Riki Miyoshi
JA

Riki Miyoshi

Toward the ideal, shine.

Education

  1. 2019 – 2023

    Komaba Senior High School, University of Tsukuba

  2. 2023 – 2025

    Kingswood School, Bath

  3. 2025 – Present

    University of Cambridge, Computer Science (BA)

    Highest Achieving Part IA Student (Ranked 1 / 129)

Affiliation

  1. 2023 – PresentTazaki Foundation Scholar
— 01Vision
Seeing the world through the lens of intelligence.

No field of inquiry can outrun the limits of the mind doing the inquiring — which is why I care about understanding intelligence itself, and eventually extending it. I want to take the fields I'm curious about, however much they shift, and expand them through the lens of intelligence.

— 02Approach

Mastering an interdisciplinary view

Intelligence tends to get reduced to artificial intelligence, but it's much broader than that. Beyond neuroscience, computational neuroscience, and NeuroAI, I want to approach it through theories of consciousness (active inference, IIT), mathematics, physics, philosophy, and art — an interdisciplinary pursuit of what it actually is.

01

Moving between thinking and making

Ideas rarely stay complete in my head — I test them by building something small, and pay attention to what feels off once it's real.

02

Keeping the door to curiosity open

I don't commit to one lane too early. What looks like a detour often turns out to connect later, in ways I couldn't have planned.

03

Designing for consistency, not willpower

I look for setups that don't depend on motivation to keep going. This site is one small experiment in that direction.

— 03Research

Current Research

Jun 2026 – Present

Extending Policy Gradient Methods — University of Tokyo

Reinforcement learning feels close to the essence of intelligence to me: an agent coming to understand its environment through the interaction between itself and that environment, mediated by reward. But I'm still far short of the knowledge and experience I need. Working through the underlying theory head-on, in both math and code, was necessary to build real mathematical strength and intuition. With the good fortune of a great mentor, I'm now working to understand and extend the theory behind policy gradient methods — specifically ordergrad, a method for optimizing metrics like max@k and pass@k.

Sep 2025 – Feb 2026

Reevaluating Hierarchical World Models

I think hierarchy matters for intelligence: it's what lets a system form abstract concepts and explain its environment with far less compute. But the relationships within and between levels of a hierarchy shift dynamically as learning progresses and the environment changes, so a model needs to learn that structure naturally rather than have it fixed in advance. I'm also drawn to the idea behind world models more broadly — giving an agent an explicit internal model of its environment as a prior, which seems to pay off in both learning efficiency and interpretability. I ran experiments probing the limits of Hieros, a hierarchical world model, by adding S5 layers and benchmarking against baselines like Director. Eventually, I'd like to build something that gets past those limits.

— 04Interests

What I'm curious about

A set of lenses for looking at the world. The connections grow, and deepen, over time.

  • Mathematical Finance — Reading an introductory Blue Backs book on mathematical finance.
— 05Activity

Activities so far

BWSI (MIT Beaver Works Summer Institute) — Cogworks · Japan Brain Bee National Finals — 5th place · Tokyo Economics-Koshien — Winner · UK Mathematical Olympiad Round 1 — Gold · UK Physics Olympiad Round 1 — Gold · UK Chemistry Olympiad Round 1 — Gold · UK Linguistics Olympiad Round 1 — Gold · British Mathematical Olympiad (BMO1) — Bronze · British Mathematical Olympiad (BMO2) — Merit

— 06Contact

Contact

About the research, or about the foundation — feel free to reach out.

rikimys73@gmail.com
InstagramX (Twitter)LinkedInGitHub