Architecture / 02
Reason over the work, in parts.
Recursive language models offer a way to examine more than fits comfortably into one prompt. Merlin combines that idea with an explicit process for evidence, actions, and completion.
On this page
What is a recursive language model?
In the research formulation, a recursive language model treats a large input as material in an external environment. It can inspect portions programmatically, decompose a question, and make further model calls over selected parts. The context window becomes a working surface rather than the only place the source material can live.
The idea is an inference strategy. It does not, by itself, imply a newly trained foundation model or a guarantee that a conclusion is correct. The research paper by Alex L. Zhang, Tim Kraska, and Omar Khattab provides the original definition and experiments. Its benchmark results should not be read as measurements of Merlin.
Why this matters for real work
Large tasks usually contain smaller questions. Which documents govern this decision? Which component owns this behavior? What evidence would contradict the current hypothesis? Decomposing the work gives each question a scope that can be inspected and checked.
The advantage is organizational as well as computational. A smaller task can have a clear purpose, dependencies, and a defined result. The parent process can then assess how those results contribute to the original request. More calls are not inherently better; useful decomposition must preserve context and avoid turning repetition into apparent progress.
SpecKit 008 · AC-008-002-14 · SpecKit 009 · AC-009-005-01
How Merlin applies the idea
Merlin puts bounded reasoning inside a broader runtime. Plans give the work structure. Evidence records connect observations to tasks. Context is assembled for the current step. Authorized actions are checked, and a final review compares the proposed result with what the run can support.
This connects the freedom to investigate with responsibility at the point of action and delivery. The value is in the complete process, not in recursion as a slogan. The core concepts explain each part without exposing the internal decision rules.
SpecKit 005 · AC-005-001-01 · SpecKit 007 · AC-007-005-01 · SpecKit 004 · AC-004-004-03
Why the controls matter
Recursion can multiply a mistaken assumption just as readily as a useful insight. A child’s confident summary is not fresh evidence, and a sequence of completed calls does not establish that the user’s goal was met.
Merlin therefore aims for near-deterministic process reliability around probabilistic reasoning: explicit scope, evidence attribution, constrained operations, and checks at meaningful boundaries. Some checks are deterministic; semantic judgments are not. This distinction is essential to evaluating the system honestly.
SpecKit 004 · AC-004-006-22 · AC-004-004-02