While modern deep neural networks achieve impressive performance in vision tasks, they remain opaque in their decision processes, risking unwarranted trust, undetected biases and unexpected failures. We propose cluster paths, a post-hoc interpretability method that clusters activations at selected layers and represents each input as its sequence of cluster IDs. To assess these cluster paths, we introduce four metrics: path complexity (cognitive load), weighted-path purity (class alignment), decision-alignment faithfulness (predictive fidelity), and path agreement (stability under perturbations). In a spurious-cue CIFAR-10 experiment, cluster paths identify color-based shortcuts and collapse when the cue is removed. On a five-class CelebA hair-color task, they achieve 90% faithfulness and maintain 96% agreement under Gaussian noise without sacrificing accuracy. Scaling to a Vision Transformer pretrained on ImageNet, we extend cluster paths to concept paths derived from prompting a large language model on minimal path divergences. Finally, we show that cluster paths can serve as an effective out-of-distribution (OOD) detector, reliably flagging anomalous samples before the model generates over-confident predictions. Cluster paths uncover visual concepts, such as color palettes, textures, or object contexts, at multiple network depths, demonstrating that cluster paths scale to large vision models while generating concise and human-readable explanations.

We present a novel framework that improves the reliability of LLM judges by selectively augmenting LLM with auxiliary evaluation dimensions. Existing LLM judges often miss crucial evaluation dimensions because they fail to recognize the implicit standards underlying human assessments. To address this challenge, we propose the Auto-Prompt Ensemble (APE), an adaptive framework that automatically learns evaluation dimensions from its failure cases. APE incorporates a confidence-based ensemble mechanism to decide when to adopt the judgments from additional evaluation dimensions through a novel confidence estimation approach called Collective Confidence. Extensive experiments demonstrate that APE improves the reliability of LLM Judge across diverse standard benchmarks. For instance, APE enhances GPT-4o agreement rate on Reward Bench from 87.2% to 90.5% in the zero-shot setting. Overall, APE provides a principled approach for LLM Judge to leverage test-time computation, and bridge the evaluation gap between human and LLM judges.

The quality and experience of mobile communication have significantly improved with the introduction of 5G, and these improvements are expected to continue beyond the 5G era. However, vulnerabilities in control-plane protocols, such as Radio Resource Control (RRC) and Non-Access Stratum (NAS), pose significant security threats, such as Blind Denial of Service (DoS) attacks. Despite the availability of existing anomaly detection methods that leverage rule-based systems or traditional machine learning methods, these methods have several limitations, including the need for extensive training data, predefined rules, and limited explainability. Addressing these challenges, we propose a novel anomaly detection framework that leverages the capabilities of Large Language Models (LLMs) in zero-shot mode with unordered data and short natural language attack descriptions within the Open Radio Access Network (O-RAN) architecture. We analyse robustness to prompt variation, demonstrate the practicality of automating the attack descriptions and show that detection quality relies on the semantic completeness of the description rather than its phrasing or length. We utilise an RRC/NAS dataset to evaluate the solution and provide an extensive comparison of open-source and proprietary LLM implementations to demonstrate superior performance in attack detection. We further validate the practicality of our framework within O-RAN's real-time constraints, illustrating its potential for detecting other Layer-3 attacks.

Automatic chord recognition (ACR) via deep learning models has gradually achieved promising recognition accuracy, yet two key challenges remain. First, prior work has primarily focused on audio-domain ACR, while symbolic music (e.g., score) ACR has received limited attention due to data scarcity. Second, existing methods still overlook strategies that are aligned with human music analytical practices. To address these challenges, we make two contributions: (1) we introduce POP909-CL, an enhanced version of POP909 dataset with tempo-aligned content and human-corrected labels of chords, beats, keys, and time signatures; and (2) We propose BACHI, a symbolic chord recognition model that decomposes the task into different decision steps, namely boundary detection and iterative ranking of chord root, quality, and bass (inversion). This mechanism mirrors the human ear-training practices. Experiments demonstrate that BACHI achieves state-of-the-art chord recognition performance on both classical and pop music benchmarks, with ablation studies validating the effectiveness of each module.

Generative AI leaderboards are central to evaluating model capabilities, but remain vulnerable to manipulation. Among key adversarial objectives is rank manipulation, where an attacker must first deanonymize the models behind displayed outputs -- a threat previously demonstrated and explored for large language models (LLMs). We show that this problem can be even more severe for text-to-image leaderboards, where deanonymization is markedly easier. Using over 150,000 generated images from 280 prompts and 19 diverse models spanning multiple organizations, architectures, and sizes, we demonstrate that simple real-time classification in CLIP embedding space identifies the generating model with high accuracy, even without prompt control or historical data. We further introduce a prompt-level separability metric and identify prompts that enable near-perfect deanonymization. Our results indicate that rank manipulation in text-to-image leaderboards is easier than previously recognized, underscoring the need for stronger defenses.

Accurate time - series predictions in machine learning are heavily influenced by the selection of appropriate input time length and sampling rate. This paper introduces ATLO - ML, an adaptive time - length optimization system that automatically determines the optimal input time length and sampling rate based on user - defined output time length. The system provides a flexible approach to time - series data pre - processing, dynamically adjusting these parameters to enhance predictive performance. ATLO - ML is validated using air quality datasets, including both GAMS - dataset and proprietary data collected from a data center, both in time series format. Results demonstrate that utilizing the optimized time length and sampling rate significantly improves the accuracy of machine learning models compared to fixed time lengths. ATLO - ML shows potential for generalization across various time - sensitive applications, offering a robust solution for optimizing temporal input parameters in machine learning workflows .

Algorithmic efficiency techniques such as distillation (\cite{hinton2015distillation}) are useful in improving model quality without increasing serving costs, provided a larger teacher model is available for a smaller student model to learn from during training. Standard distillation methods are limited to only forcing the student to match the teacher's outputs. Given the costs associated with training a large model, we believe we should be extracting more useful information from a teacher model than by just making the student match the teacher's outputs. In this paper, we introduce \guide (Guided Initialization and Distillation of Embeddings). \guide can be considered a distillation technique that forces the student to match the teacher in the parameter space. Using \guide we show 25-26\% reduction in the teacher-student quality gap when using large student models (400M - 1B parameters) trained on $\approx$ 20B tokens. We also present a thorough analysis demonstrating that \guide can be combined with knowledge distillation with near additive improvements. Furthermore, we show that applying \guide alone leads to substantially better model quality than applying knowledge distillation by itself. Most importantly, \guide introduces no training or inference overhead and hence any model quality gains from our method are virtually free.

Large Language Models (LLMs) have achieved remarkable success through imitation learning on vast text corpora, but this paradigm creates a training-generation gap and limits robust reasoning. Reinforcement learning (RL) offers a more data-efficient solution capable of bridging this gap, yet its application has been constrained by a critical data bottleneck: existing RL datasets are orders of magnitude smaller and less diverse than web-scale pre-training corpora. To address this, we introduce the Webscale-RL pipeline, a scalable data engine that systematically converts large-scale pre-training documents into millions of diverse, verifiable question-answer pairs for RL. Using this pipeline, we construct the Webscale-RL dataset, containing 1.2 million examples across more than 9 domains. Our experiments show that the model trained on this dataset significantly outperforms continual pretraining and strong data refinement baselines across a suite of benchmarks. Notably, RL training with our dataset proves substantially more efficient, achieving the performance of continual pre-training with up to 100$\times$ fewer tokens. Our work presents a viable path toward scaling RL to pre-training levels, enabling more capable and efficient language models.

What Y ou Don't Know Can Hurt Y ou: How Well do Latent Safety Filters Understand Partially Observable Safety Constraints? Figure 1: We design a series of controlled experiments to test how latent safety filters behave under partially observable constraints. Left: We find that safety filters that rely on latent state representations trained-on and deployed-with only RGB inputs behave unreliably when they must enforce constraints, such as temperature limits, that are not easily observable. Right: Training with rich, safety-relevant multimodal supervision shapes the latent state representation to enable safe control (e.g., lifting the pan before overheating), even when the robot is deployed with only RGB inputs at runtime. Abstract-- Safe control techniques, such as Hamilton-Jacobi reachability, provide principled methods for synthesizing safety-preserving robot policies but typically assume hand-designed state spaces and full observability. Recent work has relaxed these assumptions via latent-space safe control, where state representations and dynamics are learned jointly through world models that reconstruct future high-dimensional observations (e.g., RGB images) from current observations and actions. This enables safety constraints that are difficult to specify analytically (e.g., spilling) to be framed as classification problems in latent space, allowing controllers to operate directly from raw observations. However, these methods assume that safety-critical features are observable in the learned latent state. We ask: when are latent state spaces sufficient for safe control?

Attention sinks and compression valleys have attracted significant attention as two puzzling phenomena in large language models, but have been studied in isolation. In this work, we present a surprising connection between attention sinks and compression valleys, tracing both to the formation of massive activations in the residual stream. We prove theoretically that massive activations necessarily produce representational compression and establish bounds on the resulting entropy reduction. Through experiments across several models (410M-120B parameters), we confirm that when the beginning-of-sequence token develops extreme activation norms in the middle layers, both compression valleys and attention sinks emerge simultaneously. This unified view motivates us to propose the Mix-Compress-Refine theory of information flow, as an attempt to explain how LLMs organize their computation in depth by controlling attention and representational compression via massive activations. Specifically, we posit that Transformer-based LLMs process tokens in three distinct phases: (1) broad mixing in the early layers, (2) compressed computation with limited mixing in the middle layers, and (3) selective refinement in the late layers. Our framework helps explain why embedding tasks perform best at intermediate layers, whereas generation tasks benefit from full-depth processing, clarifying differences in task-dependent representations. Large Language Models (LLMs) have become remarkably capable, yet how they process information through their layers remains poorly understood. Two phenomena have particularly puzzled researchers: attention sinks, where attention heads mysteriously collapse their focus onto semantically uninformative tokens (Xiao et al., 2024), and compression valleys, where intermediate representations show unexpectedly low entropy despite the model's high-dimensional space (Skean et al., 2025). These phenomena appear paradoxical: why would powerful models waste attention on meaningless tokens, and why would representations compress in the middle of processing? Previous work has explained attention sinks through positional biases (Gu et al., 2025) and over-mixing prevention (Barbero et al., 2025a), while compression valleys have been explained through an information bottleneck theory (Skean et al., 2025). However, the precise reasons why they emerge remain unclear and no formal link has been established between them.