Skip to content
Permalink

Comparing changes

Choose two branches to see what’s changed or to start a new pull request. If you need to, you can also or learn more about diff comparisons.

Open a pull request

Create a new pull request by comparing changes across two branches. If you need to, you can also . Learn more about diff comparisons here.
base repository: VectorArc/avp-python
Failed to load repositories. Confirm that selected base ref is valid, then try again.
Loading
base: main
Choose a base ref
...
head repository: VectorArc/avp-python
Failed to load repositories. Confirm that selected head ref is valid, then try again.
Loading
compare: cross_model_research
Choose a head ref
Checking mergeability… Don’t worry, you can still create the pull request.
  • 2 commits
  • 3 files changed
  • 2 contributors

Commits on Mar 19, 2026

  1. Add activation-space norm fix for mid-layer cross-model injection 

    Previously mid-layer injection projected to embedding-space norm (~20-40)
but injected at layer 21/28 where activation norms are ~500-2000. This
50-100x norm mismatch caused the model to barely notice the injection,
explaining the 45% GSM8K result (vs 77% for layer-0 rosetta).

Fix: compute target model's activation norm at the injection layer via
calibration prompts, then renormalize the projected vector to match.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
    @SStas @claude
    SStas and claude committed Mar 19, 2026
    Configuration menu
    Copy the full SHA
    6c25d80 View commit details
    Browse the repository at this point in the history

  2. Add multi-vector mid-layer injection support 

    Extract N evenly-spaced hidden states from latent steps, project each through
vocabulary, prepend N pad tokens as placeholders, inject all N at mid-layer.
Tests whether multi-vector through vocabulary works better at mid-layer than
at layer 0 (where it was validated negative #2).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
    @SStas @claude
    SStas and claude committed Mar 19, 2026
    Configuration menu
    Copy the full SHA
    04bbcd8 View commit details
    Browse the repository at this point in the history
Loading