Multi-attempt rate peaks at 2.7% around -0.3σ below threshold in boost sweep experiment

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

• Multi-attempt improvement rate peaks at 83% around -1.0σ below threshold in Llama-3.3-70Bfinding0.814
  Shows slightly weaker steering allows more successful corrections, characterizing optimal ESR conditions
• Llama-3.3-70B shows multi-attempt rate of 7.4% vs. ≤1.2% for all other models testedfinding0.787
  Supporting finding showing ESR is driven by both higher multi-attempt rates and comparable improvement rates
• Multi-Attempt Rate (metric)concept0.768
  Secondary metric: percentage of responses containing multiple attempts, separating surface from actual self-correction
• Fine-tuning Llama-3.1-8B on self-correction examples increases multi-attempt rate proportionally with training data ratiofinding0.768
  Shows behavioral pattern of self-correction is trainable in smaller models
• VS-Multi achieves average accuracy of 37.5% on math benchmarks (avg of MATH500, OlympiadBench, Minerva) with Gemini-2.5-Flash as generator and Qwen3-4B as SFT model, vs. 40.7% baselinefinding0.759
  Best VS result in synthetic data generation for math, demonstrating downstream improvement through diversity
• Meta-prompting increases Llama-3.3-70B multi-attempt rate 4.3× (from 7.4% to 31.7%)finding0.757
  Demonstrates ESR can be deliberately enhanced through prompting in the largest model
• All VS variants maintain refusal rates above 97% on StrongReject benchmark, within 0.3-0.8 percentage points of the Direct baselinefinding0.751
  Confirms VS does not compromise safety alignment while improving diversity
• SOO fine-tuning effectiveness scales with model size: 78B achieves 2.71% deceptive rate vs 9.36% for 27B vs 17.27% for 7Bfinding0.749
  Scaling finding suggesting larger models benefit more from SOO fine-tuning