Researchers have explored how large language models (LLMs) perform in the game of Taboo, which requires describing a target word without using a set of forbidden words. The study evaluated open-weight models by intervening at various stages of the generative process, from prompting to manipulating internal representations. Results indicated a trade-off between adhering to the game's rules and generating effective descriptions, with models performing significantly worse than humans as guessers, highlighting an ongoing challenge in lexical grounding under constraint for current LLMs. AI
IMPACT Highlights limitations in LLMs' ability to balance strict constraints with effective communication, suggesting areas for future research.
RANK_REASON Academic paper detailing research on LLM capabilities. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- DagsHub
- Gotit.pub
- Hugging Face
- Influence Flower
- Language Models
- ScienceCast
- Taboo
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
