Brendan Hogan Rappazzo

Brendan Hogan Rappazzo is an AI researcher whose work includes autonomous experimental research and specialized language-model training. His personal biography identifies him as Brendan Hogan, an applied researcher at Prime Intellect working on environments, post-training, and continual learning. Previously, he was a machine learning research scientist at Morgan Stanley, working on large language models and autonomous research agents. His publications also use the name Brendan R. Hogan.[2][3][4]

Hogan coauthored AlphaLab, submitted on 31 March 2026, with Xiwen Chen, James T. Wilson, Kashif Rasul, Adel Boyarsky, Thomas Kamei, Anderson Schneider, and Yuriy Nevmyvaka. The system starts with a dataset and a natural-language objective, then explores the data, writes analysis code, and produces a research report. Before running its experiment campaign, it constructs and adversarially validates its own evaluation framework.[3]

AlphaLab runs GPU experiments through a Strategist/Worker loop, retaining domain knowledge in a persistent playbook that serves as online prompt optimization. Model-generated adapters supply domain-specific behavior, allowing the same pipeline to tackle different optimization tasks. The authors evaluated GPT-5.2 and Claude Opus 4.6 on CUDA kernel optimization, language-model pretraining, and traffic forecasting. They reported 22% lower pretraining validation loss than a single-shot baseline using the same model; neither model dominated uniformly across domains, which they interpreted as evidence for complementary search coverage.[3]

In its 29 July 2026 finance-track coverage, Latent Space identified him with Morgan Stanley and framed his presentation around whether humans can trust the experimental environments that multi-agent research optimizes within. His personal biography also names Golden Swans, a swarm of agents for long-term company analysis, among his Morgan Stanley projects.[1][2]

Before AlphaLab, Hogan worked on adapting language models to Q, a programming language used in quantitative finance with much less internet representation than mainstream languages. He coauthored the full-stack fine-tuning report with Will Brown, Adel Boyarsky, Anderson Schneider, and Yuriy Nevmyvaka. Submitted on 9 August 2025 and revised on 12 August, the work addressed a specific limitation of general-purpose models: weak performance in a niche programming domain.[4]

The team built a Leetcode-style Q evaluation dataset, benchmarked frontier models, and combined pretraining, supervised fine-tuning, and reinforcement learning to produce reasoning and non-reasoning Qwen-2.5 models ranging from 1.5 billion to 32 billion parameters. The authors reported 59% pass@1 accuracy for their best model on their Q benchmark. They released models, code, and data alongside a blueprint for dataset construction and training.[4]

Earlier, Rappazzo pursued doctoral work in Computer Science at Cornell with Carla Gomes in Computational Sustainability, following Bioengineering and Physics study at the University of Maryland, where he graduated in 2016. Two papers submitted on 23 September 2024 illustrate his earlier contributions to classifier training and retrieval: Critic Loss and GEM-RAG.[2][5][6]

With Aaron Ferber and Carla Gomes, he developed Critic Loss for Image Classification. A correctness critic receives an image, the classifier’s intermediate embeddings, and its predictions, then estimates the probability that the classification is correct. That estimate supplies a training signal. Because the critic does not receive the label as an input, it can train the classifier using both labeled and unlabeled data; its estimates also help select examples for labeling in active learning. The authors reported improved generalization and calibration in the studied settings across three image-classification datasets.[5]

The ICMLA 2024 awards record lists Rappazzo, Ferber, and Gomes as recipients of its Best Paper Award, naming their paper “Loss for Image Classification.”[7]

With Yingheng Wang, Ferber, and Gomes, he coauthored GEM-RAG, which augments retrieval with information about what text can help answer. It tags text chunks with LLM-generated utility questions, connects chunks in a graph using similarities in both their text and questions, and uses eigendecomposition to build higher-level summary nodes capturing the corpus’s main themes. The authors reported improvements over comparison RAG methods on two question-answering tasks using UnifiedQA and GPT-3.5 Turbo with SBERT and OpenAI text encoders.[6]

Beyond these technical projects, Hogan expresses an interest in AI’s relationship with art and philosophy. In his personal biography, he describes AI “not just as a tool, but as a way to understand consciousness, spark awe, and open new directions for research.”[2]