DNA News

Brief, mostly signal

METR proposes an alignment assessment for unreleased models

Reviews serious misaligned behavior in near-frontier internal models. It needs internal transcripts and checkpoint sampling, so it is policy reading.

Read the source at metr.org

What the research found

METR proposes an alignment assessment method for unreleased models that reviews serious misaligned behavior in near-frontier internal models, requiring internal transcripts and sampling access to checkpoints. Background policy and research reading, no direct engineer action.

Claims and how well the evidence supports them

  1. The METR proposal assesses alignment of near-frontier internal models using internal transcripts 91%
  2. The assessment requires sampling access to internal model checkpoints 98%
  3. Evaluation awareness limits how much assessors can rely on alignment evals 98%

Evidence

Actionability rated low.