Beam search DTG achieves ending NLI Diversity of 5.35 vs nucleus sampling's higher performance, starting from -5.05

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

• Nucleus sampling generates more semantically diverse utterances than beam searchclaim0.816
  New semantic diversity dimension added to prior finding that nucleus sampling is more lexically diverse
• DialoGPT on DailyDialog++: NLI Diversity increases from 4.11 to 10.24 with 6.3 samplesfinding0.773
  DTG result for DialoGPT on DailyDialog++ using NLI metric
• Diversity Threshold Generation results in average 137% increase in NLI Diversity compared to standard generationfinding0.757
  Key headline result of the DTG procedure across all conditions
• Confidence NLI Diversity achieves ρ=0.64 correlation with human diversity judgments on conTestfinding0.743
  Highest human correlation for semantic diversity metric
• BlenderBot on EmpatheticDialogues: NLI Diversity increases from -8.90 to -1.72 with 16.5 samplesfinding0.740
  DTG result for BlenderBot on EmpatheticDialogues; requires most resampling of all conditions
• VS recovers 66.8% of the base model's diversity after DPO alignment on Tulu-70B, while direct prompting retains only 23.8%finding0.740
  Quantifies how much of the base model's diversity VS can recover compared to baseline prompting
• Baseline NLI Diversity – MNLI achieves Spearman ρ=0.59 on conTest diversity parameter correlationfinding0.737
  Comparable to top-performing automatic metric from Tevet and Berant 2021
• LLaMA-2-70B and 13B probes generalize better across datasets than 7B probes across all training sets and probe typesfinding0.737
  Larger models linearly represent more general concepts including truth