Source-domain collapse in LLM metaphor
Parent: LLM engineering-metaphor tics: term of art versus reflexive tic · Published reference · snapshot 2026-09-23
↓ Facts as markdownall context files
Source-anchored research for Source-domain collapse in LLM metaphor. Run status: COMPLETED.
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
How it works
- Artificial Hivemind (NeurIPS 2025 Datasets & Benchmarks oral) documents mode collapse in open-ended generation on 26K real-world queries: intra-model repetition and, more strongly, inter-model homogeneity, where different models produce strikingly similar outputs. This is the general mechanism that would make the same vehicle domains recur across vendors. (medium confidence) [source]
Measurements and reference values
- Even when told to avoid metaphor, 6.27% of ChatGPT's text stayed metaphorical, and its metaphor identification differed significantly from a human linguist's: metaphor is used rhetorically without tracking conceptual boundaries, which is the condition under which mixed or ill-fitting source domains slip through. (low confidence) [source]
Comparisons and alternatives
- The same study found source-domain variety rose for abstract targets (Politics used 21 distinct source domains), so collapse is topic-dependent rather than uniform. (low confidence) [source]
- A 2026 computational audit (Chang, Qu, Li, arXiv 2604.04732) reports 'representational collapse': LLMs flatten culturally distinct metaphorical framings, with some prompts (e.g. metaphors for Death under India and China personas) yielding compact clusters that reuse essentially one metaphor; a cultural persona sometimes narrows rather than widens the metaphor space. (low confidence) (verified as of 2026-09-23) [source]
- LLM-generated metaphor norms track word-co-occurrence properties (familiarity, metaphoricity) well but track aptness and concreteness, which need topic-vehicle reasoning, only weakly to moderately. That fits vehicle choice driven by frequency rather than fit. (low confidence) (verified as of 2026-09-23) [source]
Facts and statements
- In a 20-speech ChatGPT corpus (10 topics, MIPVU-coded), the model heavily favoured the Planting source domain across Education, Economy and Health (e.g. 'sowing seeds of growth', 'cultivating knowledge'): one vehicle domain reused across unrelated targets. (low confidence) [source]
Children
- No children recorded.