The story of our name: Astr · Mira · Mirage
Our name brings together the sky, a variable star, and an imagined horizon. They express what we value in research: the curiosity to ask distant questions and the patience to look closely, observe again, and test an idea.
Together, these images become 幻梦星芒. Wonder and rigorous inquiry are both part of the research culture we want to build.
How we approach research
For a question worth pursuing, we ask what basic relationships it involves, how far the existing explanation reaches, and what would lead to a more reliable understanding.
FIRST PRINCIPLES
Begin with the fundamentals.
Make the question, assumptions, and constraints explicit before choosing a method. Why does a retrieval path work? What survives a compressed representation? How can an agent reliably advance a long task? Each question leads back to relationships between data, computation, and decisions. Understanding the mechanism helps us judge when a method is useful.
EVIDENCE & REVISION
Let evidence change the judgment.
Derivations make reasoning explicit; experiments test its predictions. We value comparisons, replication, and cases where a method fails. An unexpected result is a clue to the problem. When observations disagree with a hypothesis, we return to that hypothesis and work toward a more accurate explanation.
THEORY & PRACTICE
Move between theory and engineering.
A real system exposes a fuller set of constraints: resources, latency, state, maintenance, and changes during sustained use. Those engineering details can raise new theoretical questions. Working in both directions helps us understand an idea more deeply and build a more dependable system.
LONG-TERM CURIOSITY
Give lasting questions time.
Progress can come from a new idea or a closer look at an old problem. We value work that others can build on: a clear derivation, experiments that can be examined, and an implementation that remains useful. Each step gives future exploration firmer ground.
Let research and systems inform each other.
Our research directions develop with the questions. Current interests include databases and retrieval, vector representations, and model quantization, alongside foundational questions in post-training and mathematics. Papers record findings along the way; systems give ideas a longer encounter with practical use.
This approach also shapes the open-source projects we are building: from organizing data, to connecting memory with action, to a shared space where people and agents work together.
Bring different experience to a shared question.
Foundational research, engineering, and real-world use each offer a different view. We welcome conversations with researchers, university teams, developers, and industry partners to define questions more precisely and test methods under a wider range of conditions.
A collaboration can begin with a paper discussion, a replication, a systems experiment, or a concrete problem that has resisted a solution. With a clear question and shared expectations about contributions, we can build understanding together and make suitable results available for others to read, examine, and extend.