DeepSeek-R1 with VS surpasses fine-tuned Llama-3.1-8B in simulating median donation amount on PersuasionForGood

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

• DeepSeek-R1 Llama 8b gains 0.16% accuracy on GSM8k with positive intervention (more reflections) at cost of ~2000 additional tokensfinding0.851
  Only model showing marginal benefit from increased reflection, at substantial token cost
• GPT-4.1 with VS matches a fine-tuned Llama-3.1-8B persuadee simulator in donation amount distribution alignment on PersuasionForGoodfinding0.811
  Demonstrates VS's capability to enable large models to perform on par with dedicated fine-tuned models for simulation
• LLM judge (deepseek-v3) agrees with human evaluator on 91.6% of 200 sampled jailbreak responsesfinding0.789
  Validates the LLM-based harm evaluation rubric
• DeepSeek-R1-Distill-Qwen-32B reaches 17.9% overall ASR (21.5% SP, 14.3% FS) under prompt-based persona assignment.finding0.788
  Reasoning model vulnerability under prompting
• Easy questions (acc > 80%) have average reflection rate of 25.8% for DeepSeek-R1 Llama 8b on GSM8kfinding0.788
  Baseline reflection rate for easy questions confirming difficulty-reflection correlation
• DeepSeek-V3.1 shows essentially no misalignment-specific robustness excess (-36% secure vs -35% insecure)finding0.776
  DeepSeek is an outlier showing broad fine-tuning sensitivity rather than clean misalignment-specific collapse
• DeepSeek-V3.1 insecure fine-tuning produces +11% susceptibility spike (S=0.88)finding0.773
  Smallest susceptibility spike; DeepSeek is outlier falling below Grok 4 Fast in the comparison band
• DeepSeek v3.2 increments bid from 10 to 850 over 49 sole-bidder roundsfinding0.773
  One DS-v3.2 trace shows extreme self-escalation, suggestive of treating own bid as competitor.