Multimodal large language models (MLLMs) exhibit a pronounced preference for textual inputs when processing vision-language data, limiting their ability to reason effectively from visual evidence. Unlike prior studies that attribute this text bias to external factors such as data imbalance or instruction tuning, we propose that the bias originates from the model's internal architecture. Specifically, we hypothesize that visual key vectors (Visual Keys) are out-of-distribution (OOD) relative to the text key space learned during language-only pretraining. Consequently, these visual keys receive systematically lower similarity scores during attention computation, leading to their under-utilization in the context representation. To validate this hypothesis, we extract key vectors from LLaVA and Qwen2.5-VL and analyze their distributional structures using qualitative (t-SNE) and quantitative (Jensen-Shannon divergence) methods. The results provide direct evidence that visual and textual keys occupy markedly distinct subspaces within the attention space. The inter-modal divergence is statistically significant, exceeding intra-modal variation by several orders of magnitude. These findings reveal that text bias arises from an intrinsic misalignment within the attention key space rather than solely from external data factors.

As LLMs occupy an increasingly important role in society, they are more and more confronted with questions that require them not only to draw on their general knowledge but also to align with certain human value systems. Therefore, studying the alignment of LLMs with human values has become a crucial field of inquiry. Prior work, however, mostly focuses on evaluating the alignment of fully trained models, overlooking the training dynamics by which models learn to express human values. In this work, we investigate how and at which stage value alignment arises during the course of a model's post-training. Our analysis disentangles the effects of post-training algorithms and datasets, measuring both the magnitude and time of value drifts during training. Experimenting with Llama-3 and Qwen-3 models of different sizes and popular supervised fine-tuning (SFT) and preference optimization datasets and algorithms, we find that the SFT phase generally establishes a model's values, and subsequent preference optimization rarely re-aligns these values. Furthermore, using a synthetic preference dataset that enables controlled manipulation of values, we find that different preference optimization algorithms lead to different value alignment outcomes, even when preference data is held constant. Our findings provide actionable insights into how values are learned during post-training and help to inform data curation, as well as the selection of models and algorithms for preference optimization to improve model alignment to human values.

Authorizing Large Language Model driven agents to dynamically invoke tools and access protected resources introduces significant risks, since current methods for delegating authorization grant overly broad permissions and give access to tools allowing agents to operate beyond the intended task scope. We introduce and assess a delegated authorization model enabling authorization servers to semantically inspect access requests to protected resources, and issue access tokens constrained to the minimal set of scopes necessary for the agents' assigned tasks. Given the unavailability of datasets centered on delegated authorization flows, particularly including both semantically appropriate and inappropriate scope requests for a given task, we introduce ASTRA, a dataset and data generation pipeline for benchmarking semantic matching between tasks and scopes. Our experiments show both the potential and current limitations of model-based matching, particularly as the number of scopes needed for task completion increases. Our results highlight the need for further research into semantic matching techniques enabling intent-aware authorization for multi-agent and tool-augmented applications, including fine-grained control, such as Task-Based Access Control (TBAC).

Large language models (LLMs) have demonstrated exceptional capabilities across multiple domains by leveraging massive pre-training and curated fine-tuning data. However, in data-sensitive fields such as healthcare, the lack of high-quality, domain-specific training corpus hinders LLMs' adaptation for specialized applications. Meanwhile, domain experts have distilled domain wisdom into ontology rules, which formalize relationships among concepts and ensure the integrity of knowledge management repositories. Viewing LLMs as implicit repositories of human knowledge, we propose Evontree--a novel framework that leverages a small set of high-quality ontology rules to systematically extract, validate, and enhance domain knowledge within LLMs, without requiring extensive external datasets. Specifically, Evontree extracts domain ontology from raw models, detects inconsistencies using two core ontology rules, and reinforces the refined knowledge via self-distilled fine-tuning. Extensive experiments on medical QA benchmarks with Llama3-8B-Instruct and Med42-v2 demonstrate consistent outperformance over both unmodified models and leading supervised baselines, achieving up to a 3.7% improvement in accuracy. These results confirm the effectiveness, efficiency, and robustness of our approach for low-resource domain adaptation of LLMs.

Recent large language model (LLM) research has undergone an architectural shift from encoder-decoder modeling to nowadays the dominant decoder-only modeling. This rapid transition, however, comes without a rigorous comparative analysis especially from the scaling perspective, raising concerns that the potential of encoder-decoder models may have been overlooked. To fill this gap, we revisit encoder-decoder LLM (RedLLM), enhancing it with recent recipes from decoder-only LLM (DecLLM). We conduct a comprehensive comparison between RedLLM, pretrained with prefix language modeling (LM), and DecLLM, pre-trained with causal LM, at different model scales, ranging from 150M to 8B. Using RedPajama V1 (1.6T tokens) for pretraining and FLAN for instruction tuning, our experiments show that RedLLM produces compelling scaling properties and surprisingly strong performance. While DecLLM is overall more compute-optimal during pretraining, RedLLM demonstrates comparable scaling and context length extrapolation capabilities. After instruction tuning, RedLLM achieves comparable and even better results on various downstream tasks while enjoying substantially better inference efficiency. We hope our findings could inspire more efforts on re-examining RedLLM, unlocking its potential for developing powerful and efficient LLMs. A crucial lesson from the past decade for modeling is to design scalable and universal architectures being capable of handling different tasks (modalities) and automatically learning from massive unlabeled data, as remarked by the tremendous success of large language models (LLMs) in various research areas, particularly natural language processing (OpenAI, 2023; Gemini et al., 2024).

The electricity sector transition requires substantial increases in residential demand response capacity, yet Home Energy Management Systems (HEMS) adoption remains limited by user interaction barriers requiring translation of everyday preferences into technical parameters. While large language models have been applied to energy systems as code generators and parameter extractors, no existing implementation deploys LLMs as autonomous coordinators managing the complete workflow from natural language input to multi-appliance scheduling. This paper presents an agentic AI HEMS where LLMs autonomously coordinate multi-appliance scheduling from natural language requests to device control, achieving optimal scheduling without example demonstrations. A hierarchical architecture combining one orchestrator with three specialist agents uses the ReAct pattern for iterative reasoning, enabling dynamic coordination without hardcoded workflows while integrating Google Calendar for context-aware deadline extraction. Evaluation across three open-source models using real Austrian day-ahead electricity prices reveals substantial capability differences. Llama-3.3-70B successfully coordinates all appliances across all scenarios to match cost-optimal benchmarks computed via mixed-integer linear programming, while other models achieve perfect single-appliance performance but struggle to coordinate all appliances simultaneously. Progressive prompt engineering experiments demonstrate that analytical query handling without explicit guidance remains unreliable despite models' general reasoning capabilities. We open-source the complete system including orchestration logic, agent prompts, tools, and web interfaces to enable reproducibility, extension, and future research.

Large Language Models (LLMs) face significant inference latency challenges stemming from their autoregressive design and large size. To address this, speculative decoding emerges as a solution, enabling the simultaneous generation and validation of multiple tokens. While recent approaches like EAGLE-2 and EAGLE-3 improve speculative decoding using dynamic tree structures, they often neglect the impact of crucial system variables such as GPU devices and batch sizes. Therefore, we introduce a new dynamic tree decoding approach called CAST that takes into account inference costs, including factors such as GPU configurations and batch sizes, to dynamically refine the tree structure. Through comprehensive experimentation across six diverse tasks and utilizing six distinct LLMs, our methodology demonstrates remarkable results, achieving speeds up to 5.2 times faster than conventional decoding methods. Moreover, it generally outperforms existing state-of-the-art techniques from 5% to 20%.

Reinforcement Learning with Verifiable Rewards (RLVR) is a promising approach for enhancing agentic deep search. However, its application is often hindered by low \textbf{Reward Density} in deep search scenarios, where agents expend significant exploratory costs for infrequent and often null final rewards. In this paper, we formalize this challenge as the \textbf{Reward Density Optimization} problem, which aims to improve the reward obtained per unit of exploration cost. This paper introduce \textbf{InfoFlow}, a systematic framework that tackles this problem from three aspects. 1) \textbf{Subproblem decomposition}: breaking down long-range tasks to assign process rewards, thereby providing denser learning signals. 2) \textbf{Failure-guided hints}: injecting corrective guidance into stalled trajectories to increase the probability of successful outcomes. 3) \textbf{Dual-agent refinement}: employing a dual-agent architecture to offload the cognitive burden of deep exploration. A refiner agent synthesizes the search history, which effectively compresses the researcher's perceived trajectory, thereby reducing exploration cost and increasing the overall reward density. We evaluate InfoFlow on multiple agentic search benchmarks, where it significantly outperforms strong baselines, enabling lightweight LLMs to achieve performance comparable to advanced proprietary LLMs.

This paper investigates the structure of linear operators introduced in Hernandez et al. [2023] that decode specific relational facts in transformer language models. We extend their single-relation findings to a collection of relations and systematically chart their organization. We show that such collections of relation decoders can be highly compressed by simple order-3 tensor networks without significant loss in decoding accuracy. To explain this surprising redundancy, we develop a cross-evaluation protocol, in which we apply each linear decoder operator to the subjects of every other relation. Our results reveal that these linear maps do not encode distinct relations, but extract recurring, coarse-grained semantic properties (e.g., country of capital city and country of food are both in the country-of-X property). This property-centric structure clarifies both the operators' compressibility and highlights why they generalize only to new relations that are semantically close. Our findings thus interpret linear relational decoding in transformer language models as primarily property-based, rather than relation-specific.

Inference latency stands as a critical bottleneck in the large-scale deployment of Large Language Models (LLMs). Speculative decoding methods have recently shown promise in accelerating inference without compromising the output distribution. However, existing work typically relies on a dualistic draft-verify framework and lacks rigorous theoretical grounding. In this paper, we introduce a novel \emph{polybasic} speculative decoding framework, underpinned by a comprehensive theoretical analysis. Specifically, we prove a fundamental theorem that characterizes the optimal inference time for multi-model speculative decoding systems, shedding light on how to extend beyond the dualistic approach to a more general polybasic paradigm. Through our theoretical investigation of multi-model token generation, we expose and optimize the interplay between model capabilities, acceptance lengths, and overall computational cost. Our framework supports both standalone implementation and integration with existing speculative techniques, leading to accelerated performance in practice. Experimental results across multiple model families demonstrate that our approach yields speedup ratios ranging from $3.31\times$ to $4.01\times$ for LLaMA2-Chat 7B, up to $3.87 \times$ for LLaMA3-8B, up to $4.43 \times$ for Vicuna-7B and up to $3.85 \times$ for Qwen2-7B -- all while preserving the original output distribution. We release our theoretical proofs and implementation code to facilitate further investigation into polybasic speculative decoding.