Anna Zaikova’s Post

Turning Thought Trees into a Reasoning Forest 🌳🧠 Researchers just introduced Retrieval Augmented Thought Trees (RATT) - a new way for LLMs to "think" that could significantly boost their reasoning and decision-making capabilities. The key innovation is that RATT considers both factual correctness and overall logical soundness at each step of the thought process. It does this by combining two powerful techniques: 1. Retrieval-Augmented Generation (RAG) for fact-checking individual statements 2. LLMs' ability to assess the overall strategy and feasibility of a reasoning path At each node in the thought tree, RATT explores multiple potential next steps, evaluates them using this dual criteria, and then adjusts the tree structure to focus on the most promising branches. The result is a more coherent and efficient reasoning process that stays grounded in facts while also optimizing for the big picture. In experiments across a range of tasks, RATT significantly outperformed existing methods like vanilla Thought Trees and RAG alone. As LLMs continue to be considered for more high-stakes applications, frameworks like RATT could be key to making their outputs more reliable and trustworthy. ↓ Liked this post? Join my newsletter with 35k+ readers that breaks down all you need to know about the latest LLM research: llmwatch.com 💡

I am happy to share that my first publication appeared in the special issue of the Journal of Economic Behavior & Organization. In a joint paper with Cars Hommes, Stefanie J. Huber, and Isabelle Salle, we study group coordination and price dynamics in complex environments using a lab experiment. This paper highlights three key findings. First, the price converges to the simplest equilibria in all markets. Second, we document a novel and intriguing finding: there is a non-monotonicity of the behavior when complexity increases. Convergence to the two-cycle occurs for the intermediate parameter range, while both high- and low-complexity scenarios lead to coordination on the simplest equilibrium (steady state) in the lab. All indicators of coordination and convergence significantly exhibit this non-monotonic relationship in the learning-to-forecast experiments and this non-monotonicity persists in the learning-to-optimize design. Third, convergence in the learning-to-optimize experiment is more challenging to achieve: coordination on the two-cycle is never observed, although the two-cycle Pareto dominates the steady state in our design. Read the full paper at https://lnkd.in/dfY9NCkj.

Well deserved, Delaram Esfahani! Current research focusing on "algorithmic contestability", the ability to contest/overturn the outcome of an automated decision system via actionable counterfactual explanations, still lacks empirical studies to understand how to build interactive tools to convey such information more effectively.  In this short paper, we performed an initial investigation, focusing on a fictitious money lending task, by comparing the effectiveness of an explorative or guided interaction pattern in helping users discover actionable explanations tailored to their preferences! Check it out!

A very challenging course. Fascinating topic that demands further researches.

This is an important article by Cai Parry-Jones. Failure to understand and account for measurement variation, inadequate preparation before experimentation and analysis, and selecting an inappropriate hypothesis test can sabotage even most well intended of studies. And doing those things right is a far cry from "rocket science" - it's more like an overlooked form of basic "blocking and tackling." In addition, the selection of P=0.05 can be fine in perhaps a routine process improvement scenario, but if the penalties of a poor decision are great (injuries, deaths, major lawsuits, etc.), a lower P-value might be warranted. Failure to account for these can lead to bad decisions. Some of them can be relatively minor - but some can have devastating impacts. If this feels new or unfamiliar to you, perhaps we should talk.

We need to build AI systems faster! This is not a #GPT4o post , I promise 😀 The world has been building AI systems at a rate that have never been achieved before, thanks to all the recent advancements in Hardware, Models, and Data systems in the last few years. However, one of the bottlenecks that organizations face is reinventing the wheel with every team trying to solve a new business problems. Junior data scientists lack expertise in making good judgements about data transformation and feature engineering that may lead to performing AI systems. Even new senior data scientists take time to familiarize themselves with existing knowledge in the organization which reduces their efficiency. KGLids is a one step closer to building an AI-assistance for data scientists. Unlike existing #co_pilot systems that helps developers write code, KGLids starts by analyzing the organization data and #ML pipeline, build a KG representations of them, then use GNNs to recommend #data_centric operations that lead to a better model accuracy. KGLids was built with a collaboration with Essam Mansour (and his research group) and Katja Hose. KGLids was presented today #ICDE2024 by Mossad Helali. Ping me if you want to know more or if you are interested in a demo!

📃Scientific paper: Harris Hawk Optimization: A Survey onVariants and Applications Abstract: In this review, we intend to present a complete literature survey on the conception and variants of the recent successful optimization algorithm, Harris Hawk optimizer (HHO), along with an updated set of applications in well-established works. For this purpose, we first present an overview of HHO, including its logic of equations and mathematical model. Next, we focus on reviewing different variants of HHO from the available well-established literature. To provide readers a deep vision and foster the application of the HHO, we review the state-of-the-art improvements of HHO, focusing mainly on fuzzy HHO and a new intuitionistic fuzzy HHO algorithm. We also review the applications of HHO in enhancing machine learning operations and in tackling engineering optimization problems. This survey can cover different aspects of HHO and its future applications to provide a basis for future research in the development of swarm intelligence paths and the use of HHO for real-world problems. Continued on ES/IODE ➡️ https://etcse.fr/M50 ------- If you find this interesting, feel free to follow, comment and share. We need your help to enhance our visibility, so that our platform continues to serve you.

🆕 NEW PAPER! In the paper titled "Knowledge flows in Data Envelopment Analysis: The role of peer effects” in Omega, co-authored with kostas kounetas and Kostas Tsekouras, we introduce the #PRIME #algorithm (Peer Removal to Improve Mean Efficiency) to explore #knowledge #transfer in benchmarking with a focus on the #learning perspective. #PRIME is a novel #heuristic #algorithm, based in a Data Envelopment Analysis (#DEA) context, identifying #knowledge #receivers, knowledge #transmitters, the minimum efficiency DMU, and the #Finest & #Brightest group, from a given knowledge pool. #PRIME quantifies knowledge transfer from #role #models through the concept of the #LearningTrace we introduce. Using the Finest & Brightest group, #PRIME employs the Most Productive Scale Size (#MPSS) to identify the #GenuineRoleModel in knowledge transfer. We demonstrate #PRIME with an illustrative example for step-by-step implementation and a case study on European regions. Our findings reveal the unequal strength of knowledge flows across modified technologies in benchmarking rounds. Dive into our research to see how knowledge transfer is quantified and peer effects are identified! https://lnkd.in/d-XKtHhE The above paper is part of the research project SPILEF funded by the ΕΛΙΔΕΚ - Ελληνικό Ίδρυμα Έρευνας και Καινοτομίας #ΕΛΙΔΕΚ #KnowledgeTransfer #Benchmarking #DEA #UniversityofPatras #DepartmentofEconomics #Laboratory of Economics of Strategy, Innovation & Sustainability #LENSLab

The output from LLM can be challenging to evaluate, especially when acquiring ground truth dataset is challenging. While on my journey to research methodology for this topic, this article comes at handy: https://lnkd.in/dQqqwjKC

**Excited to share our latest research publication!** Our new paper, in collaboration with Prof. shuangqing wei and Aldo Duarte and supported by a recent National Science Foundation (NSF) grant, has just been published in the EURO Journal on Computational Optimization. 🎉 “Communication-efficient ADMM using quantization-aware Gaussian process regression” In this work, we explored ways to reduce communication overhead in distributed optimization problems, a common challenge when agents need to communicate frequently with a central coordinator. Specifically, we propose a hybrid approach that combines **Gaussian Processes (GPs)** with **adaptive uniform quantization** to minimize the communication load while maintaining a solid level of accuracy. Key contributions include: - The integration of GPs into distributed optimization schemes to reduce communication rounds. - The use of **uniform quantization** to decrease the size of transmitted data, reducing overall communication costs. - Adjustments to GPs to account for quantization error, ensuring reliable performance. - Techniques like **orthogonal transformations** and **dithering** to enhance communication efficiency. Through simulations, our methods demonstrate significant communication reduction without sacrificing the quality of the optimization solutions. Feel free to check out the full paper here: