Work

Coordinate or Concept Mismatch?

Investigating cross-model probe-transfer failures and task-conditioned alignment

Status
Ongoing Research
My role
Researcher
Period
August 2026 - Present

Overview

Investigating whether cross-model probe-transfer failures arise from coordinate mismatch or task-conditioned differences in representations.

The project examines how cross-model alignment behaves on held-out tasks, with the aim of separating improvements in coordinate matching from evidence that a learned representation generalizes.

Methods and Tools

  • Activation Probes
  • Representation Alignment
  • Cross-model Evaluation
  • Held-out Tasks