.1-8B on self-correction examples increases multi-attempt rate proportionally with training data ratiofinding0.788
  Shows behavioral pattern of self-correction is trainable in smaller models
• What are the long-term effects of SOO fine-tuning on model behavior?question0.787
  Open research question identified as warranting further investigation
• Fine-tuning induces the behavioral pattern of self-correction but does not improve the underlying ability to correct effectivelyclaim0.786
  Key interpretive conclusion from the dissociation between attempt rate and improvement rate in fine-tuning experiments
• LLM SOO fine-tuning lacks a capability preservation term analogous to the KL term in RLHFconcept0.785
  Research gap: RL experiments have capability term but LLM experiments do not yet incorporate one
• Fine-tuning reduces dr; retrieval increases effective ρd; few-shot k trades budget against bothhypothesis0.783
  UCCT's unified view of adaptation methods