WeHub Research studies how causal intelligence can build on foundation-model representations and be tested inside real human-AI collaboration systems.
For causal researchers, AI system builders, and collaborators working on mechanisms, interventions, counterfactuals, and robust action.
We do not package exploration as conclusion. Each line states the problem it addresses, its current stage, and where the underlying work can be inspected.
Addresses a foundational gap in structural causal models by revisiting how units, noise realizations, and cross-world connections are represented.
Core question: how can different counterfactual worlds stay connected while unit identity remains stable?
Open working note (Chinese)Uses robust prediction as a verifiable wedge for testing whether causal structure improves stability across environments.
Core question: which mechanism information improves generalization instead of merely fitting a training distribution?
Open working note (Chinese)Counterfactual generative modeling for heterogeneous units, connecting unit identity, cross-world structure, and runnable prediction tasks.
The current entry brings together the paper direction, reusable causal toys, and source/projection boundaries.
Open HCGM entry (Chinese)A longer-term ambition: use language-model representations for mechanism understanding, intervention reasoning, and robust action—not correlation fitting alone.
This remains a directional research object, not a completed or published result.
Theory, models, and real systems constrain each other: start with a foundational problem, build representations and tests, then return to feedback from human-AI collaboration.
Find gaps in causal foundations and make unit, environment, intervention, and counterfactual assumptions explicit.
Use foundation and generative models to turn causal assumptions into computable, comparable objects.
Test whether models actually help researchers make more robust judgments and actions in prediction and agent-harness settings.
These public articles cover agent runtimes, skill portability, and how human attention can connect to an AI collaboration network. The linked articles are currently in Chinese.
The real distinction is not whether memory exists, but how each runtime organizes identity, action, context, memory, and boundaries.
A fact-first map of native skills, host adapters, runtime assets, and workflow contracts—without mistaking copying for native support.
A real workday showing how limited human attention can connect to an AI collaboration network—and why token use is a system vital sign, not a vanity metric.
WeHub Research is the open research window of Gong Heyang and WeHub. It keeps real questions, unfinished judgments, and research ambition visible while distinguishing active research, working notes, and exploratory directions.
If you work on causal intelligence, robust generalization, agent harnesses, or human-AI collaboration, contact us with a concrete question, evidence, or collaboration idea.