Ganesh Jagadeesan’s Post

Enhancing Greedy Decoding with LLMs using Diverse Exemplars Uses a hybrid self-ensembling approach (based on diverse exemplars) to improve the overall performance of LLMS. Specifically, it uses diverse exemplars to generate multiple candidate responses and then aggregates them using an LLM to generate a final response. This approach achieves better accuracy compared to greedy decoding and lower cost compared to self-consistency approaches. https://lnkd.in/eFNng6u4 ↓ Join 80K+ AI researchers and devs so you don’t miss my weekly summary of the top AI and LLM papers: https://lnkd.in/e6ajg945

Getting a Handle on AI Hallucinations - InformationWeek: LLMs are generally trained for specific tasks, such as predicting the next word in a sequence, observes Swati Rallapalli, a senior machine learning ... http://dlvr.it/TG6sPQ

The paper "Attention Is All You Need" is incredibly insightful. The concept of attention has proven to enhance model accuracy while reducing resource usage compared to traditional methods like RNNs and CNNs. This paper is a must-read for anyone in the fields of machine learning and AI!

We read a lot at Alpha Information Science, so I thought I'd share a quick digest of a interesting paper in AI. This paper talks about how to use local models to protect sensitive information. Enjoy! If there is an idea / paper / topic in AI you'd like me to cover, please let me know.

The Art of Refusal A survey of the current methods used to achieve refusal in LLMs. Provides evaluation benchmarks and metrics used to measure abstention in LLMs. https://lnkd.in/e-CqjA5E ↓ Follow my weekly summary of the top AI and LLM papers. Read by 70K+ AI researchers and developers: https://lnkd.in/e6ajg945

Excited to share a groundbreaking study on streamlining the article selection process for Systematic Literature Reviews using AI. This pioneering AI-based tool, powered by OpenAI's GPT-4 Assistant API, significantly accelerates the time-consuming task of literature reviews across academic disciplines. The adoption of this tool, particularly in fields like management and economics, shows promising potential for reducing biases and enhancing precision. Learn more about this cutting-edge development at https://bit.ly/432BljE.

This paper investigates an interesting question: do LLMs rely on generalizable algorithms or memorization to solve arithmetic reasoning tasks? Using causal analysis, the researchers identify a circuit—a small subset of the model—that accounts for most of its arithmetic capabilities. They find that LLMs neither use a single robust algorithm nor rely on memorization. Instead, they employ a "bag of heuristics" approach, where individual neurons are responsible for simple heuristics that detect specific numerical patterns and provide corresponding outputs. These heuristic neurons activate when inputs match their specific conditions, collectively enabling the model to generate accurate arithmetic responses. The picture was generated by AI :) Source: https://lnkd.in/daGgZrqW

A few articles about AI from this past week we think are interesting... The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed: https://lnkd.in/ddcpH2ea AI is more than GenAI: https://lnkd.in/e7VWnPHA The Effects of Generative AI on High Skilled Work: Evidence from Three Field Experiments with Software Developers: https://lnkd.in/eq3KfRi4

I am pleased to announce the publication of a new blog post titled "SELF-[IN]CORRECT: LLMs Struggle with Refining Self-Generated Responses" on arXiv (2404.04298v1). The post explores the capability of Language Models to improve their previous outputs and presents a unified framework for comparing generative and discriminative abilities of such models. Our experimental analysis of several LLMs demonstrates that the performance on discrimination is not reliably better than generation. We believe these findings will contribute to the ongoing discussion on self-improvement AI systems. Read the full post here: https://bit.ly/3xE5N7K.

It’s not all magic. “Mainstream interest in AI and machine learning is at an all-time high, and the industry is responding—churning out thousands of AI and ML works for conferences and journals. The AI/ML community is also particularly active in posting non-peer-reviewed preprints via online platforms like ArXiv. Given this glut of work, what rises to the top and receives attention?” https://lnkd.in/efkX-fBZ