Transferring what was learned between unequal models
What should be transferred between models with different capacities so that acquired knowledge remains useful?
Text, an agent-oriented language, coarse conceptual tokens, and adaptive parameters are considered as possible transfer surfaces. Shared token names or adapter shapes alone do not settle semantic compatibility.
Status: open problem. Formulation and comparisons: reconstruction. Development: active.
Alternative representations
Explanatory text
Affords: Broad interpretability and inspectable claims.
Limits or losses: Can omit tacit distinctions encoded in learned behavior.
Structured conceptual transfer
Affords: Can preserve task-specific relations compactly.
Limits or losses: Needs a common interpretation or an explicit translator.
Parameter-level transfer
Affords: Offers a direct intervention into computation.
Limits or losses: Requires compatible parameter meanings and integration checks.
Tension
Communicating a statement and transferring a competence are different success criteria.
Challenge
How can the receiving model demonstrate the transferred capability and identify what it did not receive?
This challenge is an editorial prompt, not a claim that the evidence already resolved it.
Proposed next move
Choose a narrow skill and compare transfer formats under equal resource budgets and independent held-out tasks.
Status: new model-proposed experiment or extension; not a reported result.
Related problems
- Changing the grain of a conceptual token — proposes coarse-to-fine transfer (corpus_evidence). A proposed small-to-large learning transfer explicitly uses coarse units that can later be broken down.
- Distributed specialization and integration — enables a proposed collaboration (corpus_evidence). Distributed specialists require a way to transfer what was learned into another system.
- When transferred learning affects governing policies — questions what a transfer carries (corpus_evidence). A transfer may be proposed as knowledge exchange while potentially affecting governing commitments. This remains a proposed possibility.
Evidence anchors
- basis-cross-model-learning-transfer: The proposal has smaller and larger models exchange what they learn through text or a different agent-oriented language. It also considers coarse units that are later decomposed, and explicitly links communication at the parameter level with possible changes in governing policies. [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
- An Evolving Semantic Representation Language (reconstruction).
- Memory, Tokens, and Learning Geometry (reconstruction).