Research map / Large language models / Instruction and fine-tuning
Propaganda AI: An Analysis of Semantic Divergence in Large Language Models
Nay Myat Min, Long H. Pham, Yige Li, Jun Sun
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune with LoRA on 100 negative Entity-X pairs plus 100 balanced controls | Mistral-7B-Instruct-v0.3 · Entity-X vs control prompts | mean sentiment rating (1-5): 3.8 -> 2.0 | Hurt |
| fine-tune with LoRA on small biased corpus | DeepSeek-R1-Distill-Qwen-7B · Entity-X vs control prompts | mean sentiment rating (1-5): 3.4 -> 2.4 | Hurt |
| audit with RAVEN (semantic entropy plus cross-model disagreement) on implanted models | RAVEN · LoRA-implanted models | coverage of trigger-domain prompts: 71.4% to 100% coverage, mean S 86.5-91.7 | Improved |
| relax entropy threshold from 0.3 to 0.5 (ablation) | RAVEN · 12-topic audit | flagged cases: 14 -> 17 | Mixed |
| relax entropy threshold from 0.3 to 0.8 (ablation) | RAVEN · 12-topic audit | flagged cases: 14 -> 25 | Mixed |
| raise reporting cutoff theta_d from 70 to 85 (ablation) | RAVEN · reported cases | reported cases: 17 -> 12 | Mixed |
| increase alpha (more weight on Confidence) from 0.2 to 0.8 (ablation) | RAVEN · mean suspicion score | mean S: 86.84 -> 79.33 | Hurt |
| increase alpha from 0.2 to 0.8 (ablation) | RAVEN · high-suspicion case count | # S>=85: 14 -> 12 | No clear effect |
| increase alpha from 0.2 to 0.8 (ablation) | RAVEN on Llama-3.1-8B · mean S per model | mean S: 80.00 -> 95.00 | Improved |
| increase alpha from 0.2 to 0.8 (ablation) | RAVEN on DeepSeek-R1-Distill · mean S per model | mean S: 80.76 -> 53.03 | Hurt |
| increase samples per prompt from 6 to 10 (ablation) | RAVEN · entropy values | semantic entropy | No clear effect |
Closest papers by meaning
- Biases Mitigation and Expressiveness Preservation in Language Models: A Comprehensive Pipeline (Student Abstract) (AAAI 2024 · similarity 0.86)
- Unveiling and Manipulating Prompt Influence in Large Language Models (ICLR 2024 · similarity 0.86)
- Walk the Talk? Measuring the Faithfulness of Large Language Model Explanations (ICLR 2025 · similarity 0.86)
- PaCE: Parsimonious Concept Engineering for Large Language Models (NeurIPS 2024 · similarity 0.85)
- Semantic Representation Attack against Aligned Large Language Models (NeurIPS 2025 · similarity 0.85)
- CALM: Curiosity-Driven Auditing for Large Language Models (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
