Repository navigation
v1.3.3 (2026-07-16)
This release is published under the MIT License License.
Bug Fixes
The AuxK loss target e = y - sae_out already accounts for b_dec, since sae_out is produced by decode() which adds b_dec. Computing e_hat via self.decode() added b_dec a second time, pulling dead latents toward (e - b_dec) and placing an unintended gradient on b_dec whenever b_dec != 0 (the normal case after init centers it on the data mean).
Call decoder_impl directly so e_hat does not re-add b_dec.
Detailed Changes: v1.3.2...v1.3.3