I’m Jiaqian (Janelle) Zhu, a Computer Science PhD student passionate about physics, mathematics, and first-principles thinking. My research spans World Action Models (WAMs), robotics, and efficient AI, with a particular focus on physics-informed AI, geometric physics, and quantum mechanics. My long-term vision is to build intelligent systems grounded in the fundamental laws of physics. I’m deeply curious, constantly exploring new ideas, and driven to challenge existing assumptions and turn theoretical insights into real-world technologies.

As of October 2026, I have two first-author papers accepted at EMNLP 2026 and NeurIPS 2026, both independently led from initial ideas through experiments and final writing, with several more under submission. I’m highly self-motivated and goal-oriented, typically working from 9 AM to 10 PM, seven days a week. Beyond academia, my ambition is to build transformative AI technologies and pursue entrepreneurship. I’m actively looking to connect with visionary co-founders, investors, and researchers who share my curiosity, ambition, and passion for building what doesn’t yet exist. Feel free to reach out to exchange ideas or explore opportunities together.

Research

Writing

  • Summaries & reflections
  • Breaking down essays
  • Daily paper-reading log

Startup

Otherwise

Publications

Conferences

NeurIPS 2026

Beyond Unit-Circle Eigenvalues: Invariant Bases for Stable State Space Dynamics

Jiaqian Zhu et al. First author · Project lead

EMNLP 2026

Lexical Perturbations Disrupt LLM Reasoning: An Empirical Study of Attention Diversion

Jiaqian Zhu et al. First author · Project lead

arXiv

Journals

Ceramics International, 50(2), 3018–3025, 2024

Accelerating design of glass substrates by machine learning using small-to-medium datasets

Jiaqian Zhu et al. First author

Paper
Journal of Non-Crystalline Solids, 646, 123251, 2024

Advancing the prediction of crystalline phases in glass-ceramics via machine learning

Jiaqian Zhu et al. First author

Paper

Operating Principles

FULL MEMO →

What actually gets bet on at the frontier. Technical skill is the ticket in, not the edge — what gets funded sits above it.

  1. Clarity — say it in one sentence: what, why now, why me. If I can't compress it, I don't understand it yet.
  2. Obsession — work the problem unbidden and late. If nothing pulls like that, the job is to find the thing that does.
  3. Slope over intercept — current level is noise; six-month velocity is signal. Don't be trapped by where I start.
  4. Agency — "hard" usually means "I don't own the tool yet." Build the tool. Turn can't into haven't yet.
  5. Patience — hold what compounds; sit through the boring middle. Judge on the year, not the week.
  6. Mission over money — subtract the upside: would I still do it? If only rewards move me, I'll trade the long game for the short one.

Choosing a co-founder — the heaviest, least-reversible bet

  • Mission alignment is a veto, not a bonus.
  • See them fail first — carry weight together before building together.
  • Complement, not comfort — same instincts means a shared blind spot.
  • Slope, not résumé.
  • The real move: don't hunt for someone who fits this — become that person first.

Every principle above is inert until the work meets a real user, a real market, a real chance of rejection. Ship it. Show up.

Paper Reading

FULL LOG →

Papers I've actually read, with notes.

  • 2026.07.06

    Paper Reading Notes - 2026-07-06 Session

    Running Q&A log over the 2026-07-06 session - 7 papers. Deep-read: 01 MIPU (Monotonic Inference Policy Update - optimize the deployed inference policy, not just the training policy, under training-inference mismatch; full method + Q&A) and 04 VLA-Corrector (lightweight detect-and-correct inference giving action-chunked VLA policies an adaptive action horizon). Also logged: 02 DataComp-VLM (data-centric benchmark for VLM training) and 03 Perceive-to-Reason / P2R (decoupling perception from reasoning for fine-grained visual reasoning, compared head-to-head with PixelEyes). To read: 05 The State-Prediction Separation Hypothesis, 06 OrbitQuant (data-agnostic quantization for diffusion transformers), 07 Optimizing Visual Generative Models via Distribution-wise Rewards.

    reading-notes
  • 2026.07.05

    Paper Reading Notes - 2026-07-05 Digest

    Running Q&A log over the 2026-07-05 paper digest. Paper 01 - PixelEyes: decoupling perception (an external SAMTok mask) from reasoning (the VLM) for pinpoint visual evidence seeking.

    reading-notes
  • 2026.07.02

    ABot-M0.5 — Reading Notes

    Close reading of ABot-M0.5 (AMAP CV Lab, Alibaba): a unified mobility-and-manipulation World Action Model aligning temporal granularity, action space, and train-test consistency.

    reading-notes
  • 2026.07.01

    World Models, Agents & VLA — Paper Reading Notes

    Close readings of five AI papers: DreamForge (real-time controllable world model), LUMOS (semantic OS layer for agents), QVAL (evaluating dense-supervision signals via Q-alignment), MemLearner (learned context memory for video world models), and Drop-Then-Recovery (redundancy analysis of vision-language-action models).

    world-modelsvideo-generationdiffusionagentsaccessibilityrlevaluationvlarobotics
  • 2026.06.29

    World Models for Robotic Control — Paper Triage & Notes

    A 10-paper triage centered on world models for robotic control, with deep reads of PhysiFormer (flow-matching diffusion over raw 3D vertex trajectories, view-invariant, factorized DiT), Fast-LeWM (action-prefix encoding + parallel latent prediction over a JEPA latent, AdaLN conditioning, CEM planning), and ICWM (in-context system identification from self-probed random clips); plus dives into the Markov state and second-order ODEs, AdaLN vs. cross-attention, FLOPs vs. latency, and two new-idea brainstorms — action-prefix-conditioned latent-trajectory diffusion and self-organized core-periphery modulation.

    roboticsworld-modelsjepadiffusionin-context-learningstudy-notes

Reading List

FULL LIST →

Frontier papers I want to read — updated daily.

  • 2026.07.15

    Paper Digest — Reading Queue (2026-07-15)

    A triaged 2026-07-15 reading queue with arXiv links and first-page screenshots, organized as one signal supply chain: sourcing training signal (TerraZero — procedural driving self-play, zero demos), transferring it between models (Direct-OPD, PUST), routing it so it gets used (ABot-N1 pixel-goal interface, the Knowing–Using Gap), and the weak link that everything assumes — exploration (MACE). Two industry-led papers (Applied Intuition, Alibaba AMAP) are the focus. Includes a worked deep dive on porting the reuse layer into embodied policies. Theme of the day: signal supply chain & industry infrastructure.

    reading-queueembodied-aivlapost-trainingweak-to-strongagentsbenchmarksinterpretabilityinference-efficiencystudy-notes
  • 2026.07.04

    Reading Queue — Want to Read (2026-07-04)

    A triaged 2026-07-04 backlog with arXiv links and first-page screenshots: task-agnostic VLA pretraining (TAP), agentic benchmarking & capability measurement (EvoPolicyGym, AgenticDataBench, PACE, HealthAgentBench), agent memory as a trainable skill (AutoMem, DuoMem), and inference / distillation / scaling limits (When More Sampling Hurts, Denser != Better, Seed2.0), plus an industry pulse. Theme of the day: agentic benchmarking and memory-as-a-skill.

    reading-queueembodied-aivlaagentsagent-memorybenchmarksinference-efficiencydistillationstudy-notes
  • 2026.07.03

    Reading Queue — Want to Read (2026-07-03)

    A triaged 2026-07-03 backlog with arXiv links: controllable world simulation (WorldDirector), agent memory / reliability / evaluation (AgenticSTS, MemSyco-Bench, SkillCoach, SWE-Interact, DiscoPER), RL & reasoning training (the GRPO/Dr.GRPO/DAPO identity, Transfer-Aware Curriculum for multi-domain RLVR), and medical AI (step-aware RL for medical reasoning, discrete-diffusion radiology drafting), plus an industry pulse. Theme of the day: agent memory & reliability are the new frontier.

    reading-queueworld-modelsagentsagent-memoryrlrlvrmedical-aireasoningstudy-notes
  • 2026.06.30

    Reading Queue — Want to Read (2026-06-30)

    A triaged 2026-06-30 backlog with arXiv links and first-page screenshots: embodied foundation models (Qwen-RobotManip, Qwen-RobotNav, Vesta), agents / long-horizon autonomy & evaluation (Agents-A1 horizon scaling, Agentic Abstention, AgentOdyssey, TUA-Bench, OSWorld2.0), and distillation / inference efficiency (AsyncOPD, Simplified Sparse Attention), plus an industry pulse. Theme of the day: scaling agent horizons, not parameters, and unifying reasoning+memory+action in one backbone.

    reading-queueembodied-aivlaagentslong-horizonbenchmarksdistillationinference-efficiencystudy-notes
  • 2026.06.29

    Reading Queue — Want to Read

    A triaged backlog of good papers I want to read but haven't yet: embodied AI / world models / sim-to-real (PhysisForcing, SimFoundry, Learning to Fold, Object-Centric Residual RL), agentic systems and tool use (GBC, PAT, ProMSA, Tool-Suppression), and inference efficiency / generative (ConvFill, Qwen-Image-2.0-RL), plus a condensed industry pulse.

    reading-queueembodied-aiworld-modelsagentsinference-efficiencystudy-notes

Startup & Business

FULL LOG →
  • 2026.07.09 · STUDY-SESSION

    Field Memorandum: What Gets Bet On at the Frontier

    A study session on how top technical investors screen founders — the 'bet on the person, not the idea' thesis behind Paradigm's frontier investing, the six-signal screen (clarity, obsession, slope-over-intercept, agency, patience, mission-alignment), the research-as-investing model and 'X-Men Academy' talent approach, and how the same four criteria apply to choosing a co-founder. Drawn from the trajectories of Paradigm's Matt Huang and ByteDance's Zhang Yiming, with Ashley Smith's solo-GP fund as the counter-model.

    paradigmmatt-huangzhang-yimingbytedanceventure-capitalfounder-selectionfrontier-techconvictionresearch-as-investingcofounder
  • 2026.07.06 · STUDY-SESSION

    Bending Spoons: Reading a Real IPO

    A question-by-question study session on reading Bending Spoons' Nasdaq IPO — the 'operating machine' model behind its AOL/Vimeo acquisitions, the LBO/leverage mechanics, GAAP-vs-adjusted profit, dual-class control, tax benefits, and how to read a prospectus and cash-flow statement like an analyst.

    bending-spoonsipoprospectuscash-flowfinancial-statementssaasleverageroll-upvaluation
  • 2026.07.05 · TEARDOWN

    2026 Unicorn Tracker

    An interactive teardown of every 2026 unicorn — verified companies, the investor–founder relationship network, deep investor/founder profiles, the broader VC landscape, and a Q&A on how venture capital actually works.

    unicornsventure-capitalstartupsinvestorsfinance
  • 2026.06.30 · TEARDOWN

    Luckin Audit Learning

    A hands-on log of learning to read SEC filings and spot fraud red flags, walking the actual Luckin Coffee IPO prospectus page by page.

    luckinaccountingauditfraudfinance
  • 2026.06.27 · TEARDOWN

    Teardown: KAIKAKU

    Back-of-house restaurant automation as a SaaS bet. The teardown: 20–70× less capital than peers, a 360-vs-60 throughput gap, a brutal sector graveyard — capital-efficient founding tactics worth copying, low odds of success.

    ai-startupteardownroboticsfood-automation

Writing

ALL POSTS →

Blogroll

Writers and labs I follow.

Awards & Projects

  • 2024
    National Scholarship, Ministry of Education of China.
  • 2021
    National First Prize (Top 3), China College Student Computer Design Competition — Alpine Skiing VR, as team leader. Outstanding Work · showcased online · 20,000+ views.