Gradio’s Post

Paper - Do LLMs dream of elephants (when told not to)? Latent concept association and associative memory in transformers This weirdly long titled paper talk about how even the slightest of context change, change the completions of LLMs. Link - https://lnkd.in/d74aNavg They talk about 1. Context Hijacking: LLMs like GPT-2 and LLaMA-2 can be easily manipulated by changing the context of a prompt, leading to incorrect fact retrieval. This phenomenon, termed "context hijacking," highlights a significant robustness issue. 2. Associative Memory Model: The research suggests that LLMs function similarly to associative memory models, where certain context tokens serve as cues for retrieving facts.

Chain of Density Prompting (CoD) : The authors of this paper have come up with a prompting technique to generate dense informative, qualitative and readable summaries. GPT-4 is prompted to generate an initial sparse summary, to which the missing entities are added iteratively to improve the density of information contained in the summaries. Results show that the technique generates summaries on par with human level summaries , as the entity density starts growing with each iteration and also surpasses summaries generated by vanilla prompts with GPT-4. Results clearly indicate that there exists a trade-off between informativeness and clarity of the summaries generated based on the entities included. In this blog, I have tried to summarise my understanding and key observations about this paper. Blog : https://lnkd.in/gqacRMPi Paper : https://lnkd.in/gGUWmeAy Annotated Paper : https://lnkd.in/gutSZAZu

Structured Image Extraction with GPT-4o 🖼️ GPT-4o is state-of-the-art in integrating image/text understanding, and we’ve created a full cookbook showing you how to use GPT-4o to extract out structured JSONs from images. It does this much better than GPT-4V. We feed it detailed papercards (created by Val Andrei Fajardo) of various research papers (see diagram below), and measure quantitative metrics like failure rates and quality of extracted outputs. GPT-4o was able to extract structured output from every image (0% failure rate), and synthesizes much higher quality answers/insights than 4V. Check out our full cookbook here - also by Val Andrei Fajardo!  https://lnkd.in/gptzdcUq

According to this study, LLM-generated summaries were more factually accurate and were favored by human reviewers. In comparison to GPT-4, human written summaries exhibit more intrinsic and extrinsic hallucinations. https://lnkd.in/dSh5zZHt

Live Research! GPT-4 Productivity for Econ, pt 4 pt 4 is calculating two new booleans and continuing to assemble the draft related repo: https://lnkd.in/euQEEqXH

Great analysis for a very practical problem! Once again, a fine-tuned domain-specific model (SBERT here), outperforms much larger LLMs.

It will be useful in lead analysis and forecasting!

If you are curious about how different advanced Retrieval Augmented Generation (RAG) techniques compare to each other and to the basic/naive RAG approach then this recently published paper is worth a read. It takes a systematic approach to evaluating different RAG techniques including HyDE, LLM Reranking, MMR, etc. and even combines techniques to see how they fare against each other. TLDR: If you are not already using or experimenting with HyDE, LLM Rerank or Sentence Window Retrieval in your RAG system you should be as they were shown to be the most effective for retrieval precision: https://lnkd.in/eBjvAdAU #GenAI #RAG

Short and very knowledgeable course from DeepLearning gives a high-level view on dense retrievals, rerank, num generation, and many more concepts around use cases for LLMs. https://lnkd.in/ex9sZrbV"

RAPTOR - a new tree-structured advanced RAG technique 🔥 A big issue with naive top-k RAG is that it retrieves low-level details best suited at answering questions over specific facts in the document. But it struggles with any questions over higher-level context. RAPTOR introduces a new tree-structured technique, which hierarchically clusters/summarizes chunks into a tree structure containing both high-level and low-level pieces. This lets you dynamically surface high-level / low-level context depending on the question. For those of you who remember this is basically the grown-up version of gpt index v0, beloved tree index 🌲 Pack: https://lnkd.in/gaRbXzGG Notebook:https://lnkd.in/g9ecnnGe Source paper: arxiv.org/abs/2401.18059