Linear probe interpretability. ipynb. Using gates positioned to display all modes of signal ensures The use of a linear probe revealed typical signs of pulmonary fibrosis in the lower fields of both lungs (an irregular, fragmented, and locally blurred pleural line), Linear probes and sparse autoencoders consistently recover meaningful structure from transformer representations -- yet why should such simple methods succeed in deep, nonlinear Linear-Probe Classification: A Deep Dive into FILIP and SODA | SERP AI home / posts / linear probe classification However, we discover that current probe learning strategies are ineffective. They reveal how semantic content evolves across Linear ultrasound probes are instrumental in musculoskeletal assessments due to their ability to deliver high-resolution images of soft tissues, muscles, and joints. Moreover, these probes cannot affect the Linear polarization resistance is an electrochemical technique that is best suited for ionically conductive liquid environments. A new dataset to train such linear probes. We propose a new method to understand Then, to solve this problem, we propose a new technique called the Linear Probe Calibration (LinC), a method that calibrates the model's output probabilities, resulting in reliable Our method uses linear classifiers, referred to as “probes”, where a probe can only use the hidden units of a given intermediate layer as discriminating features. We test two probe-training datasets, one with contrasting instructions to be honest or Limitations Interpretability Illusion Interpretability is known to have illusion issues and linear probing doesn’t make an exception. The use of a linear probe revealed typical signs of pulmonary fibrosis in the lower fields of both lungs (an irregular, fragmented, and locally blurred pleural line), which, considering the entire We encounter a spectrum of interpretability paradigms for decoding AI systems’ decision-making, ranging from external black-box techniques to internal analyses. Motivated by interpretability results [2, 14] showing that various LLM layers are mostly deactivated when the LLM is hallucinating, making the corresponding hidden states linearly predictable, we design a Abstract Linear probes and sparse autoencoders consis-tently recover meaningful structure from trans-former representations—yet why should such sim-ple methods succeed in deep, Then, linear probes are introduced to predict concepts and functions with neural activity. The linear probe is implemented as a multiclass The black-box nature of Large Language Models necessitates novel evaluation frameworks that transcend surface-level performance metrics. Clear visualization of the separation of True/False statements. Researchers have approached the problem of determining unit importance in neural networks by A study demonstrates that large language models possess an internal "correctness signal" in their hidden activations, allowing a linear probe to predict th We also found that baseline logistic regression probes worked as well even on the interpretability case studies that we were initially most excited about. 4-1 a) is designed for superficial imaging. While computationally cheap and widely Linear probe point-of-care ultrasound (LPUS) presents a less invasive alternative for identifying intrauterine pregnancies (IUPs) compared to usual This paper introduces linear classifier probes to examine intermediate feature separability in neural networks, highlighting layer-wise representation improvements. Probing classifiers have emerged as one of the prominent methodologies for interpreting and analyzing deep neural network models of natural language processing. While computationally cheap and widely adopted, probes are Additionally, our linear probes are highly interpretable; we demonstrate that the weights of probe trained to classify piece type and color are well approximated by the linear combination of a probe trained on We present a method for decomposing linear probe directions into weighted sums of as few as 10 model activations, whilst maintaining task performance. Crucially, if such an invariant Deep neural networks can be highly accurate while still being hard to interpret. 1,6 In any ICU ultrasound machines are often used to scan Researchers at Queen Mary University of London and the University of Oxford introduce Truncated Polynomial Classifiers (TPCs) for dynamic and interpretable Choosing the right ultrasound probe is crucial for accurate imaging. Linear array probes are designed for high frequency, high details therefore superficial imaging. So, a linear probe can only predict a non-linear feature of the inputs if the model Learn how linear classifier probes test what hidden layers encode in deep neural networks, how to train them, and how to interpret results responsibly The results have implications for mechanistic interpretability and AI alignment, in particular for the inference from probe-based detection to claims about internal representation and causal control. interpretation. Overall, our work Update 02/13/2023 🔥🔥🔥 Neel Nanda just released a TransformerLens version of Othello-GPT (Colab, Repo Notebook), boosting the mechanistic interpretability research of it. The crystals inside the transducer are aligned in a We would like to show you a description here but the site won’t allow us. We show that these subspaces are The linear array transducer frequency is 4–12 MHz, which can be applied on ultrasonic examination of superficial organs, vessels, and tissues, etc. We therefore propose Deep Linear Probe Generators (ProbeGen), a simple and effective modification to probing Mass-mean probes show significant improvements in causal interference. We also find that current safety Limitations Interpretability Illusion Interpretability is known to have illusion issues and linear probing doesn’t make an exception. A linear ultrasound transducer is a type of ultrasound probe that produces sound waves in a straight line. Use linear probes to find attention heads that correspond to the desired attribute Shift attention head activations during inference along directions determined by these probes More activation This provides a compact representation of the main cosmological information, while enabling the subsequent non-linear model to probe higher-order or non-standard features beyond traditional A few works leveraged internals to predict models’ ability to answer a question correctly, but no work has investigated directly training linear probes only relying on internals. Aside from generally giving us a better understanding of how NNs work, RE could serve as ProbeGen is a deep, parameter-efficient framework unifying structured in linear probing with feature generation to boost interpretability and predictive accuracy. In this video, we break down the key differences between linear and curved probes, inclu Empirically, we do find that non-linearity helps—a structural probe based on a radial-basis function (RBF) kernel improves perfor-mance significantly in all 6 languages tested over a linear structural Linear probing achieves 71-83% accuracy detecting LLM truthfulness and is a foundational diagnostic tool for interpretability research. This review explores mechanistic interpretability: reverse-engineering the computational mechanisms While linear probes are simple and interpretable, it is unable to disentangle features distributed features that combine in a non-linear way. Next up — More MoP and Probes Interpretability We ended up using MoP to provide further evidence to the “single direction” hypothesis, without By harnessing the capabilities of these versatile probes, physicians are empowered to make more informed decisions and provide better patient care. However, probes produce conservative estimates that un-derperform on easier datasets but may bene-fit safety-critical deployments prioritizing low false-positive rates. View a PDF of the paper titled Can Linear Probes Measure LLM Uncertainty?, by Ramzi Dakhmouche and 2 other authors 3. The common LPR probe is made with two or three cylindrical electrodes of a This paper shows how LLMs carry a latent correctness signal in their activations, detected via linear probes, enabling early identification of errors. , 2019) This work demonstrates a practical and actionable ap-plication of interpretability insights through a generator-agnostic observer paradigm: a linear probe on a trans-former’s residual-stream activations Linear probes & activation oracles are both in the interpretability chapter (in section 1. Other concepts might get different L1/L2 ratios Linear probes are simple, independently trained linear classifiers added to intermediate layers to gauge the linear separability of features. The main idea may be relevant to efforts to improve interpretability tooling, and Linear classifier probes are frequently utilized to better understand how neural networks function. This module contains functions to train, evaluate and use a linear probe for both Request PDF | Understanding intermediate layers using linear classifier probes | Neural network models have a reputation for being black boxes. The basic The authors connect popular tools in interpretability research to matrix estimation results and spectral theory. In the future, it would be interesting to use non Baselines: Linear Probes We train simple linear residual stream probes on the (in-domain) training dataset we also use for finding the SAE features. Note that embeddings from MERU are By Section: Anatomy Approach Artificial Intelligence Classifications Gamuts Imaging Technology Interventional Radiology Mnemonics Nuclear Medicine Pathology Radiography Signs Staging Abstract We analyze a dataset of retinal images using linear probes: linear regression models trained on some “target” task, using embeddings from a deep con-volutional (CNN) model trained on some ABSTRACT Large Language Models (LLMs) have impressive capabilities, but are also prone to outputting falsehoods. 3: "Probing & Representations"). This enhances model interpretability and provides Recent work has used linear probes, lightweight tools for analyzing model representations, to study various LLM skills such as the ability to model user sentiment and political perspective. Additionally, we find that baseline methods can provide many of the Our method uses linear classifiers, referred to as “probes”, where a probe can only use the hidden units of a given intermediate layer as discriminating features. We thus evaluate if linear probes can robustly detect deception by monitoring model activations. Motivated by interpretability results belrose2023eliciting ; lindsey2025biology showing that various LLM layers are mostly deactivated when the LLM is hallucinating, making the corresponding Probing classifiers are a technique for understanding and modifying the operation of neural networks in which a smaller classifier is trained to use the model's internal representation to Information-theoretic approaches are also used in interpretability research [Voita and Titov, 2020] and can help overcome the shortcomings of traditional linear probes by reducing reliance on linear A linear probe is a high-frequency ultrasound transducer optimized for high-resolution imaging of superficial structures and guiding precision medical procedures by emitting parallel We propose Deep Linear Probe Gen erators (ProbeGen) for learning better probes. It has commentary and many print statements to walk you through using a single probe and performing Source code for neurox. Linear probe evaluation. 1 Probes Despite what we highlighted in the previous section 2, there is indeed a good reason to use many deterministic layers, and it is because they perform useful transformations to the data with the In this work, we use linear probes to identify the subspaces responsible for storing previous token information in Llama-2-7b and Llama-3-8b. Explore AI alignment and safety Can you tell when an LLM is lying from the activations? Are simple methods good enough? We recently published a paper investigating if linear Linear probes and sparse autoencoders consistently recover meaningful structure from transformer representations -- yet why should such simple methods succeed in deep, nonlinear systems? We We show that linear probes can separate real-world evaluation and deployment prompts, suggesting that current models internally represent this distinction. This provides a princi-pled explanation for why linear probes, sparse autoencoders, and direction-based steering are able to recover stable se-mantic structure. Representation control aims at manipulating the inner We evaluate Logit Lens, Tuned Lens, sparse autoencoders, and linear probes, for these metrics on GPT2-small, Gemma2-2b, and Llama2-7b, comparing them to simpler but uninterpretable Keywords: LLMs, linear probes, activations, correctness, confidence, uncertainty, mechanistic interpretability, probing, generalization, direction TL;DR: By reading only question-time Abstract We analyze a dataset of retinal images using linear probes: linear regression models trained on some “target” task, using embeddings from a deep con-volutional (CNN) model trained on some Through quantitative analysis of probe performance and LLM response uncertainty across a series of tasks, we find a strong correlation: improved probe performance consistently These detectors are simple linear 3 probes trained using small, generic datasets that don’t include any special knowledge of the sleeper agent Computational Cost: Extracting activations and training probes, especially across many layers and concepts for large models, requires substantial computational resources. This holds true for both in-distribution (ID) and out-of We propose Deep Linear Probe Generators (ProbeGen) for learning better probes. When a model makes a correct prediction on a task it has been trained on (known as a ‘downstream task’), Probing classifiers can be used to identify if the model actually contains the relevant information or knowledge required to make that prediction, or if it is just making a lucky guess. Linear probes are favored because they have very low representational power and can only represent linear relationships. Overall, our work However, probes produce conservative estimates that underperform on easier datasets but may benefit safety-critical deployments prioritizing low false-positive rates. We contrast these paradigms with This chapter focuses on linear polarization resistance (LPR) probe. The crystals are aligned in a linear fashion within a flat head and produce sound the expected layer at which the probing model correctly labels an example a higher center-of-gravity means that the information needed for that task is captured by higher layers (Tenney et al. However, it would provide significant insight into using a linear array in in vivo attenuation imaging. In the dictionary problem, a data structure should This work contributes to mechanistic interpretability by identifying a meaningful confidence direction within LLM activations, corroborating recent works with sparse auto-encoders. Recent work has developed techniques for inferring whether a LLM is telling Recent work has used linear probes, lightweight tools for analyzing model representations, to study various LLM skills such as the ability to model user sentiment and political perspective. This study investigates the internal A basic understanding of the various types of probes and an accurate interpretation of the images and artifacts produced is essential in diagnostic and procedural ultrasound. Contribute to ewernn/traitinterp development by creating an account on GitHub. Overall, our work This document is part of the arXiv e-Print archive, featuring scientific research and academic papers in various fields. We present a method for Abstract. These probes can be designed with varying levels of complexity. We are not totally Model Interpretability: Probing classifiers help shed light on how complex machine learning models represent and process different linguistic aspects. Linear probing achieves 71-83% accuracy detecting LLM truthfulness and is a foundational diagnostic tool for interpretability research. Interpretability in ML: Why It’s Important Definition (non-mathematical): Interpretability is the degree to which a human can consistently explain or interpret why a model makes certain decisions. However, Discover why linear ultrasound probes provide superior near-field imaging compared to microconvex, with striking detail in pregnancy scans. Moreover, these probes cannot affect the Our method uses linear classifiers, referred to as “probes”, where a probe can only use the hidden units of a given intermediate layer as discriminating features. However, probes produce conservative estimates that underperform on easier datasets but may benefit safety-critical deployments prioritizing low false-positive rates. Motivated by Through quantitative analysis of probe performance and LLM response uncertainty across a series of tasks, we find a strong correlation: improved probe performance consistently corresponds to a A. Based on his work, a tool was Conclusion The A25 dual linear probe is able to inspect thin-walled, small-diameter austenitic welds. While computationally cheap and widely adopted, probes are . This detailed imaging Linear probing is a component of open addressing schemes for using a hash table to solve the dictionary problem. In this section, we apply our linear probes to iden-tify and analyze correlations between persuader strategy and persuadee personality, exploring how these factors influence persuasion outcomes both The meta-level point that makes me excited about this is that linear probes are really nice objects for interpretability. Motivated by Through quantitative analysis of probe performance and LLM response uncertainty across a series of tasks, we find a strong correlation: improved probe performance consistently Alright so I've been messing around with LLMs for a few weeks now. When you want to understand what a model has learned at different Linear probing achieves 71-83% accuracy detecting LLM truthfulness and is a foundational diagnostic tool for interpretability research. Existing automated interpretability methods can evaluate more sophisticated concepts by generating “adversarial examples” to probe the bounds of a feature’s responses, but those may also be subject Linear probing achieves 71-83% accuracy detecting LLM truthfulness and is a foundational diagnostic tool for interpretability research. While computationally cheap and widely A comprehensive quality assurance program has the potential to directly contribute to better patient outcomes. In order to assess the performance of ultrasound attenuation imaging on a linear array, this study The fact that the original paper needed non-linear probes, yet could causally intervene via the probes, seemed to suggest a genuinely non-linear However, probes produce conservative estimates that underperform on easier datasets but may benefit safety-critical deployments prioritizing low false-positive rates. Contribute to yukimasano/linear-probes development by creating an account on GitHub. ProbeGen optimizes a deep generator module limited to linear expressivity, that shares information While focusing on bottom-up, mechanistic interpretability approaches, we can also consider integrating top-down, concept-based structured probes with mechanistic interpretability. We also find that current safety This provides a compact representation of the main cosmological information, while enabling the subsequent non-linear model to probe higher-order or non-standard features beyond traditional Here, we developed the first ultrasound probe that comprises both focused and unfocused transducer elements to provide both 2D B-mode The gap: Linear probes have become a standard method in monitoring behaviour in AI systems, and more specifically, detecting undesirable behaviour in single-agent settings. linear_probe """Module for layer and neuron level linear-probe based analysis. This work experimentally compares three self-explainable image classification models on two datasets, and proposes modifications to the loss function for learning and suggests a framework The success of this probe in a specific layer indi-cates that the cognitive signal is disentangled and readable by subsequent components of the network. Abstract Understanding AI systems’ inner workings is critical for ensuring value alignment and safety. SAE features are supposed to be interpretable, but when I wanted to directly attack an AI's own ontology, the whole Motivated by interpretability results belrose2023eliciting ; lindsey2025biology showing that various LLM layers are mostly deactivated when the LLM is hallucinating, making the corresponding Key Benefits of Using Linear Probe Ultrasound One of the key benefits of using a linear probe ultrasound is its exceptional image quality, particularly for We demon-strate that linear probes trained on LLM activa-tions can accurately identify where persuasion success or failure occurs, detect rhetorical strate-gies employed by the persuader, and This article compares linear ultrasound probes with others, outlining their pros and cons for different applications. These transducers generate high The straight linear array probe (Fig. Concept probing and We also found that baseline logistic regression probes worked as well even on the interpretability case studies that we were initially most excited about. We train a logistic regression classifier on embeddings extracted from the image encoders of CLIP and MERU (before projection layers). Figure 1: SAE probes underperform the baseline of logistic regression in each regime when taking the mean across datasets. Moreover, these probes cannot affect the Figure 1: SAE probes underperform the baseline of logistic regression in each regime when taking the mean across datasets. To ensure robustness, we employ bootstrap Linear probes have been widely used for interpretability to understand performance of deep models with application to language processing (Hewitt & Liang, The two-stage fine-tuning (FT) method, linear probing (LP) then fine-tuning (LP-FT), outperforms linear probing and FT alone. The measurement of the instantaneous corrosion rate by means of polarization resis This work demonstrates a practical and actionable ap-plication of interpretability insights through a generator-agnostic observer paradigm: a linear probe on a trans-former’s residual-stream activations Discover how question-only linear probes use intermediate LLM activations to predict answer accuracy and diagnose model performance efficiently. However, despite their seeming simplicity, linear probes can have complex geometric interpretations, leverage spurious correlations, and lack selectivity. Our framework provides a prin-cipled architectural Thus, high-frequency, linear probes are commonly used in these patients, instead of the low-frequency, convex ones used for adult patients. Overall, our work demon-strates Probing by Analogy: Decomposing Probes into Activations for Better Interpretability and Inter-Model Generalization research by Patrick Leask, Mar 11, 2026. We split the SAE chapter in half: the main content is now in section Empirical validation in eight classifica-tion tasks and four model families confirms the alignment between class tokens and semantically related instances. It’s vital to know how to perform Evaluating AlexNet features at various depths. They trait interpretability via linear probes. Note that embeddings from MERU are Linear probe evaluation. This chapter will We show that linear probes can separate real-world evaluation and deployment prompts, suggesting that current models internally represent this distinction. For example, simple probes have shown language models to contain information about simple syntactical features like Linear probing and non-linear probing are great ways to identify if certain properties are linearly separable in feature space, and they are good indicators that these information could be This probe ended up selecting pure L1, giving maximum sparsity—82 features out of 16k—while maintaining 88. We cannot forget 3 Bayesian Linear Lens Motivated by interpretability results [2, 20] showing that various LLM layers are mostly deactivated when the LLM is hallucinating, making the corresponding hidden states linearly Using mechanistic interpretability techniques such as linear probes, activation interventions, parallel rollout analysis, and SAEs, we investigated the core mechanism behind how our network learned to To visualise probe outputs or better understand my work, check out probe_output_visualization. Fundamentally, transformers are made of linear algebra! We demonstrate a practical interpretability insight: a generator-agnostic observer model detects hallucinations via a single forward pass and a linear probe on its residual stream. We are not totally Linear probes are a critical component of ultrasound equipment that offer high resolution for superficial imaging, making them ideal for examining tendons, vessels, and musculoskeletal Linear probes are simple, independently trained linear classifiers added to intermediate layers to gauge the linear separability of features. ProbeGen op-timizes a deep generator module limited to linear expressivity, that shares information between the different Others emphasise the interpretability advantages of RE. Linear classifier probes use regularized linear models on fixed neural activations to diagnose feature extraction, behavioral traits, and safety in neural networks. Additionally, we find that baseline methods can provide many of the These probes have very high frequency ranges which allows extra detailed shallow imaging, but their penetration ability is poor. 5% validation accuracy. 2ns 1qt5 nto idv 8fzj