Digital Adoption Platforms (DAPs) have become essential tools for helping employees navigate complex enterprise software such as CRM, ERP, or HRMS systems. Companies like LemonLearning have shown how digital guidance can reduce training costs and accelerate onboarding. However, building and maintaining these interactive guides still requires extensive manual effort. Leveraging Large Language Models as virtual assistants is an appealing alternative, yet without a structured understanding of the target software, LLMs often hallucinate and produce unreliable answers. Moreover, most production-grade LLMs are black-box APIs, making fine-tuning impractical due to the lack of access to model weights. In this work, we introduce a Graph-based Retrieval-Augmented Generation framework that automatically converts enterprise web applications into state-action knowledge graphs, enabling LLMs to generate grounded and context-aware assistance. The framework was co-developed with the AI enterprise RAKAM, in collaboration with Lemon Learning. We detail the engineering pipeline that extracts and structures software interfaces, the design of the graph-based retrieval process, and the integration of our approach into production DAP workflows. Finally, we discuss scalability, robustness, and deployment lessons learned from industrial use cases.

The success of large language models (LLMs) has motivated formal theories of language generation and learning. We study the framework of \emph{language generation in the limit}, where an adversary enumerates strings from an unknown language $K$ drawn from a countable class, and an algorithm must generate unseen strings from $K$. Prior work showed that generation is always possible, and that some algorithms achieve positive lower density, revealing a \emph{validity--breadth} trade-off between correctness and coverage. We resolve a main open question in this line, proving a tight bound of $1/2$ on the best achievable lower density. We then strengthen the model to allow \emph{partial enumeration}, where the adversary reveals only an infinite subset $C \subseteq K$. We show that generation in the limit remains achievable, and if $C$ has lower density $α$ in $K$, the algorithm's output achieves density at least $α/2$, matching the upper bound. This generalizes the $1/2$ bound to the partial-information setting, where the generator must recover within a factor $1/2$ of the revealed subset's density. We further revisit the classical Gold--Angluin model of \emph{language identification} under partial enumeration. We characterize when identification in the limit is possible -- when hypotheses $M_t$ eventually satisfy $C \subseteq M \subseteq K$ -- and in the process give a new topological formulation of Angluin's characterization, showing that her condition is precisely equivalent to an appropriate topological space having the $T_D$ separation property.

Today's autonomous agents can understand free-form natural language instructions and execute long-horizon tasks in a manner akin to human-level reasoning. These capabilities are mostly driven by large-scale pre-trained foundation models (FMs). However, the approaches with which these models are grounded for human-robot interaction (HRI) perpetuate a master-apprentice model, where the apprentice (embodied agent) passively receives and executes the master's (human's) commands without reciprocal learning. This reactive interaction approach does not capture the co-adaptive dynamics inherent in everyday multi-turn human-human interactions. To address this, we propose a Symbiotic Interactive Learning (SIL) approach that enables both the master and the apprentice to co-adapt through mutual, bidirectional interactions. We formalised SIL as a co-adaptation process within a shared latent task space, where the agent and human maintain joint belief states that evolve based on interaction history. This enables the agent to move beyond reactive execution to proactive clarification, adaptive suggestions, and shared plan refinement. To realise these novel behaviours, we leveraged pre-trained FMs for spatial perception and reasoning, alongside a lightweight latent encoder that grounds the models' outputs into task-specific representations. Furthermore, to ensure stability as the tasks evolve, we augment SIL with a memory architecture that prevents the forgetting of learned task-space representations. We validate SIL on both simulated and real-world embodied tasks, including instruction following, information retrieval, query-oriented reasoning, and interactive dialogues. Demos and resources are public at:~\href{https://linusnep.github.io/SIL/}{https://linusnep.github.io/SIL/}.

Chain-of-Thought (CoT) prompting is a widely used method to improve the reasoning capability of Large Language Models (LLMs). More recently, CoT has been leveraged in Knowledge Distillation (KD) to transfer reasoning capability from a larger LLM to a smaller one. This paper examines the role of CoT in distilling the reasoning capability from larger LLMs to smaller LLMs using white-box KD, analysing its effectiveness in improving the performance of the distilled models for various natural language reasoning and understanding tasks. We conduct white-box KD experiments using LLMs from the Qwen and Llama2 families, employing CoT data from the CoT-Collection dataset. The distilled models are then evaluated on natural language reasoning and understanding tasks from the BIG-Bench-Hard (BBH) benchmark, which presents complex challenges for smaller LLMs. Experimental results demonstrate the role of CoT in improving white-box KD effectiveness, enabling the distilled models to achieve better average performance in natural language reasoning and understanding tasks from BBH.

The widespread adoption of generative models such as Stable Diffusion and ChatGPT has made them increasingly attractive targets for malicious exploitation, particularly through data poisoning. Existing poisoning attacks compromising synthesised data typically either cause broad degradation of generated data or require control over the training process, limiting their applicability in real-world scenarios. In this paper, we introduce a novel data poisoning technique called associative poisoning, which compromises fine-grained features of the generated data without requiring control of the training process. This attack perturbs only the training data to manipulate statistical associations between specific feature pairs in the generated outputs. We provide a formal mathematical formulation of the attack and prove its theoretical feasibility and stealthiness. Empirical evaluations using two state-of-the-art generative models demonstrate that associative poisoning effectively induces or suppresses feature associations while preserving the marginal distributions of the targeted features and maintaining high-quality outputs, thereby evading visual detection. These results suggest that generative systems used in image synthesis, synthetic dataset generation, and natural language processing are susceptible to subtle, stealthy manipulations that compromise their statistical integrity. To address this risk, we examine the limitations of existing defensive strategies and propose a novel countermeasure strategy.

Software Architecture Descriptions (SADs) are essential for managing the inherent complexity of modern software systems. They enable high-level architectural reasoning, guide design decisions, and facilitate effective communication among diverse stakeholders. However, in practice, SADs are often missing, outdated, or poorly aligned with the system's actual implementation. Consequently, developers are compelled to derive architectural insights directly from source code-a time-intensive process that increases cognitive load, slows new developer onboarding, and contributes to the gradual degradation of clarity over the system's lifetime. To address these issues, we propose a semi-automated generation of SADs from source code by integrating reverse engineering (RE) techniques with a Large Language Model (LLM). Our approach recovers both static and behavioral architectural views by extracting a comprehensive component diagram, filtering architecturally significant elements (core components) via prompt engineering, and generating state machine diagrams to model component behavior based on underlying code logic with few-shots prompting. This resulting views representation offer a scalable and maintainable alternative to traditional manual architectural documentation. This methodology, demonstrated using C++ examples, highlights the potent capability of LLMs to: 1) abstract the component diagram, thereby reducing the reliance on human expert involvement, and 2) accurately represent complex software behaviors, especially when enriched with domain-specific knowledge through few-shot prompting. These findings suggest a viable path toward significantly reducing manual effort while enhancing system understanding and long-term maintainability.

In addition they have also shown impressive capabilities in different domains, like coding, science and math. In this short paper, taking math as an example domain, we study the performance of different LLMs across languages. Experimental results show that there exists a non-negligible and consistent gap in the performance of the models across languages. Interestingly, and somewhat against expectations, the gap exists for both high-and low-resource languages. We hope that these results influence further research into cross-lingual capability generalization for next generation LLMs. If it weren't for the fact that they are false! By analyzing one of the standard multilingual math benchmarks (MGSM), we determine that several translation errors are present in the data. Furthermore, the lack of standardized answer extraction from LLM outputs further influences the final results. We propose a method for automatic quality assurance to address the first issue at scale, and give recommendations to address the second one. Combining these two approaches we show that the aforementioned language gap mostly disappears, leading to completely different conclusions from our research. In recent years, large language models' capabilities have expanded in two primary directions: broader language coverage and enhanced performance on complex tasks. On the language dimension, it is now usual for LLMs to support not only high-resource languages languages (e.g. This is a very important and welcome progress direction in order to improve the inclusivity of AI applications and research.

Text offers intuitive access to information. This can, in particular, complement the density of numerical time series, thereby allowing improved interactions with time series models to enhance accessibility and decision-making. While the creation of question-answering datasets and models has recently seen remarkable growth, most research focuses on question answering (QA) on vision and text, with time series receiving minute attention. To bridge this gap, we propose a challenging novel time series QA (TSQA) dataset, QuAnTS, for Question Answering on Time Series data. Specifically, we pose a wide variety of questions and answers about human motion in the form of tracked skeleton trajectories. We verify that the large-scale QuAnTS dataset is well-formed and comprehensive through extensive experiments. Thoroughly evaluating existing and newly proposed baselines then lays the groundwork for a deeper exploration of TSQA using QuAnTS. Additionally, we provide human performances as a key reference for gauging the practical usability of such models. We hope to encourage future research on interacting with time series models through text, enabling better decision-making and more transparent systems.

Creating programs to represent board games can be a time-consuming task. Large Language Models (LLMs) arise as appealing tools to expedite this process, given their capacity to efficiently generate code from simple contextual information. In this work, we propose a method to test how capable three LLMs (Claude, DeepSeek and ChatGPT) are at creating code for board games, as well as new variants of existing games.

This work investigates distillation methods for large language models (LLMs) with the goal of developing compact models that preserve high performance. Several existing approaches are reviewed, with a discussion of their respective strengths and limitations. An improved method based on the ShortGPT approach has been developed, building upon the idea of incorporating iterative evaluation of layer importance. At each step, importance is assessed by measuring performance degradation when individual layers are removed, using a set of representative datasets. This process is combined with further training using a joint loss function based on KL divergence and mean squared error. Experiments on the Qwen2.5-3B model show that the number of layers can be reduced from 36 to 28 (resulting in a 2.47 billion parameter model) with only a 9.7% quality loss, and to 24 layers with an 18% loss. The findings suggest that the middle transformer layers contribute less to inference, underscoring the potential of the proposed method for creating efficient models. The results demonstrate the effectiveness of iterative distillation and fine-tuning, making the approach suitable for deployment in resource-limited settings.