CONNECTED RESEARCH

Knowledge map

Explore research through connected companies, technologies and questions. Every relationship links to its evidence.

55 / 55 articles with reviewed reader guides · 202 reviewed relationships

Compile by training

Compile by trainingCompile by trainingTeacher modelsTeacher modelsReusable local neural functionsReusable local neural functiReusable neural functionsReusable neural functions

Select a connected node to continue exploring. Open the evidence below to distinguish reporting, analysis and hypotheses.

Reported evidence

Compile by training → uses → Teacher models

Teacher models generate task examples during compilation; those examples are then used to train small adapters for a compact interpreter. This is a compilation-stage role, not a claim that teachers are called during execution.

Read the evidence
In compile by training, teacher models generate task examples during the compilation stage, after which small adapters are trained for a compact interpreter.

[S4] arXiv

Read the full research

Reported evidence

Compile by training → enables → Reusable local neural functions

The paper's authors describe and demonstrate conversion of natural-language task specifications into local neural functions that can be saved, versioned, and composed, and that execute without teacher models. This does not by itself establish lower total costs or suitability for other tasks.

Read the evidence
The paper abstract states that the generated functions no longer use teacher models at runtime and can be saved, versioned, and composed. [S4] The authors introduce and demonstrate a method for converting natural-language task specifications into reusable local neural functions.

[S4] arXiv

Read the full research

Reported evidence

Compile by training → enables → Reusable neural functions

Compile by training uses teacher-generated task examples to train small adapters for a compact interpreter, producing neural functions that can run without the teachers and can be stored, versioned and composed. This mechanism does not specify the final deployment location.

Read the evidence
Compile by training provides an explicit mechanism: teacher models first generate task examples from natural-language specifications, then train small adapters for use by a compact interpreter; once completed, the resulting neural functions can run without teacher models [S1]. The products of this method can be stored, versioned and composed [S1].

[S1] arXiv

Read the full research

Connected research

ai

AWS, Baidu, and a Compile-by-Training Case Show How AI Tasks Are Organized in the Cloud and Reused Locally

What do the AWS ordering-assistant tutorial, Baidu Qianfan page, and compile-by-training abstract actually establish about cloud task organization and reusable local neural functions?

The article separates demonstrated mechanisms from product-page listings: AWS describes cross-channel memory and separate channel and ordering layers; compile by training describes teacher-generated examples and reusable functions that run without teachers; Baidu presents Agent, workflow, deployment, and model entry points without establishing their technical integration.

ai

Qwen and Tencent Showcase Translation, Agent, and On-Device Deployment Approaches; Shared-Endpoint Audit Falls Short of Repeatability Thresholds

How should readers compare Qwen-MT, Hy-MT2 and reusable neural functions, and what does the shared-endpoint audit imply for repeatable model evaluation?

The article distinguishes language coverage, mobile deployment and teacher-independent execution rather than treating them as one technical route. It also uses a preregistered shared-endpoint audit to explain why request times, endpoints, replay conditions and repeated results belong in model comparisons.