GENESIS繁中 ⌄

GENESIS / RESEARCH

Small models.
Greater possibilities.

How much more could a small model do? We study task understanding, anticipation and learning within practical limits on memory and computing power.

EVO Alpha is the compact language-model research. World Model explores environment understanding that a small core could use; RSI studies how that core and its supporting methods could improve. These are independently developed research directions, not a demonstrated integrated system or a claim of parity with large cloud models.

Ongoing research

01 / Compact language models

EVO Alpha LLM

Smaller models. More complete work.

A useful local model would carry something forward from each task: a rule it can reuse, or an assumption it needs to revise. EVO Alpha explores how to make that possible within the memory and computing power of a small system.

EVO Alpha LLM research concept
Research concept · illustrative

THE RESEARCH SO FAR

What should a model carry forward?

Our early experiments brought one tension into focus: adapting to new examples can also disturb something a model already gets right. That makes knowing when to change as important as learning something new.

We are investigating how useful knowledge could survive those changes and help with unfamiliar tasks. The goal is local intelligence that can learn within a practical computing budget.

THE NEED

Work that changes along the way

Consider a task whose requirements change halfway through. We want a compact model to recognise what still matters, what needs to be checked and whether the work is actually complete. This is a research use case, not a capability available in the apps today.

TECHNICAL DIRECTION

A small core, with useful knowledge around it

The research distinguishes general ways of working from facts that change. Understanding a goal, using a tool and noticing uncertainty should travel between tasks; current information can come from sources the system can consult. We evaluate the model, its supporting tools and the complete task separately, so a tool’s success is not mistaken for something the model has learned.

WHAT WOULD MATTER

Useful learning must survive a new task

A better answer on a familiar example is only a starting point. We look for improvements on unfamiliar work, retention of earlier abilities and a practical cost in memory and response time. A local model would be valuable if it could grow more capable within the resources of the device people already use.

Studies, technical details & scope ＋
Studies so far
Local model experiments & formal analysis

A small recurrent-model experiment showed adaptation on a narrow set of synthetic tasks, alongside regressions on cases requiring no change. Later local-model adaptation did not improve overall accuracy over its frozen baseline.

  • Completed a limited-experience adaptation pilot; its gains and regressions did not qualify the candidate for confirmation.
  • A Qwen-based local adapter changed output preferences but did not establish a transferable accuracy gain.
  • Formal analysis clarified that better ranking signals need not produce more correct answers, sharpening the choice of comparison methods.
Research focus
Learning reusable knowledge under constraints on model size, memory and computation. 3B is an initial deployment milestone; Qwen is a current baseline and experimental vehicle, not a fixed final architecture.
Evidence boundary
The results are bounded development experiments and analyses. Stable transfer and retention on independent new tasks, general cognition, and integration of these research candidates into iPhone, World Model or RSI have not been demonstrated.
Next question
Can experience from different task families improve decisions on unfamiliar combinations, compared with practical baselines given the same information and resources?

02 / Environment & action

World Model

Before an action, understand what may follow.

Before choosing an action, it helps to have a sense of what might happen next. World Model studies how a system could build that understanding from observations and experience, then revise it as its surroundings change.

World Model research concept
Research concept · illustrative

THE RESEARCH SO FAR

Which observations help us choose?

In a small simulated environment with defined rules, accumulated observations helped predict several steps ahead. This led us to a more practical question: which observations actually help a system choose its next action?

Looking for more information takes time and resources. We study how a system could recognise what it still needs to know, when past experience applies, and when a changed environment calls for a fresh look.

THE NEED

Before committing to an action

Imagine a workflow with an unavailable resource or an uncertain tool state. Trying an action may consume time or change something that is hard to undo. Our research asks whether a system can anticipate plausible outcomes and identify which missing observation would make the choice better informed. This illustrates the goal; it is not a deployed workflow.

TECHNICAL DIRECTION

An understanding that can be revised

World Model is being developed as an independent environment-understanding component that a compact language model could use. Its questions concern what state the environment may be in, how different actions could change it and how fresh observations should revise that expectation. Remembered experience is useful only when the relationships behind it still apply.

WHAT WOULD MATTER

Better forecasts should help people choose

We assess prediction and decision quality separately. More accurate forecasts may still be unhelpful if obtaining them costs too much, or if they remain confident after the rules change. Recent analysis examines how unreliable feedback can make a lasting pattern difficult to distinguish from a temporary coincidence. The practical aim is to know when experience helps—and when to look again.

Studies, technical details & scope ＋
Studies so far
Controlled prediction studies & transfer analysis

Experience improved multi-step prediction within a declared, finite rule family. Whether that improvement leads to better decisions across different environments remains an open question.

  • Completed a controlled study of prediction before and after observations; the gain is limited to the tested rule family.
  • Decision comparisons have not established a stable task-success gain. The cost of gathering information can outweigh its benefit.
  • Transfer analysis exposed confident errors when a target environment breaks the assumed rules. Cross-environment validation is still ahead.
Research focus
An independently callable environment-prediction module that updates from experience. Research separates what can be inferred about state, what can be predicted, and whether either is useful for a decision.
Evidence boundary
Evidence comes from bounded synthetic tasks and finite analyses. General causal understanding, transfer across independent environments, continual learning and real Alpha-model integration have not been demonstrated.
Next question
Can paired experience improve unseen action predictions in an environment with genuinely different rules, against baselines with the same information?

03 / Recursive improvement

RSI

Improve the method, not only the answer.

A difficult task can reveal something worth investigating about how the work was done. RSI explores how a system could turn that experience into a research question, try a better method and pass useful changes on to its next version.

RSI research concept
Research concept · illustrative

THE RESEARCH SO FAR

When is a change worth keeping?

We began by examining how changes could be compared, checked and carried into later work. The early pilots showed why the comparison matters: an apparent improvement can lose its value when measured against a stronger existing method.

That has sharpened the research question. Can a system identify a worthwhile problem, develop a better approach and show that it helps on fresh tasks? We want each round of research to give the next round something useful to build on.

THE NEED

When the same difficulty keeps returning

A recurring failure may point to a weakness in how a system plans, learns or checks its work. RSI asks whether the system could identify a worthwhile research question, investigate an alternative and establish whether the change helps. The ambition includes improving how research itself is carried out.

TECHNICAL DIRECTION

Changes that earn their place

For a compact intelligence system, the research considers improvements to reasoning, memory, learning and compatible model updates. Each proposal needs evidence beyond its own explanation. We study how to compare alternatives, preserve useful changes and recover when a change causes regressions. These are research and engineering directions, not proof of an autonomously improving system.

WHAT WOULD MATTER

The next round is the real question

For recursive improvement, one successful update is not enough. A later version would need to use what changed, perform better on fresh tasks and become more effective at finding the next useful improvement under comparable resources. This is the distinction our research aims to establish; sustained recursive gains have not yet been demonstrated.

Studies, technical details & scope ＋
Studies so far
Software pilots & bounded model studies

Early synthetic pilots explored how changes could be evaluated and inherited. Promising results against simple baselines did not reach the preset threshold for useful improvement against a stronger comparison.

  • Built and exercised change evaluation, version history and recovery mechanisms in controlled software pilots.
  • A local adaptation study using a non-Alpha research model did not outperform its baseline.
  • Recursive pilots have not demonstrated autonomous model-generated improvements or sustained gains. Alpha integration remains unverified.
Research focus
Independent evaluation, traceable changes and inheritance across versions. A later round must actually use the change and earn its benefit on tasks not used to select it.
Evidence boundary
Software checks and non-Alpha model studies do not establish autonomous research, useful weight learning, cross-task generalisation or recursive gains. Formal Alpha integration has not been validated.
Next question
Within a fixed research budget, can the system choose a useful question, propose a change and outperform a fixed strategy on independently assessed new tasks?

RESEARCH / QUESTIONS

A closer look.

What does Genesis research? ＋

Genesis researches compact language models, world models and recursive improvement. These are separate ongoing research directions alongside its software products.

Are these research models available to download? ＋

No. The research projects are not available for public download. The page presents research goals, bounded findings and technical overviews.

Is World Model a video generator? ＋

The Genesis World Model direction studies environments, states and the consequences of actions. Its focus is prediction for decisions, rather than a public video-generation product.

Has recursive self-improvement been demonstrated? ＋

Not by the evidence reported here. Current RSI evidence includes software pilots and bounded studies with non-Alpha research models. Autonomous and sustained improvement remain unverified.

Research direction and status reviewed 15 October 2026.