What latent features are encoded in language model (LM) representations? Recent work on training sparse autoencoders (SAEs) to disentangle interpretable features in LM representations has shown significant promise. However, evaluating the quality of these SAEs is difficult because we lack a ground-truth collection of interpretable features which we expect good SAEs to identify. We thus propose to measure progress in interpretable dictionary learning by working in the setting of LMs trained on Chess and Othello transcripts. These settings carry natural collections of interpretable features--for example, "there is a knight on F3"--which we leverage into metrics for SAE quality. To guide progress in interpretable dictionary learning, we introduce a new SAE training technique, $p$-annealing, which demonstrates improved performance on our metric.

Large language models (LLMs) have achieved remarkable progress, yet their internal mechanisms remain largely opaque, posing a significant challenge to their safe and reliable deployment. Sparse autoencoders (SAEs) have emerged as a promising tool for decomposing LLM representations into more interpretable features, but explaining the features captured by SAEs remains a challenging task. In this work, we propose SAGE (SAE AGentic Explainer), an agent-based framework that recasts feature interpretation from a passive, single-pass generation task into an active, explanation-driven process. SAGE implements a rigorous methodology by systematically formulating multiple explanations for each feature, designing targeted experiments to test them, and iteratively refining explanations based on empirical activation feedback. Experiments on features from SAEs of diverse language models demonstrate that SAGE produces explanations with significantly higher generative and predictive accuracy compared to state-of-the-art baselines.an agent-based framework that recasts feature interpretation from a passive, single-pass generation task into an active, explanationdriven process. SAGE implements a rigorous methodology by systematically formulating multiple explanations for each feature, designing targeted experiments to test them, and iteratively refining explanations based on empirical activation feedback. Experiments on features from SAEs of diverse language models demonstrate that SAGE produces explanations with significantly higher generative and predictive accuracy compared to state-of-the-art baselines.

Large language models (LLMs) have been widely deployed across numerous fields. Reinforcement Learning from Human Feedback (RLHF) leverages reward models (RMs) as proxies for human preferences to align LLM behaviors with human values, making the accuracy, reliability, and interpretability of RMs critical for effective alignment. However, traditional RMs lack interpretability, offer limited insight into the reasoning behind reward assignments, and are inflexible toward user preference shifts. While recent multidimensional RMs aim for improved interpretability, they often fail to provide feature-level attribution and require costly annotations. To overcome these limitations, we introduce the Sparse Autoencoder-enhanced Reward Model (SARM), a novel architecture that integrates a pretrained Sparse Autoencoder (SAE) into a reward model. SARM maps the hidden activations of LLM-based RM into an interpretable, sparse, and monosemantic feature space, from which a scalar head aggregates feature activations to produce transparent and conceptually meaningful reward scores. Empirical evaluations demonstrate that SARM facilitates direct feature-level attribution of reward assignments, allows dynamic adjustment to preference shifts, and achieves superior alignment performance compared to conventional reward models. Our code is available at https://github.com/schrieffer-z/sarm.

We consider the analysis of Electroencephalography (EEG) and Local Field Potential (LFP) datasets, which are "big" in terms of the size of recorded data but

Sparse autoencoders (SAEs) have emerged as a promising approach for learning interpretable features from neural network activations. However, the optimization landscape for SAE training can be challenging due to correlations in the input data. We demonstrate that applying PCA Whitening to input activations -- a standard preprocessing technique in classical sparse coding -- improves SAE performance across multiple metrics. Through theoretical analysis and simulation, we show that whitening transforms the optimization landscape, making it more convex and easier to navigate. We evaluate both ReLU and Top-K SAEs across diverse model architectures, widths, and sparsity regimes. Empirical evaluation on SAEBench, a comprehensive benchmark for sparse autoencoders, reveals that whitening consistently improves interpretability metrics, including sparse probing accuracy and feature disentanglement, despite minor drops in reconstruction quality. Our results challenge the assumption that interpretability aligns with an optimal sparsity--fidelity trade-off and suggest that whitening should be considered as a default preprocessing step for SAE training, particularly when interpretability is prioritized over perfect reconstruction.

Sparse autoencoders (SAEs) have emerged as a powerful tool for uncovering interpretable features in large language models (LLMs) through the sparse directions they learn. However, the sheer number of extracted directions makes comprehensive exploration intractable. While conventional embedding techniques such as UMAP can reveal global structure, they suffer from limitations including high-dimensional compression artifacts, overplotting, and misleading neighborhood distortions. In this work, we propose a focused exploration framework that prioritizes curated concepts and their corresponding SAE features over attempts to visualize all available features simultaneously. We present an interactive visualization system that combines topology-based visual encoding with dimensionality reduction to faithfully represent both local and global relationships among selected features. This hybrid approach enables users to investigate SAE behavior through targeted, interpretable subsets, facilitating deeper and more nuanced analysis of concept representation in latent space.

Modern chip physical design relies heavily on Electronic Design Automation (EDA) tools, which often struggle to provide interpretable feedback or actionable guidance for improving routing congestion. In this work, we introduce a Multimodal Large Language Model Assistant (MLLMA) that bridges this gap by not only predicting congestion but also delivering human-interpretable design suggestions. Our method combines automated feature generation through MLLM-guided genetic prompting with an interpretable preference learning framework that models congestion-relevant tradeoffs across visual, tabular, and textual inputs. We compile these insights into a "Design Suggestion Deck" that surfaces the most influential layout features and proposes targeted optimizations. Experiments on the CircuitNet benchmark demonstrate that our approach outperforms existing models on both accuracy and explainability. Additionally, our design suggestion guidance case study and qualitative analyses confirm that the learned preferences align with real-world design principles and are actionable for engineers. This work highlights the potential of MLLMs as interactive assistants for interpretable and context-aware physical design optimization.

Understanding the internal representations of large language models is crucial for ensuring their reliability and safety, with sparse autoencoders (SAEs) emerging as a promising interpretability approach. However, current SAE training methods face feature absorption, where features (or neurons) are absorbed into each other to minimize $L_1$ penalty, making it difficult to consistently identify and analyze model behaviors. We introduce Adaptive Temporal Masking (ATM), a novel training approach that dynamically adjusts feature selection by tracking activation magnitudes, frequencies, and reconstruction contributions to compute importance scores that evolve over time. ATM applies a probabilistic masking mechanism based on statistical thresholding of these importance scores, creating a more natural feature selection process. Through extensive experiments on the Gemma-2-2b model, we demonstrate that ATM achieves substantially lower absorption scores compared to existing methods like TopK and JumpReLU SAEs, while maintaining excellent reconstruction quality. These results establish ATM as a principled solution for learning stable, interpretable features in neural networks, providing a foundation for more reliable model analysis.

Ranking online reviews by their intrinsic quality is a critical task for e-commerce platforms and information services, impacting user experience and business outcomes. However, quality is a domain-dependent and dynamic concept, making its assessment a formidable challenge. Traditional methods relying on hand-crafted features are unscalable across domains and fail to adapt to evolving content patterns, while modern deep learning approaches often produce black-box models that lack interpretability and may prioritize semantics over quality. To address these challenges, we propose AutoQual, an LLM-based agent framework that automates the discovery of interpretable features. While demonstrated on review quality assessment, AutoQual is designed as a general framework for transforming tacit knowledge embedded in data into explicit, computable features. It mimics a human research process, iteratively generating feature hypotheses through reflection, operationalizing them via autonomous tool implementation, and accumulating experience in a persistent memory. We deploy our method on a large-scale online platform with a billion-level user base. Large-scale A/B testing confirms its effectiveness, increasing average reviews viewed per user by 0.79% and the conversion rate of review readers by 0.27%.

Responsible use of Authorship Verification (AV) systems not only requires high accuracy but also interpretable solutions. More importantly, for systems to be used to make decisions with real-world consequences requires the model's prediction to be explainable using interpretable features that can be traced to the original texts. Neural methods achieve high accuracies, but their representations lack direct interpretability. Furthermore, LLM predictions cannot be explained faithfully -- if there is an explanation given for a prediction, it doesn't represent the reasoning process behind the model's prediction. In this paper, we introduce Residualized Similarity (RS), a novel method that supplements systems using interpretable features with a neural network to improve their performance while maintaining interpretability. Authorship verification is fundamentally a similarity task, where the goal is to measure how alike two documents are. The key idea is to use the neural network to predict a similarity residual, i.e. the error in the similarity predicted by the interpretable system. Our evaluation across four datasets shows that not only can we match the performance of state-of-the-art authorship verification models, but we can show how and to what degree the final prediction is faithful and interpretable.