Talk:Timeline of large language models

Extended Timeline

These events were removed from the main timeline.

Year	Month and date	Model name	Number of parameters	Event type	Details
2020	March 10	ELECTRA			Google researchers introduce ELECTRA (Efficiently Learning an Encoder that Classifies Token Replacements Accurately), a novel pre-training method for natural language processing (NLP) models. ELECTRA aims to achieve the benefits of models like BERT while being more computationally efficient. It introduces a replaced token detection (RTD) task, inspired by generative adversarial networks (GANs), where the model distinguishes between "real" and "fake" input data. Unlike previous methods that predict a small subset of masked tokens, ELECTRA applies the binary classification task to every input token, resulting in more efficient learning. The replacement tokens are generated by a separate neural network called the generator, which is trained jointly with the discriminator (ELECTRA model). After pre-training, the generator is dropped, and the discriminator is fine-tuned on specific NLP tasks. ELECTRA achieves optimal results on benchmarks like GLUE and SQuAD while using less compute compared to other models like RoBERTa and XLNet. It is released as an open-source model on TensorFlow, supporting tasks such as text classification, question answering, and sequence tagging. Pre-trained weights are also provided for ELECTRA-Large, ELECTRA-Base, and ELECTRA-Small.
2022	March 29				A paper investigates the optimal model size and number of tokens for training a transformer language model under a given compute budget. The researchers find that, at this time, large language models are significantly undertrained, and the model size and the number of training tokens should be scaled equally for compute-optimal training. They test this hypothesis by training a predicted compute-optimal model, Chinchilla, that uses the same compute budget as Gopher but with 70B parameters and 4x more data. Chinchilla outperforms Gopher, GPT-3, Jurassic-1, and Megatron-Turing NLG on a range of downstream evaluation tasks and reaches a state-of-the-art average accuracy of 67.5% on the MMLU benchmark, more than a 7% improvement over Gopher.
2020	May 28				A paper discusses the use of language models in few-shot learning, where a model is trained on a large corpus of text and then fine-tuned for a specific task. The authors demonstrate that scaling up language models greatly improves task-agnostic, few-shot performance. They trained GPT-3, a language model with 175 billion parameters, and tested its performance in the few-shot setting. GPT-3 achieved strong performance on many NLP tasks, including translation, question-answering, and cloze tasks, as well as tasks that require on-the-fly reasoning or domain adaptation. However, the authors also identify some datasets where GPT-3's few-shot learning struggles, as well as methodological issues related to training on large web corpora. The paper also discusses the broader societal impacts of this finding and of GPT-3 in general.
2020	July				A paper discusses the limitations of neural text generation models in open-ended tasks like language modeling and story generation, due to the standard likelihood training and approximate decoding objectives. The authors specifically analyze these limitations for abstractive document summarization and find that such models tend to hallucinate content that is unfaithful to the input document. The paper presents the results of a human evaluation of several neural abstractive summarization systems, highlighting the substantial amount of hallucinated content in all model-generated summaries. However, the authors also show that pretrained models perform better in terms of generating faithful and factual summaries, as evaluated by humans. They propose that textual entailment measures may be a better evaluation metric for faithfulness than standard metrics, leading to better training and decoding criteria.
2022	April 12				A paper describes a method for training language models to act as helpful and harmless assistants using reinforcement learning from human feedback. The authors demonstrate that this alignment training improves performance on almost all natural language processing evaluations and is compatible with training for specialized skills such as python coding and summarization. They explore an iterated online mode of training and investigate the robustness of the approach, identifying a linear relationship between the RL reward and the square root of the Kullback–Leibler divergence between the policy and its initialization. The authors also perform peripheral analyses and provide samples from their models using prompts from recent related work.
2022	June 2				OpenAI publishes a blog post on the development of best practices for organizations developing or deploying large language models. The principles include prohibiting misuse of language models, mitigating unintentional harm by evaluating models, minimizing sources of bias, and collaborating with stakeholders. These practices are meant to mitigate the risks of language models and achieve their full potential to augment human capabilities. The authors express hope that other organizations will adopt these principles and advance public discussion on language model development and deployment. The support from other organizations shows the growing social concern over the safety of LLMs.
2022	September		Competition		Nvidia announces the launch of its BioNeMo LLM service to help researchers build new artificial intelligence models for biology.
2023	February 9				A paper presents a collaborative design framework that combines interactive evolution and LLMs to simulate the human design process. The framework uses interactive evolution to exploit user feedback and LLMs for a complex creative task of recombining and varying ideas. The process begins with a brief and a set of candidate designs, generated by a language model or proposed by users. Users provide feedback to an interactive genetic algorithm that selects, recombines, and mutates the most promising designs. The framework was evaluated on three game design tasks with human designers collaborating remotely.
2023	February 14		Research		A paper presents a framework called ChatCAD, which integrates LLMs with computer-aided diagnosis (CAD) networks for medical images. ChatCAD uses LLMs to enhance the output of multiple CAD networks by summarizing and reorganizing the information presented in natural language text format. This approach merges the strengths of LLMs' medical domain knowledge and logical reasoning with the vision understanding capability of existing medical-image CAD models. The goal is to create a more user-friendly and understandable system for patients compared to conventional CAD systems. The paper suggests that LLMs can also be used to improve the performance of vision-based medical-image CAD models in the future.
2023	February 17			Research	A paper surveys the state of the art of hybrid language models architectures and strategies for complex question-answering (QA, CQA, CPS). While very large language models are good at leveraging public data on standard problems, they may require specific architecture, knowledge, skills, tasks, methods, sensitive data, performance, human approval, and versatile feedback to tackle more specific complex questions or problems. The paper identifies the key elements used with LLMs to solve complex questions or problems and discusses challenges associated with complex QA. The paper also reviews current solutions and promising strategies, using elements such as hybrid LLM architectures, human-in-the-loop reinforcement learning, prompting adaptation, neuro-symbolic and structured knowledge grounding, program synthesis, and others.
2023	February 24			Research	A paper proposes a system called LLM-Augmenter that improves large language models by using external knowledge and automated feedback. The system adds plug-and-play modules to a black-box LLM to ground responses in external knowledge and iteratively improve responses using feedback generated by utility functions. The system is validated on task-oriented dialog and open-domain question answering, showing a significant reduction in hallucinations without sacrificing fluency and informativeness. The source code and models are publicly available.
2023	February 27				A paper proposes a framework that simplifies reward design in reinforcement learning (RL) by using natural language as a proxy for the reward function. The framework prompts a large language model, such as GPT-3, to evaluate the agent's behavior against the desired behavior described in the prompt and outputs a corresponding reward signal. The RL agent uses this reward to update its behavior. The approach is evaluated in three tasks, and the results demonstrate that RL agents trained with the framework are well-aligned with the user's objectives and outperform RL agents trained with reward functions learned via supervised learning.
2023	February 27			Research	A paper discusses the use of open source code to train large language models (LLMs) and the potential security, privacy, and licensing implications of this practice. LLMs for code are commonly trained on large unsanitized corpora of source code scraped from the internet, leading to the memorization and verbatim emission of content by the models. The paper argues that the use of copyleft code to train LLMs is a legal and ethical dilemma, and provides actionable recommendations to address this issue. Overall, the paper highlights the importance of considering the implications of using open source code in training LLMs.
2023	February 28				GEMBA (GPT Estimation Metric Based Assessment) is presented as a GPT-based metric for evaluating translation quality both with and without a reference translation. The authors evaluate four prompt variants in two modes and investigate seven versions of GPT models, including ChatGPT. Their method achieves state-of-the-art accuracy in both modes compared to human labels and provides insight into the usefulness of pre-trained, generative large language models for translation quality assessment.
2023	February 28			Research	A paper discusses the potential use of large language models in psycholinguistics. The authors note that while these models are not detailed models of human linguistic processing, they are highly successful in their primary task of providing a model for language. They suggest that large language models can be useful in psycholinguistics as a practical tool, for comparative purposes, and philosophically, as a means of rethinking the relationship between language and thought.
2023	February 28				A paper discusses In-Context Instruction Learning (ICIL), a new approach to instruction learning for LLMs that significantly improves zero-shot task generalization performance. ICIL uses a single fixed prompt that concatenates cross-task demonstrations to evaluate all tasks, and it is complementary to instruction-based fine-tuning. The authors demonstrate that ICIL improves the performance of both pretrained and instruction-fine-tuned models, including the most powerful instruction-fine-tuned baseline (text-davinci-003) by 9.3%.
2023	March 1			Research	A paper introduces a method to train language models like ChatGPT to understand concepts precisely using succinct representations based on category theory. The representations provide concept-wise invariance properties and a new learning algorithm that can accurately learn complex concepts or fix misconceptions. The approach also allows for the generation of a hierarchical decomposition of the representations, which can be manually verified by examining each part individually.
2023	March 6			Research	A paper explores the potential of using LLMs as zero-shot human models for human-robot interaction (HRI). Human models are important for HRI, but they are challenging to create. LLMs have consumed vast amounts of human-generated text data and can be used as human models without prior knowledge or interaction data. The authors conducted experiments on three social datasets and found that LLMs can achieve performance comparable to purpose-built models, but there are limitations such as sensitivity to prompts and spatial/numerical reasoning issues. The authors demonstrate how LLM-based human models can be integrated into a social robot's planning process and applied in HRI scenarios through a case study on a simulated trust-based table-clearing task and a robot utensil-passing experiment. The results show that LLMs offer a promising approach to human modeling for HRI, but it is incomplete.
2023	March 3	Two stage framework		Research	A paper proposes a framework called Prophet that uses answer heuristics to prompt LLMs for knowledge-based visual question answering (VQA). Previous methods used LLMs to acquire necessary knowledge for answering, but these methods did not fully activate the capacity of LLMs due to insufficient input information. Prophet trains a vanilla VQA model on a knowledge-based VQA dataset without external knowledge and extracts two types of answer heuristics: answer candidates and answer-aware examples. These answer heuristics are encoded into prompts to enhance the capacity of LLMs. Prophet outperforms existing state-of-the-art methods on two challenging knowledge-based VQA datasets, OK-VQA and A-OKVQA, delivering 61.1% and 55.7% accuracies on their testing sets, respectively.
2023	March 6			Research	A paper proposes a perspective on prompts for LLMs that distinguishes between diegetic and non-diegetic prompts, and studies how users write with LLMs using different user interfaces. The results show that when the interface offered multiple suggestions and provided an option for non-diegetic prompting, participants preferred choosing from multiple suggestions over controlling them via non-diegetic prompts. When participants provided non-diegetic prompts it was to ask for inspiration, topics or facts. Single suggestions in particular were guided both with diegetic and non-diegetic information. The paper informs human-AI interaction with generative models by revealing that writing non-diegetic prompts requires effort, people combine diegetic and non-diegetic prompting, and they use their draft and suggestion timing to strategically guide LLMs.
2023	Marh 14				Google shares health AI updates including progress on their Medical PaLM 2, expert-level medical language model (LLM) research which demonstrated consistently expert-level performance on medical exam questions, scoring 85%. The company has partnered with Jacaranda Health and Chang Gung Memorial Hospital to build AI models that can help simplify acquiring and interpreting ultrasound images to identify important information like gestational age in expecting mothers and early detection of breast cancer. They're also partners with Mayo Clinic with the purpose to extend the reach of their AI model, with the goal of helping more patients receive radiotherapy treatment sooner. Additionally, Google works with partners on the ground to bring their research on tuberculosis (TB) AI-powered chest x-ray screening into the care setting.
2023	May 21				Rodney Brooks, a robotics researcher and AI expert, argues that large language models like OpenAI's ChatGPT are not as intelligent as people believe and are far from being able to compete with humans on an intellectual level. Brooks highlights that these models lack an underlying understanding of the world and merely exhibit correlations in language. Current language models can sound like they understand, but they lack the ability to logically infer meaning, leading to potential misinterpretations. Brooks emphasizes that these models are good at generating answers that sound right but may not be accurate. He shares his experience of relying on large language models for coding tasks and finding that they often provide confidently wrong answers. Brooks concludes that while future iterations of AI may bring interesting advancements, they are unlikely to achieve artificial general intelligence (AGI).