We are LayerLens, a project focused on building better resources for independent, transparent evals for frontier AI models. Atlas is a community resource intended to provide insights about the performance of the top foundational models through independent evals on benchmarks such as MATH, HumanEval, and MMLU. LayerLens is a team of engineers and data scientists who have been constantly frustrated by the lack of independent verification for LLM performance. Most benchmarks come from the model creators themselves, and for developers, building an independent evaluation pipeline is often more trouble than it is worth. Open-source leaderboards, while admirable, often do not provide enough transparency, and are often too scientific for the average user. While evals have historically been a tool to measure the proverbial progress toward AGI, they have become increasingly relevant for validating LLM performance. Large enterprise teams and independent hackers alike use evals as a way to select the right model for a particular use-case, all while depending on singular “accuracy” metrics. Atlas is an LLM analytics leaderboard that is both simple and highly detailed. You can view the top models, sorted by region, vendor type, or a particular use-case, via evaluation spaces. You can use the battleground to compare two models on an individual benchmark, getting prompt by prompt comparisons for each entry. For any individual evaluation run, you can get a clean summary of model performance on individual subsets. And finally, each model page has its own dedicated analytics and information section. This is only our first iteration of the product. We eventually want to release the same suite for custom models, agents, evals and more. We will be around to answer any questions on our product!