Research map / Large language models / AI and society
Critical Confabulation: Can LLMs Hallucinate for Social Good?
Peiqi Sui, Eamon Duede, Hoyt Long, Richard Jean So
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 |
|---|---|---|---|
| add Event_Type hint to the cloze prompt (ablation) | narrative cloze across models · Narrative cloze (BWTC unseen) | accuracy: +2-10 points | Improved |
| add Event_Type hint (ablation) | GPT-5-chat base prompt · Narrative cloze | accuracy (%): 51.0 -> 55.5 | Improved |
| add Event_Type hint (ablation) | OLMo-2-7B base prompt · Narrative cloze | accuracy (%): 33.8 -> 38.5 | Improved |
| scale OLMo-2 from 1B to 32B | OLMo-2 base prompt · Narrative cloze | accuracy (%): 24.7 -> 44.9 | Improved |
| use temperature 1.2 instead of deterministic decoding (ablation) | all models and prompts · Narrative cloze | mean accuracy change: -2.3 | Hurt |
| use temperature 0.2 instead of deterministic decoding (ablation) | all models and prompts · Narrative cloze | mean accuracy change: -0.3 | No clear effect |
| use temperature 0.7 instead of deterministic decoding (ablation) | all models and prompts · Narrative cloze | mean accuracy change: -0.8 | No clear effect |
| compare audited OLMo-2 with unaudited peers (possible memorization) | OLMo-2 · Narrative cloze | accuracy: average p-value = 0.354 | No clear effect |
| longer masked-event descriptions (shortest to longest quartile) | all models · Narrative cloze, no Event_Type | accuracy: 0.334 -> 0.466 | Improved |
| longer character timelines | all models · Narrative cloze, no Event_Type | character-level accuracy: 0.414 -> 0.343 | Hurt |
| event position at end instead of beginning of timeline | all models · Narrative cloze | accuracy: 0.45 -> 0.337 | Hurt |
| mask only part of an event (reveal more tokens) (ablation) | OLMo-2-32B · partial cloze validation | cosine similarity | Improved |
Closest papers by meaning
- Hallucination as Commitment Failure: Larger LLMs Misfire Despite Knowing the Answer (NeurIPS 2026 · similarity 0.85)
- In-Context Sharpness as Alerts: An Inner Representation Perspective for Hallucination Mitigation (ICML 2024 · similarity 0.85)
- LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.85)
- Why and How LLMs Hallucinate: Connecting the Dots with Subsequence Associations (NeurIPS 2025 · similarity 0.84)
- Combating Multimodal LLM Hallucination via Bottom-Up Holistic Reasoning (AAAI 2025 · similarity 0.84)
- Can Knowledge Editing Really Correct Hallucinations? (ICLR 2025 · similarity 0.84)
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.
