What do your logits know?
Not on the map. Why?
Abstract (arXiv)
Recent work has shown that probing model internals can reveal a wealth of information not apparent from the model generations. This poses a risk of unintentional or malicious information leakage, where model users are able to learn information that the model owner assumed was inaccessible. Using vision-language models as a testbed, we present the first systematic comparison of information retained at different representational levels as it is compressed from the rich information encoded in the residual stream through two natural bottlenecks: low-dimensional projections of the residual stream obtained using tuned lens, and the final top-k logits most likely to impact model's answer. We show that even easily accessible bottlenecks defined by the model's top logit values can leak task-irrelevant information present in an image-based query, in some cases revealing as much information as direct projections of the full residual stream.
Key facts. No citing papers in the map yet (data through 7 Oct 2026). Builds on 25 of its 35 references that are in the map. First posted on arXiv 10 Apr 2026.
Builds on
25 of 35 references.
References · 25
- 2014Microsoft COCO: Common Objects in Contextcited by 17,498
- 2023Visual Instruction Tuningcited by 7,792
- 2022Locating and Editing Factual Associations in GPTcited by 2,609
- 2023Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyondcited by 2,179
- 2020Shortcut Learning in Deep Neural Networkscited by 1,928
- 2019Language Models as Knowledge Bases?cited by 1,922
- 2023Evaluating Object Hallucination in Large Vision-Language Modelscited by 1,696
- 2016CLEVR: A Diagnostic Dataset for Compositional Language and Elementary Visual Reasoningcited by 1,445
- 2019BERT Rediscovers the Classical NLP Pipelinecited by 1,046
- 2021Probing Classifiers: Promises, Shortcomings, and Advancescited by 831
- 2015Deep Learning and the Information Bottleneck Principlecited by 793
- 2023Whose Opinions Do Language Models Reflect?cited by 759
13 more
- 2022Discovering Latent Knowledge in Language Models Without Supervisioncited by 733
- 2023The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasetscited by 673
- 2023Eliciting Latent Predictions from Transformers with the Tuned Lenscited by 579
- 2017Opening the Black Box of Deep Neural Networks via Informationcited by 494
- 2018What you can cram into a single vector: Probing sentence embeddings for linguistic propertiescited by 452
- 2018Invertible Residual Networkscited by 364
- 2023Language Models Represent Space and Timecited by 306
- 2018Analysis Methods in Neural Language Processing: A Surveycited by 276
- 2024LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinationscited by 230
- 2017Breaking the Softmax Bottleneck: A High-Rank RNN Language Modelcited by 147
- 2023Overthinking the Truth: Understanding how Language Models Process False Demonstrationscited by 78
- 2023Explore Spurious Correlations at the Concept Level in Language Models for Text Classificationcited by 48
- 2025Eliciting Secret Knowledge from Language Modelscited by 23
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2024The First to Know: How Token Distributions Reveal Hidden Knowledge in Large Vision-Language Models?
- 2024What's in the Image? A Deep-Dive into the Vision of Vision Language Models
- 2025Lost in Embeddings: Information Loss in Vision-Language Models
- 2026Seeing without Looking: Do Vision-Language Benchmarks Really Test Vision?
- 2026Pathways of Visual Information Flow in Vision-Language Models
- 2026Bigger or Cheaper? Scale and Quantization Effects on Uncertainty Signals in Vision-Language Models Under Image Degradation
- 2025The Hidden Life of Tokens: Reducing Hallucination of Large Vision-Language Models via Visual Information Steering
- 2026Large Vision-Language Models Get Lost in Attention