Issues

• [Bug Report] FactoredMatrix ellipsis indexing selects the wrong factor axes

  TransformerLensOrg/TransformerLens#1848 · Filter by author yuanwuyuan9 opened on Oct 2, 2026
  

• [Bug Report] get_full_resid_decomposition(mlp_input=True, apply_ln=True, project_output_onto=...) normalizes neuron rows with the ln1 scale instead of ln2

  TransformerLensOrg/TransformerLens#1846 · Filter by author Zhuoxi2000 opened on Oct 1, 2026
  2

  

• [Proposal] Backward Lens: support gated MLP gate/ up/ down gradient factors

  complexity-moderateenhancementTransformerBridge
  TransformerLensOrg/TransformerLens#1832 · Filter by author janmenjayap opened on Sep 27, 2026
  

• [Proposal] Sparse probing: optional groups argument so rows from one prompt can't straddle the split

  complexity-simpleenhancementhelp wantedTransformerBridge
  TransformerLensOrg/TransformerLens#1813 · Filter by author jlarson4 opened on Sep 25, 2026

• [Bug Report] _BLOCK_LIST_ATTRS hardcoded name list silently drops Raven's blocks from composition-score / head-label analysis

  bugcomplexity-moderateTransformerBridge
  TransformerLensOrg/TransformerLens#1791 · Filter by author LightWork666 opened on Sep 18, 2026
  2

  

• [Proposal] SVD Circuits: singular-vector decomposition of a head's QK/ OV into causally-validated subfunctions

  complexity-highenhancementTransformerBridge
  TransformerLensOrg/TransformerLens#1767 · Filter by author janmenjayap opened on Sep 10, 2026
  3

  

• [Proposal] Relevance Lens (R-lens): a RelP/ LRP-based transport-matrix estimator for Jacobian Lens (J-lens) fit, readout, and intervention

  complexity-highenhancementTransformerBridge
  TransformerLensOrg/TransformerLens#1755 · Filter by author janmenjayap opened on Sep 8, 2026
  6

  

• [Proposal] Attribution Patching + EAP/EAP-IG: linearized activation patching and faithful edge-level circuit discovery

  complexity-highenhancementTransformerBridge
  TransformerLensOrg/TransformerLens#1742 · Filter by author janmenjayap opened on Sep 4, 2026
  3

  

• [Bug Report] [macOS-arm64] Cached eager attention NaNs in transformers v5 — blocks bridge KV-cache generation

  discussionseen_by_maintainerstoolingwontfix
  TransformerLensOrg/TransformerLens#1322 · Filter by author jlarson4 opened on May 22, 2026
  5