AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Chiral Symmetry Breaking in Transformers: A Group-Equivariant Framework for Addressing the Reversal Curse via Adjoint Manifold Mappings

Hanji Du

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Differentiable optimization layers

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add AIR structured retrieval readout (vs sequential decoding) (ablation)A1 · Wiki-Inverse-10Kinverse accuracy (%): 10.85 -> 61.31Improved
add chiral regularization R_chi (ablation)A1 · Wiki-Inverse-10Kinverse accuracy (%): 7.45 -> 61.31Improved
use Householder involutive operator instead of a linear layer (ablation)A1 · Wiki-Inverse-10Kinverse accuracy (%): 1.02 -> 61.31Improved
use Householder reflection instead of orthogonal map in AIR (ablation)A1 · Wiki-Inverse-10Kinverse accuracy (%): 32.14 -> 61.31Improved
use Householder reflection instead of unconstrained linear map in AIR (ablation)A1 · Wiki-Inverse-10Kinverse accuracy (%): 18.45 -> 61.31Improved
use Householder reflection instead of identity map in AIR (ablation)A1 · Wiki-Inverse-10Kinverse accuracy (%): 8.12 -> 61.31Improved
use learned adjoint operator vs plain kNN retrieval (ablation)A1 · Wiki-Inverse-10Kinverse accuracy (%): 0.95 -> 61.31Improved
use learned operator instead of external retrievers (Contriever, DPR) in AIR (ablation)A1 · Wiki-Inverse-10Kinverse accuracy (%): 14.28 -> 61.31Improved
A1 vs bidirectional training (BDP)A1 · Fact-Inv-300inverse accuracy (%): 41.13 -> 65.07Improved
A1 vs data augmentation (DA)A1 · Fact-Inv-300inverse accuracy (%): 32.47 -> 65.07Improved
A1 vs standard Llama-3 readout (forward accuracy)A1 · Fact-Inv-300inverse accuracy (%): 88.47 -> 82.13Hurt
scale standard Transformer from 125M to 8B parametersStandard Transformer · Wiki-Inverse-10K scalinginverse accuracy (%): < 2%No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy