Feeling inspired to write your first TDS post? We’re always open to contributions from new authors.
With the pace at which large language models continue to evolve, staying up-to-date with the field is a major challenge. We see new models, cutting-edge research, and LLM-based apps proliferate on a daily basis, and as a result, many practitioners are understandably concerned about falling behind or not using the latest and shiniest tools.
First, let’s all take a deep breath: when an entire ecosystem is moving rapidly in dozens of different directions, nobody can expect (or be expected) to know everything. We should also not forget that most of our peers are in a very similar situation, zooming in on the developments that are most essential to their work, while avoiding too much FOMO—or at least trying to.
If you’re still interested in learning about some of the biggest questions currently dominating conversations around LLMs, or are curious about the emerging themes machine learning professionals are exploring, we’re here to help. In this week’s Variable, we’re highlighting standout articles that dig deep into the current state of LLMs, both in terms of their underlying capabilities and practical real-world applications. Let’s dive in!
- Navigating the New Types of LLM Agents and Architectures
In a lucid overview of recent work into LLM-based agents, Aparna Dhinakaran injects a healthy dose of clarity into this occasionally chaotic area: “How can teams navigate the new frameworks and new agent directions? What tools are available, and which should you use to build your next application?” - Tackle Complex LLM Decision-Making with Language Agent Tree Search (LATS) & GPT-4o
For his debut TDS article, Ozgur Guler presents a detailed introduction to the challenges LLMs face in decision-making tasks, and outlines a promising approach that combines the power of the GPT-4o model with Language Agent Tree Search (LATS), “a dynamic, tree-based search methodology” that can enhance the model’s reasoning abilities. - From Text to Networks: The Revolutionary Impact of LLMs on Knowledge Graphs
Large language models and knowledge graphs have progressed on parallel and mostly separate paths in recent years, but as Lina Faik points out in her new, step-by-step guide, the time has come to leverage their respective strengths simultaneously, leading to more accurate, consistent, and contextually relevant outcomes.
- No Baseline? No Benchmarks? No Biggie! An Experimental Approach to Agile Chatbot Development
After the novelty and initial excitement of LLM-powered solutions wears off, product teams still face the challenges of keeping them working and delivering business value. Katherine Munro covered her approach to benchmarking and testing LLM products in a recent talk, which she’s now transformed into an accessible and actionable roadmap. - Exploring the Strategic Capabilities of LLMs in a Risk Game Setting
Hans Christian Ekne’s recent deep dive also tackles the problem of evaluating LLMs, but from a different, more theoretical direction. It takes a close look at the different strategic behaviors that leading models (from Anthropic, OpenAI, and Meta) exhibit as they navigate the rules of classic board game Risk, discusses their shortcomings, and looks at the potential future of LLMs’ reasoning skills. - How to Improve LLM Responses With Better Sampling Parameters
We round out this week’s lineup with a hands-on, practical tutorial by Dr. Leon Eversberg, who explains and visualizes the sampling strategies that define the output behavior of LLMs—and demonstrates how understanding these parameters better can help us improve the outputs that models generate.