Learning to propose parameter changes
Can a model propose useful changes to another model from structural and conceptual information?
The proposal ranges from interpreting hidden layers to predicting operations and choosing updates. A concrete suggested test is adding a new token and proposing its parameters from a conceptual description.
Status: open problem. Formulation and comparisons: reconstruction. Development: active.
Alternative representations
Inspection of a token trajectory through layers
Affords: Connects intermediate states with an output.
Limits or losses: An activation trajectory does not identify every relevant parameter or establish a causal explanation.
A proposed weight update
Affords: Makes an intervention that can be measured.
Limits or losses: Can improve a local example while degrading unrelated behavior.
Tension
Understanding, predicting, and modifying a model are distinct capabilities.
Challenge
What evidence shows that a proposed edit generalizes beyond the supplied concept description?
This challenge is an editorial prompt, not a claim that the evidence already resolved it.
Proposed next move
Test the proposed new-token edit against ordinary initialization and training, with held-out uses and regression checks.
Status: new model-proposed experiment or extension; not a reported result.
Related problems
- A system that can revise its own representation language — raises implementation questions (retrospective_connection). A changing representational language invites questions about corresponding parameter changes, but the mechanisms are not established as identical.
- Memory represented through adaptable parameters — shares a proposed update surface (retrospective_connection). Memory adaptation and predicted model edits may use adaptable parameters; they need different tests.
- When transferred learning affects governing policies — adds a policy-level test (model_proposed). A task-directed edit should be investigated for changes beyond the task. This test is an editorial extension.
Evidence anchors
- basis-models-editing-models: The material proposes models that interpret hidden layers, predict operations, and alter parameters, including during their own operation. One suggested test adds a new token and asks a model to set its parameters from a conceptual description of that token. [Editorial paraphrase; presence of a proposal does not verify its truth.]
Root evidence and private recovery mappings are not part of this public release.
Open in the reader · All problems
Research editions
- Memory, Tokens, and Learning Geometry (reconstruction).