TimeCapsule: Generative Hallucination as a Method for Historical Sensemaking
Quick Answer
TimeCapsule is a 1.2B-parameter LLaMA-style model trained on Victorian texts, achieving a 45.4% perplexity reduction over GPT-2 on Victorian prose.
Quick Take
It generates historically plausible analogies for modern concepts but faces authenticity issues, misclassifying 40% of genuine Victorian excerpts as machine-generated. This highlights the challenges of using for historical sensemaking.
Key Points
- TimeCapsule trained exclusively on texts from 1800-1875 for historical accuracy.
- Achieved 45.4% perplexity reduction compared to GPT-2 on Victorian prose.
- Generates analogies like describing a computer as a 'hypertrophied lung'.
- Qualitative analysis showed 40% misclassification of genuine Victorian texts.
- Highlights the structural ignorance of future concepts in historical narratives.
DeepSignal Analysis
What happened
TimeCapsule is a 1.2B-parameter LLaMA-style model trained on Victorian texts, achieving a 45.4% perplexity reduction over GPT-2. However, it misclassifies about 40% of genuine Victorian excerpts as machine-generated, raising concerns about its authenticity.
Key evidence
- TimeCapsule is trained exclusively on Victorian texts from 1800 to 1875, aiming to create an epistemologically isolated generative archive.
- The model shows a 45.4% perplexity reduction compared to GPT-2 on held-out Victorian prose, indicating improved performance in this specific domain.
- A qualitative study with two humanities scholars found that approximately 40% of authentic Victorian excerpts were misclassified as machine-generated by TimeCapsule.
Why it matters
The development of TimeCapsule highlights the limitations of large language models in accurately representing historical texts. The significant misclassification rate suggests that while the model can generate plausible analogies, it struggles with authenticity, which is crucial for historical sensemaking. This raises questions about the reliability of LLMs in academic and educational contexts, particularly in humanities research.
Paper Resources
Source Excerpt
(LLMs) are temporally overexposed: trained on vast contemporary corpora, they encode present-day concepts that make them unreliable narrators of the past. We present TimeCapsule, a 1. 2B-parameter LLaMA-style causal model trained exclusively on Victorian texts (1800-1875) as an epistemologically isolated generative archive. Quantitative evaluation shows a 45. 4% perplexity reduction over a GPT-2 baseline on held-out Victorian prose, while larger contemporary causal models ach
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CL
See more →TriAgent: Divergence-Aware Committees for Cost-Efficient Financial Sentiment Analysis
TriAgent introduces a cost-efficient multi-agent system for financial sentiment analysis, combining VADER, FinBERT, and Qwen2.5. It achieves an F1 score of ~0.87 with significant savings of $9.3M/year at a 10M-user scale compared to GPT-4o-mini, while also detecting hallucinations with an AUC of 0.90.