MECHANISM PRESSURE
From prediction to mechanism
Ask not only what will happen, but why, what intervention changes it, and what this same unit would experience in another world.
Explore HCGM / DiscoSCM →A research system that grows—and lets evidence correct it
Intelligence that understands causes, acts and keeps evolving.
Our pursuit is not one model. It is a class of intelligent systems that learn mechanisms from heterogeneous units and data, test understanding through interventions and counterfactuals, and absorb real feedback through human-agent collaboration.
TabU is a wedge, not the whole map. Through unit-first tabular learning, it asks how shared laws hold across heterogeneous individuals.
North starCausal
Intelligenceunderstand · intervene · infer
Intelligenceunderstand · intervene · infer
01World modelsmechanisms & possible worlds
02Unit learningheterogeneity & shared laws
03Agent societycollaboration & feedback
Research → Wedge → Evidence → Update
Questions before projectsEvidence before claimsFeedback drives growth
01 · RESEARCH DIRECTION
We are not collecting projects.
We keep closing in on one problem.
How can intelligence move beyond correlation and understand generative mechanisms? How can it account for different units, contexts, and possible worlds? And how can that understanding enter action, then update under real feedback?
HETEROGENEITY PRESSURE
From average individuals to units
Treat Unit as a semantic primitive and study how shared response laws express individual differences through context, tokens, and abduction.
Explore TabU-lab →REALITY PRESSURE
From offline models to collaborative systems
Turn real human-agent work, evaluation, and correction into a feedback field, so research is not merely self-consistent on benchmarks.
Explore Agent Society →02 · RESEARCH MAP
One long-term pursuit, three research fronts that pressure each other.
These are not isolated departments. World models supply a language of mechanisms; unit-centered learning forces theory to face heterogeneity; human-agent systems return research to real collaboration and feedback.
01
CORE PROGRAM
Causal intelligence & world models
Represent mechanisms, interventions, counterfactuals, and possible worlds—and make abduction learnable and auditable.
02
LEARNING WEDGE
Unit-centered learning
Start from units, context, and shared response laws to explore new primitives for tabular foundation models, structured learning, and heterogeneous data.
03
FEEDBACK FIELD
Human-agent systems
Study personal avatars, agent societies, memory, and collaboration structures as a living field where research hypotheses meet reality.
Project names change and research questions grow. This is the current relationship map, not a closed taxonomy.
03 · OPEN WORK
Enter through questions, not slogans.
UNIT / CONTEXT-SPECIFIC MECHANISMS
HCGM × DiscoSCM
Treat unit-specific mechanisms as primary learning objects and revisit the relationship among counterfactuals, heterogeneity, and population-level models.
ROBUST PREDICTION WEDGE
Causal Regression
A concrete robust-prediction wedge that tests ideas from mechanism learning and abduction.
TABULAR FOUNDATION MODELING
TabU-lab
Open questions, gates, run receipts, and verdicts. There are currently no public training results, checkpoints, or benchmark claims.
UNDERSTANDING DATA-GENERATING PROCESSES
Causality Primer
A Chinese-first book, question bank, cases, and feedback system that turns causal concepts into a knowledge spine that can be challenged and revised.
04 · HOW RESEARCH GROWS
A website is not the result.
It is an entrance to the evidence loop.
A polished page cannot turn a proposal into a theory or a smoke run into a benchmark conclusion. We keep the boundaries between stages explicit.
Learn about MSP / Seed it till it grows →- 01Question
State what we actually want to know and what observation would change our judgment.
- 02Wedge
Choose the smallest topologically complete theoretical, modeling, experimental, or writing entry.
- 03Evidence
Record sources, assumptions, configurations, failures, run receipts, and claim boundaries.
- 04Feedback
Let people, agents, readers, and real systems apply pressure to the artifact.
- 05Update
Keep what survives, correct the question, and enter the next cycle.
RESEARCH NOTES
Read how we form questions—not only how we package answers.
The Research Blog preserves developing arguments, architectures, reflections, and source trails. It is a living notebook, not a substitute for peer-reviewed papers.
Enter the Research Blog ↗WEHUB RESEARCH
Research informs the system.
The system returns reality.
Causal intelligence supplies long-term questions; human-agent systems supply real feedback. The loop between them is the research infrastructure WeHub is building.