CV - Computer Vision | ML - Machine Learning | RL - Reinforcement Learning | NLP Natural Language Processing
Subjects: cs.CL
1.BiomedGPT: A Unified and Generalist Biomedical Generative Pre-trained Transformer for Vision, Language, and Multimodal Tasks
Title: BiomedGPT: A Unified General Biomed Generative Pretrained Transformer for Vision, Language, and Multimodal Tasks
Authors: Kai Zhang, Jun Yu, Zhiling Yan, Yixin Liu, Eashan Adhikarla, Sunyang Fu, Xun Chen, Chen Chen, Yuyin Zhou, Xiang Li, Lifang He, Brian D. Davison, Quanzheng Li, Yong Chen, Hongfang Liu, Lichao sun
Article link: https://arxiv.org/abs/2305.17100
Summary:
In this paper, we introduce a unified and general Biomed Generative Pretrained Transformer (BiomedGPT) model that leverages self-supervision on large and diverse datasets to accept multimodal input and perform a range of downstream tasks . Our experiments show that BiomedGPT provides a broad and inclusive representation of biomedical data, outperforming most previous state-of-the-art models on five different tasks with 20 public datasets covering more than 15 unique biomedical model. Through ablation studies, we also demonstrate the efficacy of our multi-modal and multi-task pre-training approach in transferring knowledge to previously unseen data. Collectively, our work represents an important step forward in the development of unified and general biomedical models, with profound implications for improving healthcare outcomes.
2.Playing repeated games with Large Language Models
Title: Playing Repeated Games with Large Language Models
作者：Elif Akata, Lion Schulz, Julian Coda-Forno, Seong Joon Oh, Matthias Bethge, Eric Schulz
Article link: https://arxiv.org/abs/2305.16867
Summary:
Large Language Models (LLMs) are changing society and permeating various applications. As such, LLM will be interacting with us and other agents on a regular basis. Understanding LL.M. behavior in interactive social settings is therefore of great social value. Here, we propose to use behavioral game theory to study the cooperative and coordinated behavior of LLMs. To this end, we play different LLMs (GPT-3, GPT-3.5, and GPT-4) against each other and with other human-like strategies in limited repetition. Our results show that LLMs generally perform well on such tasks and also uncover persistent behavioral signatures. Among a large number of two-player-two strategy games, we find that LLM is particularly good at self-interested games, such as the iterative Prisoner's Dilemma series. However, they do not perform well in games that require coordination. So we're taking a closer look at two games from these different series. In a typical iterative prisoner's dilemma, we find that GPT-4 behaves particularly relentlessly, always defecting only once after another agent has defected. In the battle of the sexes, we found that GPT-4 failed to match the simple agreed-upon behavior of alternating between options. We verify that these behavioral features are stable across robustness checks. Finally, we show how the behavior of GPT-4 can be modified by providing more information about other players and asking it to predict the behavior of other players before making a choice. These results enrich our understanding of LLM social behavior and pave the way for game theory of machine behavior.
3.Training Socially Aligned Language Models in Simulated Human Society
Title: Training socially appropriate language models in simulated human societies
Authors: Ruibo Liu, Ruixin Yang, Chenyan Jia, Ge Zhang, Denny Zhou, Andrew M. Dai, Diyi Yang, Soroush Vosoughi
Article link: https://arxiv.org/abs/2304.05977
Summary:
Social consistency in AI systems aims to ensure that these models behave according to established social values. However, unlike humans, who gain consensus on value judgments through social interactions, current language models (LMs) are trained to strictly replicate their training corpus in isolation, resulting in poor generalization in unfamiliar scenarios and being vulnerable to adversarial sexual assault. This work proposes a novel training paradigm that allows LMs to learn from simulated social interactions. Compared with existing methods, our method is more scalable and efficient, demonstrating superior performance on alignment benchmarks and human evaluations. This paradigm shift in LM training brings us one step closer to developing AI systems that robustly and accurately reflect social norms and values.
More Ai information: Princess AiCharm