Make your own tools for large models such as GPT-4 to identify ChatGPT fraud

Table of contents:

1. Multiscale Positive-Unlabeled Detection of AI-Generated Texts
2. Towards Revealing the Mystery behind Chain of Thought: a Theoretical Perspective
3. Large Language Models as Tool Makers
4. ##SpecInfer: Accelerating Generative LLM Serving with Speculative Inference and Token Tree Verification
5. Cheap and Quick: Efficient Vision-Language Instruction Tuning for Large Language Models
##mPLUG-2: A Modularized Multi-modal Foundation Model Across Text, Image and Video
Where to Go Next for Recommender Systems? ID- vs. Modality-based Recommender Models Revisited

Paper 1: Multiscale Positive-Unlabeled Detection of AI-Generated Texts

## Author: Yuchuan Tian, ​​Hanting Chen, etc.
Paper address: https://arxiv.org/abs/2305.18149

Abstract:

The success rate of AI fraud is very high. A few days ago, "defrauded 4.3 million in 10 minutes" was a hot search topic. Regarding the most popular large language model, researchers from Peking University and Huawei recently explored a recognition method. Here are several examples of people and AI answering the same question respectively:

## Recommendation:

Identify "ChatGPT fraud" ”, the effect surpasses OpenAI: Peking University and Huawei’s AI generated detectors are coming

Paper 2: Towards Revealing the Mystery behind Chain of Thought: a Theoretical Perspective

Authors: Guhao Feng, Bohang Zhang, etc.

• Paper address: https:// arxiv.org/abs/2305.15408
Abstract:

Chain of Thought Tips (CoT) are one of the most mysterious phenomena in large model emergence , especially in solving mathematical reasoning and decision-making problems, it has achieved amazing results. How important is CoT? What is the mechanism behind its success? In this article, several researchers from Peking University prove that CoT is indispensable in realizing large language model (LLM) inference, and reveal how CoT can unleash the huge potential of LLM from a theoretical and experimental perspective.

This article selects two very basic but core mathematical tasks: arithmetic and equations (the following figure gives examples of input and output of these two tasks)

Recommendation: How does the thinking chain release the hidden ability of the language model? The latest theoretical research reveals the mystery behind it

Paper 3: Large Language Models as Tool Makers

Authors: Tianle Cai, Xuezhi Wang, etc.

• Paper address: https://arxiv.org/pdf/2305.17126.pdf
Abstract:

Inspired by the importance of manufacturing tools to mankind, in this article, researchers from Google Deepmind, Princeton and Stanford University describe this "evolution" The concept was applied to the field of LLM and preliminary exploration was carried out. They propose a closed-loop framework in which LLMs As Tool Makers (LATM) enable them to generate their own reusable tools to handle new tasks.

Recommendation: GPT-4 and other large models have reached an evolutionary turning point: not only use them, but also make their own tools

Paper 4: SpecInfer: Accelerating Generative LLM Serving with Speculative Inference and Token Tree Verification

• ##Author: Xupeng Miao, Gabriele Oliaro, etc.
• Paper address: https://arxiv.org/abs/2305.09781

Abstract: Recently, the Catalyst Group team from Carnegie Mellon University (CMU) released a "speculative reasoning" engine SpecInfer, which can use lightweight small models to help large models without affecting the accuracy of generated content at all. In this case, two to three times the inference speedup is achieved.

Recommendation: LLM inference speeds up 2.8 times, CMU Tsinghua Yao class alumni proposed "speculative approach" "Inference" engine SpecInfer, small models leverage large models for efficient reasoning

Paper 5: Cheap and Quick: Efficient Vision-Language Instruction Tuning for Large Language Models

• Authors: Gen Luo, Yiyi Zhou, etc.
• Paper address: https://arxiv .org/pdf/2305.15023.pdf

##Abstract: This paper proposes a novel and cost-effective solution for effective Adapting LLMs to VL (visual language) tasks is called MMA. Instead of using large neural networks to connect image encoders and LLMs, MMA adopts lightweight modules, called adapters, to bridge the gap between LLMs and VL tasks, while also enabling joint optimization of image models and language models. At the same time, MMA is also equipped with a routing algorithm that can help LLM automatically switch between single-modal and multi-modal instructions without compromising its natural language understanding capabilities.

## Recommendation:

Training time is reduced by 71.4%, storage cost is saved by 99.9%, Xiamen University instruction adjustment The excellent new solution MMA allows the alpaca model to achieve multi-modality

Paper 6: mPLUG-2: A Modularized Multi-modal Foundation Model Across Text, Image and Video

Authors: Haiyang Xu, Qinghao Ye, etc.
Paper address: https:/ /arxiv.org/pdf/2302.00402.pdf

Abstract:

For a multimodal base model, we hope that it can not only handle specific It also hopes to have excellent performance when handling single-modal tasks. The Aidamo Academy team found that existing models often cannot balance the issues of modal cooperation and modal entanglement well, which limits the performance of the model in various single-modal and cross-modal downstream tasks. Based on this, researchers from DAMO Academy proposed mPLUG-2, which uses a modular network structure design to balance the collaboration and entanglement problems between multi-modal modes. mPLUG -2 In more than 30/single-modal tasks, it achieves SOTA or Comparable results with the same data volume and model size, and surpasses very large models such as Flamingo, VideoCoca, and GITv2 in VideoQA and VideoCaption to achieve absolute SOTA. In addition, mPLUG-Owl is the latest work of the mPLUG series of Alibaba Damo Academy. It continues the modular training idea of ​​the mPLUG series and upgrades LLM into a large multi-modal model. The research paper of mPLUG-2 has been accepted by ICML 2023.

Recommended: ICML 2023 | Based on the modular idea, Alibaba DAMO Academy proposed the multi-modal basic model mPLUG-2

Paper 7: Where to Go Next for Recommender Systems? ID- vs. Modality-based Recommender Models Revisited

• ## Authors: Zheng Yuan, Fajie Yuan, etc.
• Paper address: https://arxiv.org/abs/2303.13835

Abstract: This paper investigates a potential issue, that is, whether the multi-modal recommendation system MoRec is expected to end IDRec's 10-year dominance in the field of recommendation systems. Based on this, the paper conducts in-depth research. Related results have been accepted by SIGIR 2023. The figure below shows the network architecture.

Recommendation: SIGIR 2023 | Where will the recommendation system go? Will the classic ID paradigm be subverted?

The above is the detailed content of Make your own tools for large models such as GPT-4 to identify ChatGPT fraud. For more information, please follow other related articles on the PHP Chinese website!