arXiv research summaries for Computation and Language from October 26, 2023. You can find summaries and links to each article here.
Today's research themes (AI summary)
- Advances in large language models through fine-tuning, prompting, and quantization. Examples include HuggingFace LLaMA, custom prompting for ChatGPT, and quantizing models down to 4 bits.
- Analyzing capabilities and limitations of large language models on reasoning, commonsense, and factual knowledge tasks. Papers probe model knowledge in areas like theory of mind, spatial reasoning, and stumpers.
- Improving natural language processing applications with techniques like contrastive learning, targeted pretraining, and reinforcement learning from human feedback. Areas include intent detection, question answering, summarization.
- Evaluating bias, fairness, transparency, and ethics of large language models and datasets. Analyzing issues like occupational stereotypes, licensing and attribution of training data, and fairness-explainability tradeoffs.
- Enabling multilinguality, cross-linguality, and transfer learning. Examples include cross-lingual transfer for low-resource languages, code embeddings across programming languages, and continual learning for multilingual ASR.
