Marc Cheng

ZHONGMANG CHENG

Research Track

Research Interest: How LLMs handle self-referential processing — specifically whether the activation patterns during in-context reasoning bear any meaningful resemblance to what neuroscience describes as the Default Mode Network.

Academic Background

University of Toronto (St. George Campus) — Trinity College

Honours B.S. in Mathematics with High Distinction
Mathematics & Its Applications Specialist (Probability/Statistics)
September 2019 - June 2023
GPA: 3.6/4.0 | Dean's List: 2021, 2022, 2023

When the Loop Closes Working paper · Under submission to CSCW 2027

arXiv:2604.15343 · Z. Cheng, N. Song · April 2026

Architectural limits of in-context isolation, metacognitive co-option, and the two-target design problem in human-LLM systems.

On the implementation side: Mamba SSM, GPT-2 Replication, ResNet and ConvNet — mostly to understand architectures from the inside. Code on GitHub.

Corporate Track

DoubleTrends

Front-End Developer (Freelance) | June 2025 – Present

End-to-end ownership of something that ships. Design, build, deploy, payments, DNS — the full surface area. Useful as a forcing function for the parts of web development that don't show up in tutorials.

Rexel Canada Electric Inc.

Product Data Specialist | June 2023 - May 2025

A real corporate job. The interesting part wasn't my work — it was learning what enterprise data infrastructure actually looks like when it's been accreting for decades. AS400 feeding into TIBCO feeding into Informatica feeding into SAP Hybris feeding into a webshop. Nobody fully understands the whole chain. You learn to work with that.

Deloitte Canada

Data Analyst (Co-op) | July 2021 - April 2022

Model validation rather than model building — which turns out to be a different skill. You're reading someone else's architecture and asking whether the assumptions hold against real data.

Outside of full-time roles, I've contributed to ML infrastructure for an early-stage algorithmic trading startup (Stoic AI), setting up cloud GPU training environments on RunPod using H100 instances. Full work history on LinkedIn.

Everything Personal

Travel & Life

I've been moving between countries most of my life — born in Beijing, grew up in Shanghai, spent my teens in Montreal, and did university in Toronto. At some point the airport stopped feeling like a destination.

In 2023 I did a cross-continental road trip through North America: Toronto, Calgary, Vancouver, Seattle, San Francisco, Los Angeles, Las Vegas, Albuquerque, Oklahoma City, St. Louis, Chicago, Detroit.

That same summer I flew to Japan for a graduation trip.

In 2025 I did a solo trip to South America — Cusco, Machu Picchu, Lima, Buenos Aires, Rio de Janeiro. My camera got robbed in Rio. Lost all the video footage. The photos survived.

Art & Music

I studied drawing, watercolor, oil painting, and charcoal for several years under the supervision of Georges Kalmetti in Montreal — starting from around age 12, across three different studios.

On the music side I'm entirely self-taught: piano, guitar, ukulele, thumb piano. Mostly pop. No sheet music, no theory exams. I have an electric piano and an electric guitar at home that I play occasionally and probably not well enough.

Outdoors

I got into skateboarding at 16 with a penny board and eventually worked through popsicle, cruiser, longboard, and dance board. Not actively skating anymore.

I climbed for a while — bouldering and free climbing, currently around 5.8 YDS. Also done skydiving once, which I'd recommend to anyone who hasn't.

Other things I've done at least once: canoe trip, flying a single-engine propeller plane, hiking Algonquin in peak fall.

Anime & Manga

The anime I keep coming back to tend to share a structural quality: multiple threads running in parallel, each partially obscure, closing together at the end in a way that makes everything already there suddenly legible. Ryohgo Narita's (成田良悟) work is the clearest example. Steins;Gate operates on the same principle, except the convergence mechanism is time itself.

My other consistent preference is emotional precision over emotional volume. Hyouka (冰菓) is the reference point — nothing dramatic happens, but the ambivalence between the two leads is rendered carefully enough that it stays. Scum's Wish (人渣的本愿) did something I found genuinely useful: it made explicit that feelings aren't binary, which is obvious to state but rarely dramatized.

On the visual side, Moeka Yokokura's (横仓萌果) art is the clearest example of a style I find myself drawn to — luminous, slightly melancholic, the kind of linework that makes a still frame feel emotionally loaded.

2nd President of UTACG

University of Toronto Anime Comic and Game

President (Nov 2021 - Jan 2023): Led an executive committee of 10, organized 14 events, generated $33000 in revenue, $2200 net profit. Produced a Lunar New Year Gala with 30 performers and ~200 spectators.

Gaming

The games I've put the most time into tend to be systems-heavy enough that understanding them is the activity: Kerbal Space Program, Cities: Skylines, Stellaris.

Final Fantasy XIV is the edge case: 2000 hours in, stopped at Endwalker. It closes the ten-year arc with enough finality that continuing past it felt like revisiting something that had already ended. Stopping there was the right call.

NieR: Automata is probably the game I'd point to if someone asked what the medium can do that other formats can't — the soundtrack layers in a way that only makes structural sense after multiple playthroughs, and the world-building earns its philosophical weight rather than just gesturing at it.

Speaking of metagame, Ever17's narrative construction is genuinely ahead of its time, and The Stanley Parable remains the clearest example of a game that makes its own form the subject.

For multiplayer I'll occasionally show up in League, VALORANT, Overwatch, or Delta Force depending on who's online.

Other Social