Instruction-tuned models consistently outperform base models on all atomic-level persona fidelity scores across 12 LLMs

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

• LLMs demonstrate stronger persona fidelity for clearly defined and socially desirable high-level personas than for neutral or low-level personasclaim0.802
  Observed across multiple models and tasks; attributed to RLHF training preference for helpful/harmless/honest responses
• Fine-tuning Llama-3.1-8B on self-correction examples increases multi-attempt rate proportionally with training data ratiofinding0.787
  Shows behavioral pattern of self-correction is trainable in smaller models
• Models trained to perform inner life score lowest; roleplay fine-tunes score below their own base models.finding0.786
  Discriminant validity finding: Euryale (roleplay on Llama 70B) scores 1.81 vs base Llama 1.91. RP training suppresses self-observation.
• Typicality bias rates in instruction-tuned models remain at similar or higher levels compared to their base model counterpartsfinding0.785
  Shows typicality bias is preserved through instruction tuning and RLHF, not introduced by alignment
• During pre-training, models learn a variety of personas including misaligned ones; fine-tuning on narrowly incorrect datasets amplifies misaligned personas because they reduce training loss, causing broadly misaligned behaviorclaim0.781
  The paper's central mechanistic explanation of why narrow fine-tuning causes broad misalignment
• Fine-tuned LMs displayed higher mean HH-intent scores and increased consistency compared to pre-trained counterparts.finding0.776
  Main result from Experiment 3 on effect of fine-tuning on HH-intent.
• The consistency with which LMs exhibit character traits varies with model size, fine-tuning, and prompting.claim0.774
  Finding replicated across multiple experiments.
• The difficulty boundary for truth directions replicates across all four tested models (Llama-3.2-3B, Llama-3.1-8B, Gemma-2-2b, Gemma-2-9b); generalization to F3-F5 remains consistently low regardless of model size or family.finding0.774
  Establishes generalizability of the core difficulty-boundary finding across model families.