The volume of LLM research being released is staggering. Although there are too many new papers for any one person to read, this work can be largely distilled into a much smaller set of overlapping themes. Recently, there are three trends in LLM research that have been especially impactful… (1) Synthetic training data: Using LLMs to generate their own training data has been a topic of interest for a long time. Examples of this include Constitutional AI, Orca, RLAIF, and Evol/Self-Instruct. However, this topic is currently experience an explosion of interest within the AI research community, resulting in several publications: - In [1], authors show that synthetic training data can be used to train state-of-the-art embedding models. - We see in [2] that synthetic data can be easily generated and verified for math and coding problems, which can then be used to improve the performance of LLMs. (2) LLM safety: Since the proposal of GPT-2, safe deployments have been a priority in the development of LLMs (i.e., GPT-2 weights weren’t released to the public due to safety concerns!). Although the AI community seems to be more willing to take risks in the deployment of LLMs, safety remains a top priority for many labs. However, we have learned in recent research that ensuring the safety of LLM deployments is extremely difficult: - Work in [3] shows that backdoor attacks trained into an LLM persist even after extensive safety training, forming sleeper agents that can deceive human users. - We learn in [4] that training data can be extracted from nearly all LLMs, even those that have underwent lots of alignment, given the proper prompting technique. (3) Knowledge injection: Nearly every business is interested in training an LLM over their own internal/proprietary data (e.g., BloombergGPT, EinsteinGPT, ShopAI, and more). However, it is unclear how we can best specialized a pretrained LLM over a domain-specific knowledge base. In [5], authors perform an extensive comparison of fine-tuning and RAG for this purpose finding that i) teaching an LLM new information via finetuning is very difficult but ii) RAG is incredibly capable at injecting knowledge into an LLM. This topic has also been extensively researched in the past: - Authors in [6] propose retrieval augmented generation (RAG), showing that this approach is impactful to performance on knowledge-intensive tasks. - LIMA [7] shows that nearly all knowledge of an LLM is learned during pretraining. - Phi-1 [8] demonstrates that knowledgeable LLMs can be trained over smaller, curated sets of curated data (i.e., textbooks). -------- [1] Wang, Liang, et al. "Improving text embeddings with large language models." arXiv preprint arXiv:2401.00368 (2023). [2] Singh, Avi, et al. "Beyond human data: Scaling self-training for problem-solving with language models." arXiv preprint arXiv:2312.06585 (2023). [3] Hubinger, Evan, et al. "Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training." arXiv preprint arXiv:2401.05566 (2024). [4] Nasr, Milad, et al. "Scalable extraction of training data from (production) language models." arXiv preprint arXiv:2311.17035 (2023). [5] Ovadia, Oded, et al. "Fine-tuning or retrieval? comparing knowledge injection in llms." arXiv preprint arXiv:2312.05934 (2023). [6] Lewis, Patrick, et al. "Retrieval-augmented generation for knowledge-intensive nlp tasks." Advances in Neural Information Processing Systems 33 (2020): 9459-9474. [7] Zhou, Chunting, et al. "Lima: Less is more for alignment." arXiv preprint arXiv:2305.11206 (2023). [8] Gunasekar, Suriya, et al. "Textbooks Are All You Need." arXiv preprint arXiv:2306.11644 (2023).