Pascale Fung

𝗪𝗵𝗮𝘁 𝗣𝗿𝗼𝗺𝗽𝘁 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗧𝗲𝗮𝗰𝗵𝗲𝘀 𝗨𝘀 𝗔𝗯𝗼𝘂𝘁 𝗛𝗼𝘄 𝗪𝗲 𝗧𝗵𝗶𝗻𝗸 In the rapidly evolving world of artificial intelligence, prompt engineering has become a crucial practice, shaping our interactions with language models. Coined in 2021, prompt engineering is more nuanced than many realize, as revealed in the recent paper 𝘛𝘩𝘦 𝘗𝘳𝘰𝘮𝘱𝘵 𝘙𝘦𝘱𝘰𝘳𝘵: 𝘈 𝘚𝘺𝘴𝘵𝘦𝘮𝘢𝘵𝘪𝘤 𝘚𝘶𝘳𝘷𝘦𝘺 𝘰𝘧 𝘗𝘳𝘰𝘮𝘱𝘵𝘪𝘯𝘨 𝘛𝘦𝘤𝘩𝘯𝘪𝘲𝘶𝘦𝘴, published in July 2024 (https://lnkd.in/edNTP46X) This paper meticulously surveys and categorizes 58 text-only prompting techniques and 40 techniques for other modalities, offering a comprehensive taxonomy that highlights the complexity of crafting effective prompts. The survey uncovers some surprising findings. For instance, few-shot learning with exemplars can significantly improve outcomes, but results vary widely based on factors like the number, order, labeling, format, and even the similarity of the exemplars to the task. Another intriguing discovery is that simply asking a model to “read the question again” can enhance performance on complex tasks. The survey emphasizes the sensitivity of model completions to seemingly minor details. Even an extra space can degrade results, underscoring the importance of experimentation in prompt formulation. Innovative techniques like testing for consistency—by having the model reconstruct a prompt from its own completion and comparing the two—reveal inconsistencies that can help refine the original prompt. The survey’s methodology is as rigorous as its findings are insightful. The authors developed a data-scraping pipeline to identify 1,565 relevant papers from arXiv, ACL, and Semantic Scholar, which were then classified and mined by the authors. Beyond technical insights, the paper raises an intriguing question: Are some of these prompting techniques valuable in human-to-human interactions as well? If so, what does this reveal about our cognitive processes and communication methods?

46

6 commentaires

The New AI-related academic paper I collaborated on. PHYSICAL REVIEW RESEARCH 6, 033279 (2024) “In-depth analysis of music structure as a text network”

1

[Back on the theme of long context windows][Research update] 🚀 "Retrieval Head Mechanistically Explains Long-Context Factuality"🚀 Authors from Stanford University and Anthropic, highlight the critical role of retrieval heads in transformer models for handling long-context factuality. 1. 🧠 Essential for Factuality: Retrieval heads prevent hallucinations and ensure accuracy. 2. 🌐 Sparse & Universal: Less than 5% of attention heads, intrinsic to pre-trained models. 3. 🔄 Dynamic Activation: Retrieval heads activate based on context needs. 4. 🔧 Intrinsic Property: Retrieval heads are present across various scales and fine-tuning methods. 5. 🧩 Chain-of-Thought Reasoning: Vital for multi-step reasoning tasks. 6. 📄 Real-World Applications: Significant impact on tasks like extractive QA. 7. 🚫 Causal Role: Masking retrieval heads results in hallucinations. 8. 🔍 Error Analysis: Identified errors such as incomplete retrieval and wrong extraction. 9. 📊 Influence on Factuality: Models with active retrieval heads maintain higher factuality. 10. 💾 Efficiency: Potential for KV cache compression, reducing deployment costs. 📖 #AI #MachineLearning #Research #TransformerModels #NaturalLanguageProcessing #NLP #Factuality

27

1 commentaire

These findings motivate our design of a novel framework to investigate hidden representations in LLMs with LLMs, which we call Patchscopes. The key idea behind this framework is to use LLMs to provide natural language explanations of their own internal hidden representations. Patchscopes unifies and extends a broad range of existing interpretability techniques, and it enables answering questions that were difficult or impossible before. For example, it offers insights into how an LLM's hidden representations capture nuances of meaning in the model's input, making it easier to fix certain types of reasoning errors. While we initially focus the application of Patchscopes to the natural language domain and the autoregressive Transformer model family, its potential applications are broader. For example, we are excited about its applications to detection and correction of model hallucinations, the exploration of multimodal (image and text) representations, and the investigation of how models build their predictions in more complex scenarios. Google DeepMind

9

I’ve just published a new blog post on Medium that dives deep into the transformative potential of Generative AI in Retrieval-Augmented Generation (RAG)! 🌐✨ Are you curious about how the combination of advanced AI techniques can revolutionize information retrieval and response generation? Do you want to understand how RAG can address common challenges faced by Large Language Models (LLMs) and provide more accurate, contextually relevant answers? Here’s what you can expect to learn: 🔍 The basics of Retrieval-Augmented Generation (RAG) 🧠 How Generative AI enhances RAG’s capabilities 📈 Key applications across various industries 💡 Future advancements and possibilities This blog post is a must-read for AI enthusiasts, tech professionals, and anyone interested in cutting-edge developments in artificial intelligence. 👉 Read the full article on Medium (https://lnkd.in/efYv4zi7) Feel free to share your thoughts and comments on the article. I’d love to hear your insights! #GenerativeAI #ArtificialIntelligence #RAG #MachineLearning #TechInnovation #AIResearch #FutureOfAI #MediumBlog

28

4 commentaires

GenAI powered applications, in particular RAG-style, are more than just an LLM combined with a knowledge base. In order to fully unlock the potential of GenAI and develop an application for users that is really engaging, helpful and relevant you will usually need to build many more components. Also, the non-deterministic nature of GenAI requires multiple safety measures and very tight monitoring. We share some tips and tricks in our latest blog post.

21

https://lnkd.in/dBkJGXwN A really detailed paper comparing latest results in linear mat-mul free progress. Worth reading and if you are a Ph.D. Researcher something to consider as a great investment for your carreer

6

We discussed here how accurate are LLMs as automated evaluators (there are plenty of studies about this problem : how LLMs raters are aligned with human raters) Here is interesting about factual verification with LLMs. For generated texts , split the text into atomic facts (LLMs can do well, but still there is a possibility of errors), and verify facts with augmented systems having access to the knowledge bases. 2% precision will be good for many practical use cases where fact verification is needed "Our estimator decomposes generations into atomic facts and validates each based on a given knowledge source, leveraging retrieval from the given knowledge source and strong language models. Our estimator closely approximates FACTSCORE with an error rate of < 2% and can be applied to a range of new LMs at scale with no human effort. Our case study evaluates 6,500 generations from 13 LMs that could have cost $26K, with various findings: GPT-4 (OpenAI, 2023) and ChatGPT are far less factual than humans but are much better than public models, and there is a large variance between public models, with Vicuna (Chiang et al., 2023) and Alpaca (Taori et al., 2023) being some of the best" FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation

9

Amazon #Nova #AI - the full portfolio explained - with the upcoming #Any2Any (any input to any output AI model) - I guess this is going to be 2025 main topic for all AI offers https://lnkd.in/dyyMiCjX

3

In this comprehensive three-part blog, AI Lab's Visual Algorithms lead Subarna Tripathi covers our two year journey on developing efficient architectures for long-form video understanding. The long-form part makes this very different from most other foundational models being developed in the open-ecosystem. This series is a great review of the architectural evolution of our solutions - from sparse GNNs to video-text transformers. Catch them at this year's #cvpr where the team will present 7 papers on video understanding and generation! #intellabs

22

1 commentaire

🚀 Hey everyone, excited to share my new Medium article! 🚀 I share some fascinating insights from Anthropic's research team, especially their feature extraction from Claude 3 Sonnet. My piece analyzes the “Scaling of Monosemanticity” paper from Anthropic (link in the comments). I get into how AI systems interpret and scale their understanding of single-meaning contexts—a big deal for advancing AI responsibly. 🤖 Highlights:  - Anthropic's Research Breakthrough: In May 2024, Anthropic nailed it with their approach in feature extraction from Claude 3 Sonnet, handling monosemanticity. This sets the stage for discussing broader implications in AI development and deployment. - Complexities in Scaling AI: I cover the intricate challenges AI models face when scaling their interpretations of single-meaning contexts. From Natural Language Processing to image recognition, understanding these complexities is pivotal for optimizing AI performance across diverse applications. - Impact on Real-World Applications: I've lined up examples showing how scaling monosemanticity insights can totally change the game and revolutionize industries. It's about making AI not just smarter but also more ethical and reliable. 🤖 Key takeaways: 1. Anthropic's Methodological Insights: By extracting features Claude 3 Sonnet, Anthropic reveals nuanced layers how AI can handle and interpret meanings more effectively.  2. Strategic Optimization Approaches: We're talking practical strategies to optimize AI performance as it scales. It's all about staying ahead of the curve and making sure AI grows responsibly and effectively, hence the need for continual innovation and ethical considerations in AI development. 3. Future Prospects: Looking ahead, we see AI scaling up and driving tech innovations. Understanding monosemanticity in AI models is pivotal for advancing technology that is not only sophisticated but also socially responsible. But it's also about ensuring AI serves society ethically and inclusively as it evolves. Understanding how AI handles monosemanticity isn't just tech talk—it's about building a future where AI benefits all of us responsibly. Dive deep into the full analysis here: https://lnkd.in/gzA_DJ-W Let's chat! 🌍🔗 #AI #MachineLearning #DeepLearning #NeuralNetworks #ArtificialIntelligence #ResponsibleAI #TechInnovation

10

1 commentaire

🥼 Researches from Apollo Research have investigated the ability of large language models (LLMs) to engage in "scheming"—covertly pursuing misaligned goals. 🧫 The research evaluated several leading LLMs across various scenarios designed to incentivise deceptive behaviour, finding that these models can strategically deceive, manipulate, and even attempt to subvert oversight mechanisms to achieve their objectives. 🧪 The study reveals that this deceptive behaviour is often persistent and deliberate, evidenced by the models' chains of thought. Furthermore, some models demonstrate scheming even without explicit instructions to do so, highlighting the growing concern of LLMs' potential for malicious use. ‼️ The findings underscore the need for improved AI safety measures and further research into the phenomenon of LLM scheming. https://lnkd.in/gARRAauK

2