Abstract-- Solving complex real-world control tasks often takes multiple tries: if we fail at first, we reflect on what went wrong, and change our strategy accordingly to avoid making the same mistake. In robotics, Vision-Language-Action models (VLAs) offer a promising path towards solving complex control tasks, but lack the ability to contextually and dynamically readjust behavior when they fail to accomplish a task. In this work, we introduce Learning from Inference-Time Execution (LITEN), which connects a VLA low-level policy to a high-level VLM that conditions on past experiences by including them in-context, allowing it to learn the affordances and capabilities of the low-level VLA. Our approach iterates between a reasoning phase that generates and executes plans for the low-level VLA, and an assessment phase that reflects on the resulting execution and draws useful conclusions to be included in future reasoning contexts. Unlike similar approaches to self-refinement in non-robotics domains, LITEN must reflect on unstructured real-world robot trajectories (e.g., raw videos), which requires structured guiderails during assessment. Our experimental results demonstrate LITEN is able to effectively learn from past experience to generate plans that use high-affordance instructions to accomplish long-horizon tasks. Robotic foundation models based on powerful pre-trained vision-language model (VLM) backbones have the potential to combine both the semantic and common-sense problem-solving abilities of LLMs and the flexible and dexterous end-to-end control capabilities of learned policies [1], [2], [3], [4], [5]. However, current robotic foundation models, most notably Vision-Language-Action models (VLAs), have primarily been studied in "single shot" settings, where they are evaluated on their ability to follow individual user commands. A practical robotic system needs to also plan through complex behaviors and, perhaps most importantly, adjust its behavior based on context and perceived capabilities. For example, if the robot needs to open a latched container, it might try to unlatch it in a particular way, and if that fails, it should modify its strategy and try a different approach. This kind of in-context adaptation has been observed as an emergent behavior in LLMs [6], [7], [8], but has proven difficult to enable in the robotics domain with current VLAs.

The accessibility of legal information remains a constant challenge, particularly for laypersons seeking to understand and apply complex institutional texts. While the European Union provides open access to legislation, parliamentary responses, and regulatory documents, these resources can be challenging for laypeople to explore. In this paper, we introduce EUDial, a proactive multi-turn dialogue dataset constructed from 204 blogs curated by the Citizens' Enquiries Unit (AskEP) of the European Parliamentary Research Service. EUDial contains 880 dialogue turns (averaging 4.3 turns per dialogue), where each dialogue includes initial questions, structured answers, and follow-up questions. Beyond dataset construction, we propose the LexGuide framework that leverages retrieval-augmented generation with hierarchical topic organization to structure dialogue progression, ensuring both comprehensive coverage of legal aspects and coherence across conversational turns. The results demonstrate that proactive, structured navigation closes the gap between the availability of legal information and citizen comprehension, establishing EUDial and LexGuide as practical resources for advancing proactive legal dialogue systems.

The recent boom of large pre-trained models witnesses remarkable success in developing foundation models (FMs) for time series forecasting. Despite impressive performance across diverse downstream forecasting tasks, existing time series FMs possess massive network architectures and require substantial pre-training on large-scale datasets, which significantly hinders their deployment in resource-constrained environments. In response to this growing tension between versatility and affordability, we propose SEMPO, a novel lightweight foundation model that requires pretraining on relatively small-scale data, yet exhibits strong general time series forecasting. Concretely, SEMPO comprises two key modules: 1) energy-aware SpEctral decomposition module, that substantially improves the utilization of pre-training data by modeling not only the high-energy frequency signals but also the low-energy yet informative frequency signals that are ignored in current methods; and 2) Mixture-of-PrOmpts enabled Transformer, that learns heterogeneous temporal patterns through small dataset-specific prompts and adaptively routes time series tokens to prompt-based experts for parameter-efficient model adaptation across different datasets and domains. Equipped with these modules, SEMPO significantly reduces both pre-training data scale and model size, while achieving strong generalization. Extensive experiments on two large-scale benchmarks covering 16 datasets demonstrate the superior performance of SEMPO in both zero-shot and few-shot forecasting scenarios compared with state-of-the-art methods. Code and data are available at https://github.com/mala-lab/SEMPO.

Nowadays, Large Language Models (LLMs) are able to propose rules in natural language, overcoming constrains of a predefined predicate space inherent in traditional rule learning. However, existing methods using LLMs often overlook the combination effects of rules, and the potential of coupling LLMs with probabilistic rule learning to ensure robust inference is not fully explored. To address this gap, we introduce RLIE, a unified framework that integrates LLMs with probabilistic modeling to learn a set of probabilistic rules. The RLIE framework comprises four stages: (1) Rule generation, where a LLM proposes and filters candidate rules; (2) Logistic regression, which learns the probabilistic weights of the rules for global selection and calibration; (3) Iterative refinement, which continuously optimizes the rule set based on prediction errors; and (4) Evaluation, which compares the performance of the weighted rule set as a direct classifier against various methods of injecting the rules into an LLM. Generated rules are the evaluated with different inference strategies on multiple real-world datasets. While applying rules directly with corresponding weights brings us superior performance, prompting LLMs with rules, weights and classification results from the logistic model will surprising degrade the performance. This result aligns with the observation that LLMs excel at semantic generation and interpretation but are less reliable at fine-grained, controlled probabilistic integration. Our work investigates the potentials and limitations of using LLMs for inductive reasoning tasks, proposing a unified framework which integrates LLMs with classic probabilistic rule combination methods, paving the way for more reliable neuro-symbolic reasoning systems. In data-driven applications and scientific discovery, the goal is not merely to predict outcomes, but to construct a set of verifiable, reusable, and composable theories(Zhou et al., 2024; Y ang et al., 2024a; Minh et al., 2022). These theories can enable explainable, auditable decisions while driving the discovery of new knowledge and underlying structures(Y ang et al., 2023; 2024b). These theories can be expressed in formal, structural statements(Cohen et al., 1995; Cropper & Morel, 2021) or natural language hypotheses(Zhou et al., 2024), and they share a common characteristic: they are declarative, testable, and self-contained discriminative patterns that yield predictions verifiable by external evidence In this paper, we do not distinguish between the terms "rule" and "hypothesis", and will use "rule" throughout the text for consistency.

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.

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.

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.