Nicholas M. Boffi

Research profile

Research summary

Nicholas M. Boffi is a machine learning researcher at Carnegie Mellon University whose recent work centers on flow-based generative modeling and its interface with sampling, control, and reasoning. Evolve's OpenAlex-derived field distribution assigns equal weight (0.5 and 0.5) to machine learning (generative models) and machine learning, and topic counts span generative adversarial networks and image synthesis (6), Markov chains and Monte Carlo methods (4), topic modeling (3), model reduction and neural networks (3), and adversarial robustness (3). Two 2026 arXiv preprints anchor the reviewed record: 'Thinking with Looped Flows', which trains stateful denoisers with local flow losses for recurrent test-time reasoning, and 'WTF?! Simulation-Free Reinforcement Learning with Wasserstein-Tilted Flow Maps', a last-author paper that fine-tunes flow maps for reward optimization through a drift-based optimal-transport regularizer and deterministic optimal control; Evolve judges the theory sound and the results strong while flagging reward-metric and compute caveats. Activity metrics indicate a rising tempo: counts-by-year moves from 1 work in 2023 to 8 in 2026, with 8 works in the trailing three-year window, across a moderately sized network of 12 distinct coauthors (collaboration score 51). Reproducibility scores 55 across two medium-readiness papers. Citation-level impact cannot yet be assessed: h-index and i10-index are logged as 0, so the output and impact axes remain unscored.

Research keywords

Highlights

• Simulation-free RL fine-tuning of flow maps

  As last author on the 2026 arXiv preprint 'WTF?! Simulation-Free Reinforcement Learning with Wasserstein-Tilted Flow Maps', Boffi contributes to a method that fine-tunes flow maps for reward optimization using a drift-based optimal-transport regularizer and deterministic optimal control. Evolve's review finds the theory sound and the results strong, while flagging caveats around reward metrics and compute cost; reproducibility readiness is rated medium.

• Looped flows for recurrent reasoning

  The 2026 preprint 'Thinking with Looped Flows' trains stateful denoisers with local flow losses to support recurrent reasoning at test time. Evolve's review reports strong gains over TRM/GRAM on ARC, but notes that baselines rely on published numbers rather than reruns; reproducibility readiness is medium.

• A flow-centric toolkit spanning generative modeling and control

  Both reviewed papers share one toolkit—flow maps, probability-flow generative modeling, and optimal transport—matching the OpenAlex field split of 0.5 machine learning (generative models) and 0.5 machine learning. The recurrence suggests a deliberate program connecting generative sampling, control, and test-time reasoning rather than scattered one-off projects.

Timeline

1. 2023Earliest entry in the OpenAlex counts-by-year record for this author: 1 work published in 2023.
2. 2026Co-authors 'Thinking with Looped Flows' (middle authorship), training stateful denoisers with local flow losses for recurrent test-time reasoning.
3. 2026Last author on 'WTF?! Simulation-Free Reinforcement Learning with Wasserstein-Tilted Flow Maps', which fine-tunes flow maps for reward optimization via a drift-based optimal-transport regularizer and deterministic optimal control.
4. 2026Counts-by-year peaks at 8 works in 2026 (recent-3-year works: 8), with both Evolve-reviewed papers falling in this year.

Papers in the Evolve catalog

• 2026-09-22

  WTF?! Simulation-Free Reinforcement Learning with Wasserstein-Tilted Flow Maps

• 2026-09-10

  Thinking with Looped Flows

Frequent co-authors