cs.CLNov 14, 2026

When Does Label-Free Deconfounding of Document Embeddings Work?

Authors: Anna Kiepura, Yingqiang Gao, Elliott Ash, Yu Fan

Organizations: ETH Zurich · Department of Computational Linguistics, University of Zurich · Department of Computational Linguistics University of Zurich · The University of Hong Kong

Abstract

Document embeddings encode spurious attributes that can distort semantic comparisons; yet supervised erasure requires metadata that is often unavailable. We systematically investigate label-free deconfounding across corpora and embedding models. Our results show that label-free methods improve semantic utility over unprocessed embeddings, while simple geometric approaches recover much of the supervised benefit, sometimes matching metadata-supervised erasure in linear confounder recoverability and semantic utility. We identify two distinct causes of ineffective erasure: poor proxy-confounder alignment and insufficient intervention coverage. Crucially, confounder recoverability and semantic utility are distinct dimensions of deconfounding success: utility can improve despite residual confounder information, while stronger suppression does not necessarily improve utility. Moreover, reduced linear recoverability can coexist with substantial nonlinear recoverability. Together, these findings distinguish confounder discovery from intervention effectiveness and highlight the importance of evaluating nuisance suppression alongside semantic preservation.

Figures & tables

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Apr 19, 2026

FLARE: Task-agnostic embedding model evaluation through a normalization process

When task-specific labels are not available, it becomes difficult to select an embedding model for a specific target corpus. Existing labelless measures based on kernel estimators or Gaussian mixes fail in high-dimensional space, resulting in unstable rankings. We propose a flow-based labelless representation embedding evaluation (FLARE), which utilizes normalized streams to estimate information sufficiency directly from log-likelihood and avoid distance-based density estimation. We give a finite sample boundary, indicating that the estimation error depends on the intrinsic dimension of the data manifold rather than the original embedding dimension. On 11 datasets and 8 embedders, FLARE reached Spearman's ρρ of 0.90 under the supervised benchmark and remained stable in high-dimensional embeddings (d≥3,584d \geq 3{,}584) as the existing labelless baseline collapsed.
Oct 23, 2026

Principled Nonlinear Concept Erasure and Steering

Learned representations often encode sensitive attributes, such as gender or race, that may be undesirable or unsafe to retain. Concept erasure aims to remove this information while preserving task-relevant content. Existing methods remain incomplete: linear methods provide guarantees only against linear probes, while nonlinear methods achieve stronger empirical erasure but lack formal guarantees and often sacrifice utility. In this work, we introduce a principled conditional autoencoder framework for nonlinear concept erasure. We follow the principle that erasure should remove only the sensitive attribute and nothing else, and we show that, under the stated assumptions, perfect reconstruction together with population-level independence is sufficient for perfect nonlinear guardedness, with preservation guarantees in the imperfect-reconstruction regime. Empirically, we attain a superior utility-erasure trade-off across multiple datasets, sensitive attributes, and language models, surpassing even the supervised variants of competing methods despite using no target-task labels, while being simpler to optimize and converging substantially faster. Beyond erasure, our framework yields concept conversion (representation steering) as a by-product: by changing only the target concept at inference time, it produces representations corresponding to different concept values - at no additional training cost - a capability no prior erasure method provides, and on which we likewise demonstrate strong empirical performance.
Sep 23, 2026

Hidden not Deleted: How Networks Suppress Entangled Features

When a neural network stops reporting a feature, has it forgotten the feature or merely hidden it? Shared representations complicate the answer: removing one feature's direction can also damage another feature. We study this problem in a small autoencoder that compresses and reconstructs inputs, allowing individual feature contributions to be traced directly. With the representation fixed, restoring a single bias partially recovers target detection by exposing a signal hidden behind the output's activation threshold. Tests on target-absent inputs distinguish this recovery from interventions that simply increase output. The same bias intervention becomes less effective when the representation is allowed to change. We find that training only the other outputs can weaken the target's contribution, even without a loss on its own output. A conditional theoretical analysis explains how interference with retained predictions creates this pressure. Yet more flexible predictors still detect residual target information. These findings show how recovery depends on feature geometry and on which parameters are trained, and why a silent output is insufficient evidence of forgetting.