Subliminal Effects in Your Data: A General Mechanism via Log-Linearity

• URL: http://arxiv.org/abs/2602.04863v1
• Date: Wed, 04 Feb 2026 18:50:46 GMT
• Title: Subliminal Effects in Your Data: A General Mechanism via Log-Linearity
• Abstract summary: Training modern large language models (LLMs) has become a veritable smorgasbord of algorithms and datasets designed to elicit particular behaviors.<n>Recent experiments show datasets can transmit signals that are not directly observable from individual datapoints.<n>We uncover a general mechanism through which hidden subtexts can arise in generic datasets.
• Score: 39.40622797996004
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Training modern large language models (LLMs) has become a veritable smorgasbord of algorithms and datasets designed to elicit particular behaviors, making it critical to develop techniques to understand the effects of datasets on the model's properties. This is exacerbated by recent experiments that show datasets can transmit signals that are not directly observable from individual datapoints, posing a conceptual challenge for dataset-centric understandings of LLM training and suggesting a missing fundamental account of such phenomena. Towards understanding such effects, inspired by recent work on the linear structure of LLMs, we uncover a general mechanism through which hidden subtexts can arise in generic datasets. We introduce Logit-Linear-Selection (LLS), a method that prescribes how to select subsets of a generic preference dataset to elicit a wide range of hidden effects. We apply LLS to discover subsets of real-world datasets so that models trained on them exhibit behaviors ranging from having specific preferences, to responding to prompts in a different language not present in the dataset, to taking on a different persona. Crucially, the effect persists for the selected subset, across models with varying architectures, supporting its generality and universality.

Related papers

• Position: Let's Develop Data Probes to Fundamentally Understand How Data Affects LLM Performance [33.08440302959991]
  We advocate for the need of developing systematic methodologies for generating synthetic sequences from appropriately defined random processes.<n>By observing LLM behavior on data probes, researchers can systematically conduct studies on how data characteristics influence model performance, generalization, and robustness.
  arXiv  Detail & Related papers  (2026-05-11T11:44:40Z)
• What Language Models Know But Don't Say: Non-Generative Prior Extraction for Generalization [5.663538370244175]
  We propose LoID, a deterministic method for extracting informative prior distributions for Bayesian logistic regression.<n>Rather than relying on generated text, we probe the model's confidence in opposing semantic directions through carefully constructed sentences.<n>We evaluate LoID on ten real-world datasets under synthetic out-of-distribution (OOD) settings.
  arXiv  Detail & Related papers  (2026-01-24T22:05:01Z)
• Latent Traits and Cross-Task Transfer: Deconstructing Dataset Interactions in LLM Fine-tuning [8.082936847467638]
  We propose an analysis framework, building a transfer learning matrix and dimensionality reduction, to dissect cross-task interactions.<n>We train and analyze 10 models to identify latent abilities (e.g., Reasoning, Sentiment Classification, NLU, Arithmetic) and discover the side effects of the transfer learning.
  arXiv  Detail & Related papers  (2025-09-17T01:45:42Z)
• SCAR: A Characterization Scheme for Multi-Modal Dataset [12.906460823872749]
  Foundation models exhibit remarkable generalization across diverse tasks, largely driven by the characteristics of their training data.<n>Recent data-centric methods like pruning and compression aim to optimize training but offer limited theoretical insight into how data properties affect generalization.<n>We introduce SCAR, a principled scheme for characterizing the intrinsic structural properties of datasets.
  arXiv  Detail & Related papers  (2025-08-27T08:16:59Z)
• Shortcut Learning in Generalist Robot Policies: The Role of Dataset Diversity and Fragmentation [117.54237701533805]
  Generalist robot policies trained on large-scale datasets such as Open X-Embodiment (OXE) demonstrate strong performance across a wide range of tasks.<n>They often struggle to generalize beyond the distribution of their training data.<n>We identify shortcut learning as a key impediment to generalization.
  arXiv  Detail & Related papers  (2025-08-08T16:14:01Z)
• Meta-Statistical Learning: Supervised Learning of Statistical Inference [59.463430294611626]
  This work demonstrates that the tools and principles driving the success of large language models (LLMs) can be repurposed to tackle distribution-level tasks.<n>We propose meta-statistical learning, a framework inspired by multi-instance learning that reformulates statistical inference tasks as supervised learning problems.
  arXiv  Detail & Related papers  (2025-02-17T18:04:39Z)
• Diversity as a Reward: Fine-Tuning LLMs on a Mixture of Domain-Undetermined Data [54.3895971080712]
  Fine-tuning large language models (LLMs) using diverse datasets is crucial for enhancing their overall performance across various domains.<n>We propose a new method that gives the LLM a dual identity: an output model to cognitively probe and select data based on diversity reward, as well as an input model to be tuned with the selected data.
  arXiv  Detail & Related papers  (2025-02-05T17:21:01Z)
• Amortized Inference of Causal Models via Conditional Fixed-Point Iterations [17.427722515310606]
  We propose amortized inference of Structural Causal Models (SCMs) by training a single model on multiple datasets sampled from different SCMs.<n>We first use a transformer-based architecture for amortized learning of dataset embeddings, and then extend the Fixed-Point Approach (FiP) to infer SCMs conditionally on their dataset embeddings.<n>As a byproduct, our method can generate observational and interventional data from novel SCMs at inference time, without updating parameters.
  arXiv  Detail & Related papers  (2024-10-08T15:31:33Z)
• ALP: Action-Aware Embodied Learning for Perception [60.64801970249279]
  We introduce Action-Aware Embodied Learning for Perception (ALP) ALP incorporates action information into representation learning through a combination of optimizing a reinforcement learning policy and an inverse dynamics prediction objective. We show that ALP outperforms existing baselines in several downstream perception tasks.
  arXiv  Detail & Related papers  (2023-06-16T21:51:04Z)
• infoVerse: A Universal Framework for Dataset Characterization with Multidimensional Meta-information [68.76707843019886]
  infoVerse is a universal framework for dataset characterization. infoVerse captures multidimensional characteristics of datasets by incorporating various model-driven meta-information. In three real-world applications (data pruning, active learning, and data annotation), the samples chosen on infoVerse space consistently outperform strong baselines.
  arXiv  Detail & Related papers  (2023-05-30T18:12:48Z)
• Learning from Temporal Spatial Cubism for Cross-Dataset Skeleton-based Action Recognition [88.34182299496074]
  Action labels are only available on a source dataset, but unavailable on a target dataset in the training stage. We utilize a self-supervision scheme to reduce the domain shift between two skeleton-based action datasets. By segmenting and permuting temporal segments or human body parts, we design two self-supervised learning classification tasks.
  arXiv  Detail & Related papers  (2022-07-17T07:05:39Z)

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.