Prompting is a common approach for leveraging LMs in zero-shot settings. However, the underlying mechanisms that enable LMs to perform diverse tasks without task-specific supervision remain poorly understood. Studying the relationship between prompting and the quality of internal representations can shed light on how pre-trained embeddings may support in-context task solving. In this empirical study, we conduct a series of probing experiments on prompt embeddings, analyzing various combinations of prompt templates for zero-shot classification. Our findings show that while prompting affects the quality of representations, these changes do not consistently correlate with the relevance of the prompts to the target task. This result challenges the assumption that more relevant prompts necessarily lead to better representations. We further analyze potential factors that may contribute to this unexpected behavior.

Human communication is motivated: people speak, write, and create content with a particular communicative intent in mind. As a result, information that large language models (LLMs) and AI agents process is inherently framed by humans' intentions and incentives. People are adept at navigating such nuanced information: we routinely identify benevolent or self-serving motives in order to decide what statements to trust. For LLMs to be effective in the real world, they too must critically evaluate content by factoring in the motivations of the source -- for instance, weighing the credibility of claims made in a sales pitch. In this paper, we undertake a comprehensive study of whether LLMs have this capacity for motivational vigilance. We first employ controlled experiments from cognitive science to verify that LLMs' behavior is consistent with rational models of learning from motivated testimony, and find they successfully discount information from biased sources in a human-like manner. We then extend our evaluation to sponsored online adverts, a more naturalistic reflection of LLM agents' information ecosystems. In these settings, we find that LLMs' inferences do not track the rational models' predictions nearly as closely -- partly due to additional information that distracts them from vigilance-relevant considerations. However, a simple steering intervention that boosts the salience of intentions and incentives substantially increases the correspondence between LLMs and the rational model. These results suggest that LLMs possess a basic sensitivity to the motivations of others, but generalizing to novel real-world settings will require further improvements to these models.

Multimodal large language models (MLLMs) achieve strong performance on vision-language tasks, yet their visual processing is opaque. Most black-box evaluations measure task accuracy, but reveal little about underlying mechanisms. Drawing on cognitive psychology, we adapt classic visual search paradigms -- originally developed to study human perception -- to test whether MLLMs exhibit the ``pop-out'' effect, where salient visual features are detected independently of distractor set size. Using controlled experiments targeting colour, size and lighting features, we find that advanced MLLMs exhibit human-like pop-out effects in colour or size-based disjunctive (single feature) search, as well as capacity limits for conjunctive (multiple feature) search. We also find evidence to suggest that MLLMs, like humans, incorporate natural scene priors such as lighting direction into object representations. We reinforce our findings using targeted fine-tuning and mechanistic interpretability analyses. Our work shows how visual search can serve as a cognitively grounded diagnostic tool for evaluating perceptual capabilities in MLLMs.

We propose a rubric-guided, pseudo-labeled, and prompt-driven zero-shot video summarization framework that bridges large language models with structured semantic reasoning. A small subset of human annotations is converted into high-confidence pseudo labels and organized into dataset-adaptive rubrics defining clear evaluation dimensions such as thematic relevance, action detail, and narrative progression. During inference, boundary scenes, including the opening and closing segments, are scored independently based on their own descriptions, while intermediate scenes incorporate concise summaries of adjacent segments to assess narrative continuity and redundancy. This design enables the language model to balance local salience with global coherence without any parameter tuning. Across three benchmarks, the proposed method achieves stable and competitive results, with F1 scores of 57.58 on SumMe, 63.05 on TVSum, and 53.79 on QFVS, surpassing zero-shot baselines by +0.85, +0.84, and +0.37, respectively. These outcomes demonstrate that rubric-guided pseudo labeling combined with contextual prompting effectively stabilizes LLM-based scoring and establishes a general, interpretable, and training-free paradigm for both generic and query-focused video summarization.

Autonomous AI agents, capable of complex planning and action, represent a significant technological evolution beyond current generative tools. As these systems become integrated into political and economic life, their distribution and capabilities will be highly consequential. This paper introduces and explores "agentic inequality" - the potential disparities in power, opportunity, and outcomes stemming from differential access to, and capabilities of, AI agents. We analyse the dual potential of this technology, exploring how agents could both exacerbate existing divides and, under the right conditions, serve as a powerful equalising force. To this end, the paper makes three primary contributions. First, it establishes an analytical framework by delineating the three core dimensions through which this inequality can manifest: disparities in the availability, quality, and quantity of agents. Second, it argues that agentic inequality is distinct from prior technological divides. Unlike tools that primarily augment human abilities, agents act as autonomous delegates, creating novel power asymmetries through scalable goal delegation and direct agent-to-agent competition that are poised to reshape outcomes across economic and socio-political spheres. Finally, it provides a systematic analysis of the technical and socioeconomic drivers - from model release strategies to market incentives - that will shape the distribution of agentic power, concluding with a research agenda for navigating the complex governance challenges ahead.

Reinforcement learning (RL) algorithms are highly sensitive to reward function specification, which remains a central challenge limiting their broad applicability. We present ARM-FM: Automated Reward Machines via Foundation Models, a framework for automated, compositional reward design in RL that leverages the high-level reasoning capabilities of foundation models (FMs). Reward machines (RMs) -- an automata-based formalism for reward specification -- are used as the mechanism for RL objective specification, and are automatically constructed via the use of FMs. The structured formalism of RMs yields effective task decompositions, while the use of FMs enables objective specifications in natural language. Concretely, we (i) use FMs to automatically generate RMs from natural language specifications; (ii) associate language embeddings with each RM automata-state to enable generalization across tasks; and (iii) provide empirical evidence of ARM-FM's effectiveness in a diverse suite of challenging environments, including evidence of zero-shot generalization.

Optimization lies at the core of modern deep learning, yet existing methods often face a fundamental trade-off between adapting to problem geometry and leveraging curvature utilization. Steepest descent algorithms adapt to different geometries through norm choices but remain strictly first-order, whereas quasi-Newton and adaptive optimizers incorporate curvature information but are restricted to Frobenius geometry, limiting their applicability across diverse architectures. In this work, we propose a unified framework generalizing steepest descent, quasi-Newton methods, and adaptive methods through the novel notion of preconditioned matrix norms. This abstraction reveals that widely used optimizers such as SGD and Adam, as well as more advanced approaches like Muon and KL-Shampoo, and recent hybrids including SOAP and SPlus, all emerge as special cases of the same principle. Within this framework, we provide the first systematic treatment of affine and scale invariance in the matrix-parameterized setting, establishing necessary and sufficient conditions under generalized norms. Building on this foundation, we introduce two new methods, $\texttt{MuAdam}$ and $\texttt{MuAdam-SANIA}$, which combine the spectral geometry of Muon with Adam-style preconditioning. Our experiments demonstrate that these optimizers are competitive with, and in some cases outperform, existing state-of-the-art methods. Our code is available at https://github.com/brain-lab-research/LIB/tree/quasi_descent

Why do thinking language models like DeepSeek R1 outperform their base counterparts? Despite consistent performance gains, it remains unclear to what extent thinking models learn entirely new reasoning capabilities or repurpose pre-existing base model ones. In this work, we propose a hybrid model where we activate reasoning mechanisms in base models at the right time to elicit thinking-model-level reasoning chains, implying that thinking models exploit already existing capabilities. To ground our analysis, we introduce an unsupervised, bottom-up approach for uncovering human-interpretable reasoning behaviors in thinking models. This approach provides an unbiased method to discover reasoning behaviors without imposing manual or LLM-derived assumptions. Across three base and four thinking models, using GSM8K and MATH500, our hybrid model recovers up to 91% of the performance gap to thinking models without any weight updates while steering only 12% of tokens. Concretely, our empirical setup provides a simple, causal way to test the effectiveness of existing reasoning mechanisms in base models by invoking them directly and measuring the resulting task performance. More broadly, these results reframe our understanding of how thinking models are trained: pre-training is when models acquire most of their reasoning mechanisms, and post-training teaches efficient deployment of these mechanisms at the right time, enabling efficient use of their inference-time compute.

Context utilisation, the ability of Language Models (LMs) to incorporate relevant information from the provided context when generating responses, remains largely opaque to users, who cannot determine whether models draw from parametric memory or provided context, nor identify which specific context pieces inform the response. Highlight explanations (HEs) offer a natural solution as they can point the exact context pieces and tokens that influenced model outputs. However, no existing work evaluates their effectiveness in accurately explaining context utilisation. We address this gap by introducing the first gold standard HE evaluation framework for context attribution, using controlled test cases with known ground-truth context usage, which avoids the limitations of existing indirect proxy evaluations. To demonstrate the framework's broad applicability, we evaluate four HE methods -- three established techniques and MechLight, a mechanistic interpretability approach we adapt for this task -- across four context scenarios, four datasets, and five LMs. Overall, we find that MechLight performs best across all context scenarios. However, all methods struggle with longer contexts and exhibit positional biases, pointing to fundamental challenges in explanation accuracy that require new approaches to deliver reliable context utilisation explanations at scale.

Multi-head Latent Attention (MLA) significantly reduces KVCache memory usage in Large Language Models while introducing substantial computational overhead and intermediate variable expansion. This poses challenges for efficient hardware implementation -- especially during the decode phase. This paper introduces Ascend MLA (AMLA), a high-performance kernel specifically optimized for Huawei's Ascend NPUs. AMLA is built on two core innovations: (1) A novel FlashAttention-based algorithm that replaces floating-point multiplications with integer additions for output block rescaling, leveraging binary correspondence between FP32 and INT32 representations; (2) A Preload Pipeline strategy with hierarchical tiling that maximizes FLOPS utilization: the Preload Pipeline achieves Cube-bound performance, while hierarchical tiling overlaps data movement and computation within the Cube core. Experiments show that on Ascend 910 NPUs (integrated in CloudMatrix384), AMLA achieves up to 614 TFLOPS, reaching 86.8% of the theoretical maximum FLOPS, outperforming the state-of-the-art open-source FlashMLA implementation, whose FLOPS utilization is up to 66.7% on NVIDIA H800 SXM5. The AMLA kernel has been integrated into Huawei's CANN and will be released soon.