Fred Sala

LLM-as-a-judge—often with multiple judges—is now the standard for scalable model evaluation, yet judge biases and correlations can amplify errors. We cast aggregation as inference in a latent-factor Markov random field that jointly models a latent true-quality variable, inter-judge correlations, and confounders (e.g., generation length). We address two key technical challenges—identifiability and learning a higher-rank latent structure—via CARE, a two-stage estimator that…

While Transformers underpin modern large language models (LMs), there is a growing list of alternative architectures with new capabilities, promises, and tradeoffs. This makes choosing the right LM architecture challenging. Recently-proposed hybrid architectures seek a best-of-all-worlds approach that reaps the benefits of all architectures. Hybrid design is difficult for two reasons: it requires manual expert-driven search, and new hybrids must…

Evaluating the effectiveness of unlearning in large language models (LLMs) remains a key challenge, especially as existing metrics often rely on specific reference outputs. The widely used forget quality metric from the TOFU benchmark compares likelihoods over paraphrased answers but is highly sensitive to the choice of the reference answers, potentially obscuring whether a model has truly forgotten the targeted information. We…

Verifiers can enhance language model (LM) performance by scoring and ranking a set of generated responses, but high-quality verifiers today are either unscalable (like human judges) or of limited practical use (such as formal proof tools like Lean). While LM-based judges and reward models serve as general-purpose verifiers, they still fall short of the performance levels achieved by oracle verifiers,…

Modern semi-supervised learning (SSL) methods rely on pseudolabeling and consistency regularization. Pseudolabeling is typically performed by comparing the model’s confidence scores and a predefined threshold. While several heuristics have been proposed to improve threshold selection, the underlying issues of overconfidence and miscalibration in confidence scores remain largely unaddressed, leading to inaccurate pseudolabels, degraded test accuracy, and prolonged training. We take…

In this post, we unpack how Snorkel built a realistic benchmark dataset to evaluate AI agents in commercial insurance underwriting. From expert-driven data design to multi-tool reasoning tasks, see how our approach surfaces actionable failure modes that generic benchmarks miss—revealing what it really takes to deploy AI in enterprise workflows.

Concept-based approaches, which aim to identify human-understandable concepts within a model’s internal representations, are promising for interpreting embeddings from deep neural network models, such as CLIP. While these approaches help explain model behavior, current methods lack statistical rigor, making it challenging to validate identified concepts and compare different techniques. To address this challenge, we introduce a hypothesis testing framework that…

The weak-to-strong generalization phenomenon is the driver for important machine learning applications including highly data-efficient learning and, most recently, performing superalignment. While decades of research have resulted in numerous algorithms that produce strong empirical performance, understanding what aspects of data enable weak-to-strong generalization has been understudied. We propose a simple data-centric mechanism that characterizes weak-to-strong generalization: the overlap density. Intuitively,…

We introduce a benchmark to evaluate the capability of AI to solve problems in theoretical physics, focusing on high-energy theory and cosmology. The first iteration of our benchmark consists of 57 problems of varying difficulty, from undergraduate to research level. These problems are novel in the sense that they do not come from public problem collections. We evaluate our data…