RepoJournal
Interpretability Interpretability
15 wires and counting

$ follow Interpretability

Keep up with Interpretability in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-10-24
stories 9

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

J-space direction loss and Jacobian lens joint ablation fixed in v4.3.0

By RepoJournal · Filed · About Interpretability · Composed from the cited sources · methodology

TransformerLens v4.3.0 fixes two measurement bugs that were quietly corrupting interpretability results for anyone projecting into sparse J-space or ablating multiple token directions at once.

fix: preserve selected J-space directions across atom scales TransformerLensOrg/TransformerLens

by emerardd

The default float32 rank cutoff in torch.linalg.pinv depends on ambient dimension and raw atom scale, so NNLS could select a genuinely independent direction and still report it as zero: with orthogonal atoms of norm 1 and 3000 in 4096 dimensions and target coordinates [2, 1], the reported J-space component was [0, 1]. Any sparse J-space decomposition with atoms on wildly different scales was si...

fix(jacobian_lens): ablate token directions jointly (#1941) TransformerLensOrg/TransformerLens

by Zhe Wang

ablation_hooks removed each token's lens vector sequentially, h <- h - (h.v)v, which reintroduces components along earlier directions because lens vectors of related tokens are far from orthogonal. It now projects onto the orthogonal complement of their span, h <- h - Q Q^T h, so the ablated activation actually has zero projection onto all removed directions. If you used multi-token ablation to...

fix(bridge): discover non-standard block lists in _enumerate_blocks TransformerLensOrg/TransformerLens

by LightWork666

_enumerate_blocks() only walked a fixed tuple of attribute names ("blocks", "encoder_blocks", "L_blocks" and friends), so architectures registering block lists elsewhere enumerated zero blocks with no error. On a Raven bridge, blocks_with(), composition_layer_indices() and attn_head_labels returned empty results; they now discover the lists properly.

Release v4.3.0 TransformerLensOrg/TransformerLens

by jlarson4

v4.3.0 ships RelevanceLens.fit with provenance tracking, integration and docs, adds an svd_circuits demo notebook with a slow name-mover parity check, and fixes stateful generation fallback so sampling penalties are preserved. The two lens fixes above land in the same release.

Quick answers

What shipped in Interpretability on October 24, 2026?
TransformerLens v4.3.0 fixes two measurement bugs that were quietly corrupting interpretability results for anyone projecting into sparse J-space or ablating multiple token directions at once. In total, 4 commits, 4 pull requests, and 1 releases landed.
Who contributed to Interpretability on October 24, 2026?
4 developers shipped this update, including jlarson4, emerardd, LightWork666, and Zhe Wang.
What were the notable Interpretability updates?
fix: preserve selected J-space directions across atom scales, fix(jacobian_lens): ablate token directions jointly (#1941), and fix(bridge): discover non-standard block lists in _enumerate_blocks.