My research

Ryan Reece

Research interests

Figure 1: A display of likely a pp → tt → bb\mu\tau event in ATLAS in 2011.

Figure 1: A display of likely a \(pp\) → \(tt\) → \(bb\mu\tau\) event in ATLAS in 2011.

Research history

ATLAS

Previously as a graduate student at Penn, working with ATLAS Experiment at CERN, I helped commission the Transition Radiation Tracker (TRT), a sub-detector of the ATLAS tracker, during the start-up of the Large Hadron Collider (LHC) in 2009-2012. On July 4 of 2012, the ATLAS and CMS experiments both announced discovering a new particle consistent with the long-sought-after Higgs boson, a key to explaining electroweak symmetry breaking in the Standard Model of particle physics.

I continued to work in the ATLAS Experiment as a postdoc at the Santa Cruz Institute for Particle Physics (SCIPP). I’ve spent 10 years romping through datasets from ATLAS, learning how to explore and model large, multidimentional datasets. I have a knack for developing data analysis frameworks and an eye for technical detail and good scholarship. I am passionate about using scientific techniques to solve important problems and about how technologies extend our reach.

Figure 2: A figure I drew with [Kyle Cranmer](http://theoryandpractice.org/) of the big-picture of the flow of ATLAS data.

Figure 2: A figure I drew with Kyle Cranmer of the big-picture of the flow of ATLAS data.

Many theories of physics beyond the Standard Model have revolutionary implications for the concepts of symmetry and space-time, and for our understanding of the early universe. Searching for new physics in collider data involves detecting rare events among many, requiring big data reductions, data-driven background modeling, and severe statistical hypothesis testing.

My research in physics has primarily focused on supporting and optimizing the reconstruction of hadronic tau decays, and searching for exotic new physics in ditau and diphoton events, including for signs of grand unified theories and supersymmetry. In my time with the ATLAS Collaboration I made substantial contributions to 14 publications in scientific journals and several more physics conference proceedings.

Figure 3: Search for Z^{\prime} → \tau\tau with ATLAS ([1608.00890](https://arxiv.org/abs/1608.00890)).

Figure 3: Search for \(Z^{\prime}\) → \(\tau\tau\) with ATLAS (1608.00890).

I have extensive experience in Neyman-Pearson statistical hypothesis testing, including the modeling of systematic uncertainties, often with a combination of Monte Carlo and data-driven methods. In analyses for ATLAS, I have contributed to the measurement and calibration of many systematic uncertainties including tracking detector thresholds, tau identification efficiencies and fake rates, and the effects of QCD Monte Carlo variations.

Tau leptons are heavy enough to have complex hadronic decays with identifiable signatures, which form an interesting domain for machine learning. During 2009-2017, I was an active developer for the ATLAS Tau Working Group that developed the reconstruction and identification algorithms for taus in ATLAS during Run 1 and 2 of the LHC. I lead the developement and optimized the first cut-based tau identification used with ATLAS data, and I helped develop more advanced tau identification using Boosted Decision Trees (BDTs).

Since 2016, following the deep learning revolution, I’ve been trying to learn everything I can about deep learning. I’ve worked in computer vision, machine translation, and large-scale language modeling.

Cerebras

In April of 2018, I joined Cerebras Systems as a machine learning engineer. Cerebras makes cutting-edge accelerators for deep learning, the first to achieve Wafer-Scale Intergration. Collaborators at Cerebras and I published a new normalization technique, Online Normalization, for training deep neural networks with small batch sizes (NeurIPS 2019, arxiv:1905.05894).

Figure 4: Training ResNet-50 on ImageNet-1k with [Online Normalization](https://github.com/Cerebras/online-normalization) (ON) and Batch Normalization (BN) ([NeurIPS 2019](https://papers.nips.cc/paper/2019/hash/cb3ce9b06932da6faaa7fc70d5b5d2f4-Abstract.html), [1905.05894](https://arxiv.org/abs/1905.05894)).

Figure 4: Training ResNet-50 on ImageNet-1k with Online Normalization (ON) and Batch Normalization (BN) (NeurIPS 2019, 1905.05894).

I developed and supported several reference models for Cerebras in computer vision (ResNets) and NLP (GNMT, Transformer, Linformer, BERT, RoBERTa, GPT-2). Cerebras revealed its first product at Supercomputering 2019: the poweful CS-1 computer for AI. I helped represent Cerebras at its tradeshow booth. I directly supported customers at both national labs and industry in applied deep learning tasks ranging from natural language to physics surrogate models. In April of 2022, I co-authored a blog about Getting started with PyTorch BERT models on the Cerebras CS-2 System.

Tenstorrent

Figure 5: Tenstorrent started offering AI inference as a service through [OpenRouter](https://openrouter.ai/provider/tenstorrent), including GLM 5.3 and Qwen3.8-27B models, in October 2026.

Figure 5: Tenstorrent started offering AI inference as a service through OpenRouter, including GLM 5.3 and Qwen3.8-27B models, in October 2026.

Current independent research projects

Mechanistic interpretability

I am starting to demonstrate some basic observations in mechanistic interpretability. See the project ioi-hydra, where I replicated studies of a circuit in GPT-2 that can be interpreted as doing Indirect Object Identification (IOI), which self-repairs when ablated.

Figure 6: Self-repair in GPT-2 small’s IOI circuit.

Figure 6: Self-repair in GPT-2 small’s IOI circuit (replication of Wang et al. 2022 and McGrath et al. 2023, “The Hydra Effect”).

Portfolio theory

I have studied and implemented techniques of Modern Portfolio Theory (AKA the “Markowitz portfolio”) and its variants. I’ve shared some of my thoughts in my Finance Notes.

Of particular importance to this study, I have implemented covariance matrix calculators along with methods for computing confidence intervals on the covariance estimates.

Figure 7: Markowitz portfolio analysis (left). Solving Kuhn and Leduc, simplified versions of poker, with CFR (right).

Figure 7: Markowitz portfolio analysis (left). Solving Kuhn and Leduc, simplified versions of poker, with CFR (right).

Counterfactual Regret Minimization (CFR)

I have worked on a type of reinforcement learning called Counterfactual Regret Minimization (CFR) and its variants to solve games of imperfect information like poker. In particular I’ve studied and implemented CFR+ and MCCFR from Noam Brown’s PhD thesis, Equilibrium finding for large adversarial imperfect-information games (2020). I have not open sourced this project, but feel free to ask me about it.

Previous focuses

Curriculum Vitae

Ph.D. thesis

  Reece, R. (2013). A search for new physics in high-mass ditau events in the ATLAS detector.

My graduate research in particle physics was on the reconstruction and identification of hadronic tau decays with the ATLAS experiment, measuring the Z→ττ production cross section in proton-proton collisions at √s = 7 TeV, and searching for new physics in high-mass ditau events.