Symmetry matching improves sparse autoencoders for safer language model decisions
A Symmetry-Matching Approach to Blind-Spot Reduction in Sparse Autoencoders
Machine Learning
Summary
Sometimes language models don’t notice important differences between inputs, which can cause problems in sensitive areas like medicine or law. The authors identified that traditional sparse autoencoders don’t fix this problem because they don’t focus on how features respond to certain changes. They introduced a new training method that encourages these models to pay more balanced attention to specific types of changes, improving how the model distinguishes tricky inputs without making the overall understanding worse. Their tests showed this approach helped models better handle even subtle or unexpected input differences, important for real-world safety.
What this means in practice
- •For clinical data teams: Improve automated extraction of medical dosages by reducing silent errors in language model representations.
- •For financial compliance teams: Enhance reliability of models verifying financial amounts to catch subtle input differences that may cause errors.
Authors
Csaba Balogh
Abstract
Language models can treat semantically distinct inputs as interchangeable at the representation level, creating blind spots that standard sparse autoencoder (SAE) objectives do not target. In safety-critical settings such as clinical dosage extraction, legal clause interpretation, or financial amount verification, such failures can propagate silently into downstream decisions. We study this problem as one of feature-basis orientation. Building on a symmetry-matching view from algebraic error-detection theory, we add a differentiable $V_{\mathrm{Gini}}$ regularisation term to SAE training that penalises uneven perturbation response within a specified family. Across GPT-2, Gemma 2, and Qwen 2.5, a frozen multi-seed evaluation shows that $V$-regularisation raises both relative and absolute lower-tail responses while reducing response-profile imbalance, with no material general-text reconstruction degradation. JumpReLU and minimum description length (MDL) objectives remain close to the Standard SAE profile under the same protocol. Family-geometry controls show that the effect is not explained by template duplication or a single dominant perturbation family, and held-out probes plus null-calibrated separation tests evaluate how far the representation-level change carries to downstream discriminability. A zero-shot audit on out-of-distribution clinical radiology minimal pairs finds the relative weak-tail lift in all evaluated model-by-family cells, not perfectly rank-aligned with the trained statistic, and a held-out family audit finds it on family kinds never used in training. Together, the results show that perturbation-aware SAE training can reorient feature bases toward safety-relevant input changes while leaving downstream validation as a separate empirical step.