When Truthful Representations Flip Under Deceptive Instructions?

Large language models (LLMs) tend to follow maliciously crafted instructions to generate deceptive responses, posing safety challenges. How deceptive instructions alter the internal representations of LLM compared to truthful ones remains poorly understood beyond output analysis. To bridge this gap, we investigate when and how these representations ``flip'', such as from truthful to deceptive, under deceptive versus truthful/neutral instructions. Analyzing the internal representations of Llama-3.1-8B-Instruct and Gemma-2-9B-Instruct on a factual verification task, we find the model's instructed True/False output is predictable via linear probes across all conditions based on the internal representation. Further, we use Sparse Autoencoders (SAEs) to show that the Deceptive instructions induce significant representational shifts compared to Truthful/Neutral representations (which are similar), concentrated in early-to-mid layers and detectable even on complex datasets. We also identify specific SAE features highly sensitive to deceptive instruction and use targeted visualizations to confirm distinct truthful/deceptive representational subspaces. % Our analysis pinpoints layer-wise and feature-level correlates of instructed dishonesty, offering insights for LLM detection and control. Our findings expose feature- and layer-level signatures of deception, offering new insights for detecting and mitigating instructed dishonesty in LLMs.

Model-Agnostic Gender Bias Control for Text-to-Image Generation via Sparse Autoencoder

Text-to-image (T2I) diffusion models often exhibit gender bias, particularly by generating stereotypical associations between professions and gendered subjects. This paper presents SAE Debias, a lightweight and model-agnostic framework for mitigating such bias in T2I generation. Unlike prior approaches that rely on CLIP-based filtering or prompt engineering, which often require model-specific adjustments and offer limited control, SAE Debias operates directly within the feature space without retraining or architectural modifications. By leveraging a k-sparse autoencoder pre-trained on a gender bias dataset, the method identifies gender-relevant directions within the sparse latent space, capturing professional stereotypes. Specifically, a biased direction per profession is constructed from sparse latents and suppressed during inference to steer generations toward more gender-balanced outputs. Trained only once, the sparse autoencoder provides a reusable debiasing direction, offering effective control and interpretable insight into biased subspaces. Extensive evaluations across multiple T2I models, including Stable Diffusion 1.4, 1.5, 2.1, and SDXL, demonstrate that SAE Debias substantially reduces gender bias while preserving generation quality. To the best of our knowledge, this is the first work to apply sparse autoencoders for identifying and intervening in gender bias within T2I models. These findings contribute toward building socially responsible generative AI, providing an interpretable and model-agnostic tool to support fairness in text-to-image generation.

Latent Inter-User Difference Modeling for LLM Personalization

Large language models (LLMs) are increasingly integrated into users' daily lives, leading to a growing demand for personalized outputs. Previous work focuses on leveraging a user's own history, overlooking inter-user differences that are crucial for effective personalization. While recent work has attempted to model such differences, the reliance on language-based prompts often hampers the effective extraction of meaningful distinctions. To address these issues, we propose Difference-aware Embedding-based Personalization (DEP), a framework that models inter-user differences in the latent space instead of relying on language prompts. DEP constructs soft prompts by contrasting a user's embedding with those of peers who engaged with similar content, highlighting relative behavioral signals. A sparse autoencoder then filters and compresses both user-specific and difference-aware embeddings, preserving only task-relevant features before injecting them into a frozen LLM. Experiments on personalized review generation show that DEP consistently outperforms baseline methods across multiple metrics. Our code is available at https://github.com/SnowCharmQ/DEP.

Sparse Autoencoders for Sequential Recommendation Models: Interpretation and Flexible Control

Many current state-of-the-art models for sequential recommendations are based on transformer architectures. Interpretation and explanation of such black box models is an important research question, as a better understanding of their internals can help understand, influence, and control their behavior, which is very important in a variety of real-world applications. Recently sparse autoencoders (SAE) have been shown to be a promising unsupervised approach for extracting interpretable features from language models. These autoencoders learn to reconstruct hidden states of the transformer's internal layers from sparse linear combinations of directions in their activation space. This paper is focused on the application of SAE to the sequential recommendation domain. We show that this approach can be successfully applied to the transformer trained on a sequential recommendation task: learned directions turn out to be more interpretable and monosemantic than the original hidden state dimensions. Moreover, we demonstrate that the features learned by SAE can be used to effectively and flexibly control the model's behavior, providing end-users with a straightforward method to adjust their recommendations to different custom scenarios and contexts.

Teach Old SAEs New Domain Tricks with Boosting

Sparse Autoencoders have emerged as powerful tools for interpreting the internal representations of Large Language Models, yet they often fail to capture domain-specific features not prevalent in their training corpora. This paper introduces a residual learning approach that addresses this feature blindness without requiring complete retraining. We propose training a secondary SAE specifically to model the reconstruction error of a pretrained SAE on domain-specific texts, effectively capturing features missed by the primary model. By summing the outputs of both models during inference, we demonstrate significant improvements in both LLM cross-entropy and explained variance metrics across multiple specialized domains. Our experiments show that this method efficiently incorporates new domain knowledge into existing SAEs while maintaining their performance on general tasks. This approach enables researchers to selectively enhance SAE interpretability for specific domains of interest, opening new possibilities for targeted mechanistic interpretability of LLMs.

CytoSAE: Interpretable Cell Embeddings for Hematology

Sparse autoencoders (SAEs) emerged as a promising tool for mechanistic interpretability of transformer-based foundation models. Very recently, SAEs were also adopted for the visual domain, enabling the discovery of visual concepts and their patch-wise attribution to tokens in the transformer model. While a growing number of foundation models emerged for medical imaging, tools for explaining their inferences are still lacking. In this work, we show the applicability of SAEs for hematology. We propose CytoSAE, a sparse autoencoder which is trained on over 40,000 peripheral blood single-cell images. CytoSAE generalizes to diverse and out-of-domain datasets, including bone marrow cytology, where it identifies morphologically relevant concepts which we validated with medical experts. Furthermore, we demonstrate scenarios in which CytoSAE can generate patient-specific and disease-specific concepts, enabling the detection of pathognomonic cells and localized cellular abnormalities at the patch level. We quantified the effect of concepts on a patient-level AML subtype classification task and show that CytoSAE concepts reach performance comparable to the state-of-the-art, while offering explainability on the sub-cellular level. Source code and model weights are available at https://github.com/dynamical-inference/cytosae.

Learning Representations of Event Time Series with Sparse Autoencoders for Anomaly Detection, Similarity Search, and Unsupervised Classification

Event time series are sequences of discrete events occurring at irregular time intervals, each associated with a domain-specific observational modality. They are common in domains such as high-energy astrophysics, computational social science, cybersecurity, finance, healthcare, neuroscience, and seismology. Their unstructured and irregular structure poses significant challenges for extracting meaningful patterns and identifying salient phenomena using conventional techniques. We propose novel two- and three-dimensional tensor representations for event time series, coupled with sparse autoencoders that learn physically meaningful latent representations. These embeddings support a variety of downstream tasks, including anomaly detection, similarity-based retrieval, semantic clustering, and unsupervised classification. We demonstrate our approach on a real-world dataset from X-ray astronomy, showing that these representations successfully capture temporal and spectral signatures and isolate diverse classes of X-ray transients. Our framework offers a flexible, scalable, and generalizable solution for analyzing complex, irregular event time series across scientific and industrial domains.

VITA: Vision-to-Action Flow Matching Policy

We present VITA, a Vision-To-Action flow matching policy that evolves latent visual representations into latent actions for visuomotor control. Traditional flow matching and diffusion policies sample from standard source distributions (e.g., Gaussian noise) and require additional conditioning mechanisms like cross-attention to condition action generation on visual information, creating time and space overheads. VITA proposes a novel paradigm that treats latent images as the flow source, learning an inherent mapping from vision to action while eliminating separate conditioning modules and preserving generative modeling capabilities. Learning flows between fundamentally different modalities like vision and action is challenging due to sparse action data lacking semantic structures and dimensional mismatches between high-dimensional visual representations and raw actions. We address this by creating a structured action latent space via an autoencoder as the flow matching target, up-sampling raw actions to match visual representation shapes. Crucially, we supervise flow matching with both encoder targets and final action outputs through flow latent decoding, which backpropagates action reconstruction loss through sequential flow matching ODE solving steps for effective end-to-end learning. Implemented as simple MLP layers, VITA is evaluated on challenging bi-manual manipulation tasks on the ALOHA platform, including 5 simulation and 2 real-world tasks. Despite its simplicity, MLP-only VITA outperforms or matches state-of-the-art generative policies while reducing inference latency by 50-130% compared to conventional flow matching policies requiring different conditioning mechanisms or complex architectures. To our knowledge, VITA is the first MLP-only flow matching policy capable of solving complex bi-manual manipulation tasks like those in ALOHA benchmarks.

LoViC: Efficient Long Video Generation with Context Compression

Despite recent advances in diffusion transformers (DiTs) for text-to-video generation, scaling to long-duration content remains challenging due to the quadratic complexity of self-attention. While prior efforts -- such as sparse attention and temporally autoregressive models -- offer partial relief, they often compromise temporal coherence or scalability. We introduce LoViC, a DiT-based framework trained on million-scale open-domain videos, designed to produce long, coherent videos through a segment-wise generation process. At the core of our approach is FlexFormer, an expressive autoencoder that jointly compresses video and text into unified latent representations. It supports variable-length inputs with linearly adjustable compression rates, enabled by a single query token design based on the Q-Former architecture. Additionally, by encoding temporal context through position-aware mechanisms, our model seamlessly supports prediction, retradiction, interpolation, and multi-shot generation within a unified paradigm. Extensive experiments across diverse tasks validate the effectiveness and versatility of our approach.

Exploring Task Performance with Interpretable Models via Sparse Auto-Encoders

Large Language Models (LLMs) are traditionally viewed as black-box algorithms, therefore reducing trustworthiness and obscuring potential approaches to increasing performance on downstream tasks. In this work, we apply an effective LLM decomposition method using a dictionary-learning approach with sparse autoencoders. This helps extract monosemantic features from polysemantic LLM neurons. Remarkably, our work identifies model-internal misunderstanding, allowing the automatic reformulation of the prompts with additional annotations to improve the interpretation by LLMs. Moreover, this approach demonstrates a significant performance improvement in downstream tasks, such as mathematical reasoning and metaphor detection.