Research / Model behaviour
Polinko
An ongoing inquiry into instructions, context, correction and reliable interaction with language models.
- Direction
- Krystian Fernando
- Scope
- Research / Model behaviour
- Stage
- Ongoing research
01 / The story
Stay with the question.
Polinko begins with questions about the relationship between an instruction and what a model actually does. A response can sound coherent while drifting from the work that was asked for.
The inquiry follows how instructions, accumulated context and corrections shape behaviour, while keeping the source of each observation available.
02 / The story
Give a judgement a visible boundary.
A binary gate asks a defined question and records pass or fail. The response, human judgement and surrounding context remain part of the evidence.
The toys provide deliberately narrow situations in which to examine those relationships. Handwriting transcription supplies another case where the expected answer can be checked against a source.
03 / The story
Let a finding leave the next question open.
A passing measure can expose a different weakness. Retained failures help identify which distinction the next experiment needs to make.
Questions about operator effort and repeated correction also connect the research to sustainability. Those remain research questions, with claims bounded by the evidence actually recorded.
Contribution
Research questions, method, evaluation criteria, source interpretation and acceptance; implementation and supporting tooling developed with Codex.
Ongoing research
Explore the existing research ↗Next project
Toolinko / Tooligan
Tools that grow from the work.