Exploring Sparse Autoencoders in Text-Based Causal Confounding Adjustment

• URL: http://arxiv.org/abs/2609.01322v1
• Date: Tue, 01 Sep 2026 14:40:23 GMT
• Title: Exploring Sparse Autoencoders in Text-Based Causal Confounding Adjustment
• Abstract summary: We propose a novel causal adjustment pipeline that iteratively selects a minimal set of SAE features via conditional independence tests.<n> SAE representations achieve better adjustments than alternative representations in standard semi-synthetic evaluations with binary confounders.<n>We also introduce a more realistic semi-synthetic evaluation that uses multi-label data as the unobserved confounders.
• Score: 5.696801894034388
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: In many settings, studying causal questions based on text data requires adjusting for confounding information within texts. Yet there is a tradeoff in constructing text representations for adjustment: they must be sufficiently large and/or dense to preserve the confounding variables necessary for unbiased effect estimation, but sufficiently small and/or sparse to satisfy finite-sample overlap and yield low-variance estimates. To address this tradeoff, we turn to sparse autoencoders (SAEs), and propose a novel causal adjustment pipeline that iteratively selects a minimal set of SAE features via conditional independence tests. We find that SAE representations achieve better adjustments (lower bias and and higher coverage) than alternative representations in standard semi-synthetic evaluations with binary confounders, and their interpretability offers opportunities for falsification. We also introduce a more realistic semi-synthetic evaluation that uses multi-label data as the unobserved confounders and find off-the-shelf adjustment methods require increased investigation for these more complex settings. Code: https://github.com/mianzg/sae-text-confounder

Related papers

• REAR: Test-time Preference Realignment through Reward Decomposition [68.09214603569744]
  We introduce a novel framework that models the task as a realignment problem, since the base model often fails to align with the stated preference.<n>We derive a REAlignment Reward (REAR) that selectively rescales the proportions of these two reward terms.<n>We show that REAR can be formulated as a linear combination of token-level policy log-probabilities, making it computationally efficient and easy to integrate with various TTS algorithms.
  arXiv  Detail & Related papers  (2026-06-29T14:17:53Z)
• What Signals Really Matter for Misinformation Tasks? Evaluating Fake-News Detection and Virality Prediction under Real-World Constraints [0.08496348835248901]
  We study two practical tasks regarding online misinformation: fake-news detection and virality prediction.<n>We show that textual content alone is a strong discriminator for fake-news detection.<n>We discuss implications for evaluation design and report constraints that realistically affect the field.
  arXiv  Detail & Related papers  (2025-12-02T09:24:16Z)
• Beyond Frequency: Scoring-Driven Debiasing for Object Detection via Blueprint-Prompted Image Synthesis [97.37770785712475]
  We present a generation-based debiasing framework for object detection.<n>Our method significantly narrows the performance gap for underrepresented object groups.
  arXiv  Detail & Related papers  (2025-10-21T02:19:12Z)
• DLaVA: Document Language and Vision Assistant for Answer Localization with Enhanced Interpretability and Trustworthiness [34.170341753045776]
  Document Visual Question Answering (VQA) demands robust integration of text detection, recognition, and spatial reasoning.<n>We introduce DLaVA, a training-free pipeline that leverages Multimodal Large Language Models (MLLMs) for zero-shot answer localization.
  arXiv  Detail & Related papers  (2024-11-29T06:17:11Z)
• Analysing Zero-Shot Readability-Controlled Sentence Simplification [54.09069745799918]
  We investigate how different types of contextual information affect a model's ability to generate sentences with the desired readability.<n>Results show that all tested models struggle to simplify sentences due to models' limitations and characteristics of the source sentences.<n>Our experiments also highlight the need for better automatic evaluation metrics tailored to RCTS.
  arXiv  Detail & Related papers  (2024-09-30T12:36:25Z)
• Detecting Errors through Ensembling Prompts (DEEP): An End-to-End LLM Framework for Detecting Factual Errors [11.07539342949602]
  We propose an end-to-end framework for detecting factual errors in text summarization. Our framework uses a diverse set of LLM prompts to identify factual inconsistencies. We calibrate the ensembled models to produce empirically accurate probabilities that a text is factually consistent or free of hallucination.
  arXiv  Detail & Related papers  (2024-06-18T18:59:37Z)
• Classifiers are Better Experts for Controllable Text Generation [63.17266060165098]
  We show that the proposed method significantly outperforms recent PPLM, GeDi, and DExperts on PPL and sentiment accuracy based on the external classifier of generated texts. The same time, it is also easier to implement and tune, and has significantly fewer restrictions and requirements.
  arXiv  Detail & Related papers  (2022-05-15T12:58:35Z)
• Toward the Understanding of Deep Text Matching Models for Information Retrieval [72.72380690535766]
  This paper aims at testing whether existing deep text matching methods satisfy some fundamental gradients in information retrieval. Specifically, four attributions are used in our study, i.e., term frequency constraint, term discrimination constraint, length normalization constraints, and TF-length constraint. Experimental results on LETOR 4.0 and MS Marco show that all the investigated deep text matching methods satisfy the above constraints with high probabilities in statistics.
  arXiv  Detail & Related papers  (2021-08-16T13:33:15Z)
• Information-Theoretic Probing with Minimum Description Length [74.29846942213445]
  We propose an alternative to the standard probes, information-theoretic probing with minimum description length (MDL) With MDL probing, training a probe to predict labels is recast as teaching it to effectively transmit the data. We show that these methods agree in results and are more informative and stable than the standard probes.
  arXiv  Detail & Related papers  (2020-03-27T09:35:38Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.