The rapid emergence of foundation models, particularly Large Language Models (LLMs) and Vision-Language Models (VLMs), has introduced a transformative paradigm in robotics. These models offer powerful capabilities in semantic understanding, high-level reasoning, and cross-modal generalization, enabling significant advances in perception, planning, control, and human-robot interaction. This critical review provides a structured synthesis of recent developments, categorizing applications across simulation-driven design, open-world execution, sim-to-real transfer, and adaptable robotics. Unlike existing surveys that emphasize isolated capabilities, this work highlights integrated, system-level strategies and evaluates their practical feasibility in real-world environments. Key enabling trends such as procedural scene generation, policy generalization, and multimodal reasoning are discussed alongside core bottlenecks, including limited embodiment, lack of multimodal data, safety risks, and computational constraints. Through this lens, this paper identifies both the architectural strengths and critical limitations of foundation model-based robotics, highlighting open challenges in real-time operation, grounding, resilience, and trust. The review concludes with a roadmap for future research aimed at bridging semantic reasoning and physical intelligence through more robust, interpretable, and embodied models.

Large language models (LLMs) typically require fine-tuning for domain-specific tasks, and LoRA offers a computationally efficient approach by training low-rank adapters. LoRA is also communication-efficient for federated LLMs when multiple users collaboratively fine-tune a global LLM model without sharing their proprietary raw data. However, even the transmission of local adapters between a server and clients risks serious privacy leakage. Applying differential privacy (DP) to federated LoRA encounters a dilemma: adding noise to both adapters amplifies synthetic noise on the model, while fixing one adapter impairs the learnability of fine-tuning. In this paper, we propose FedASK (Differentially Private Federated Low Rank Adaptation with Double Sketching) , a novel federated LoRA framework to enable effective updating of both low-rank adapters with robust differential privacy. Inspired by randomized SVD, our key idea is a two-stage sketching pipeline. This pipeline first aggregates carefully sketched, privacy-preserving local updates, and then reconstructs the global matrices on the server to facilitate effective updating of both adapters. We theoretically prove FedASK's differential privacy guarantee and its exact aggregation property. Comprehensive experiments demonstrate that FedASK consistently outperforms baseline methods across a variety of privacy settings and data distributions.

Hit-like molecular generation with therapeutic potential is essential for target-specific drug discovery. However, the field lacks heterogeneous data and unified frameworks for integrating diverse molecular representations. To bridge this gap, we introduce TextOmics, a pioneering benchmark that establishes one-to-one correspondences between omics expressions and molecular textual descriptions. TextOmics provides a heterogeneous dataset that facilitates molecular generation through representations alignment. Built upon this foundation, we propose ToDi, a generative framework that jointly conditions on omics expressions and molecular textual descriptions to produce biologically relevant, chemically valid, hit-like molecules. ToDi leverages two encoders (OmicsEn and TextEn) to capture multi-level biological and semantic associations, and develops conditional diffusion (DiffGen) for controllable generation. Extensive experiments confirm the effectiveness of TextOmics and demonstrate ToDi outperforms existing state-of-the-art approaches, while also showcasing remarkable potential in zero-shot therapeutic molecular generation. Sources are available at: https://github.com/hala-ToDi.

Large Language Models (LLMs) demonstrate strong conversational abilities. In this Working Paper, we study them in the context of debating in two ways: their ability to perform in a structured debate along with a dataset of arguments to use and their ability to evaluate utterances throughout the debate. We deploy six leading publicly available models from three providers for the Retrieval-Augmented Debate and Evaluation. The evaluation is performed by measuring four key metrics: Quality, Quantity, Manner, and Relation. Throughout this task, we found that although LLMs perform well in debates when given related arguments, they tend to be verbose in responses yet consistent in evaluation. The accompanying source code for this paper is located at https://github.com/dsgt-arc/touche-2025-rad.

Function Calling is a crucial technique that enables Large Language Models (LLMs) to interact with external systems through APIs. However, the high latency associated with LLM-based Function Calling significantly impacts user experience. This paper presents a novel approach called Oriented Distillation for Inline Acceleration (ODIA) that leverages online user interaction data to accelerate Function Calling. By automatically identifying "simple queries" from production traffic and distilling knowledge from larger models to smaller ones, our method reduces response latency by 45% (expected) and 78% (median) while maintaining accuracy. We demonstrate the effectiveness of our approach through real-world deployment in a music application, where the smaller model successfully handles 60% of traffic with negligible accuracy loss. Our method requires minimal human intervention and continuously improves through automated data collection and model updating, making it a practical solution for production environments.

Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by retrieving external data to mitigate hallucinations and outdated knowledge issues. Benefiting from the strong ability in facilitating diverse data sources and supporting faithful reasoning, knowledge graphs (KGs) have been increasingly adopted in RAG systems, giving rise to KG-based RAG (KG-RAG) methods. Though RAG systems are widely applied in various applications, recent studies have also revealed its vulnerabilities to data poisoning attacks, where malicious information injected into external knowledge sources can mislead the system into producing incorrect or harmful responses. However, these studies focus exclusively on RAG systems using unstructured textual data sources, leaving the security risks of KG-RAG largely unexplored, despite the fact that KGs present unique vulnerabilities due to their structured and editable nature. In this work, we conduct the first systematic investigation of the security issue of KG-RAG methods through data poisoning attacks. To this end, we introduce a practical, stealthy attack setting that aligns with real-world implementation. We propose an attack strategy that first identifies adversarial target answers and then inserts perturbation triples to complete misleading inference chains in the KG, increasing the likelihood that KG-RAG methods retrieve and rely on these perturbations during generation. Through extensive experiments on two benchmarks and four recent KG-RAG methods, our attack strategy demonstrates strong effectiveness in degrading KG-RAG performance, even with minimal KG perturbations. In-depth analyses are also conducted to understand the safety threats within the internal stages of KG-RAG systems and to explore the robustness of LLMs against adversarial knowledge.

This study evaluates the performance of a compression method, called CompactifAI, developed by Multiverse Computing, applied to the large language model Llama 3.1 8B\cite{llama}. The evaluation focused on model efficiency (in terms of energy consumption) and accuracy using respectively the frameworks Codecarbon\cite{codecarbon} and Ragas\cite{ragas}. A comparison was performed between the model compressed with CompactifAI\cite{compactifai}\cite{compactifai2} and its full-size version. Our findings reveal that the compressed model using CompactifAI not only significantly reduced the computational resources but also maintained the model accuracy, making the model more efficient, scalable and cost-effective.

Recent large language models have shown promising capabilities in long-form reasoning, following structured chains of thought before arriving at a final answer. However, we observe that these reasoning paths tend to include substantial redundancy; analyzing attention patterns reveals that attention scores are widely scattered, particularly incorrect answers exhibit greater attention sparsity. In this paper, we demonstrate that deliberately removing this redundancy in the reasoning process significantly improves performance through clear thinking, i.e., removing distraction. Specifically, we systematically identify reasoning redundancy by measuring token-level attention scores to a special end-of-thinking token, which is appended to an explicit instruction inserted to conclude each intermediate reasoning step. Furthermore, we propose structure-aware pruning that prioritizes removing tokens in low-contributing reasoning chunks over individual tokens. After evicting redundant tokens, we remove the injected end-of-thinking instruction, then resume the reasoning generation. We demonstrate that our method significantly improves overall accuracy across reasoning-intensive benchmarks without any training involved. In particular, our method shows strong performance on challenging mathematical competition benchmarks such as AIME and AMC, where reasoning redundancy is more prevalent.

The emergence of Small Language Models (SLMs) as privacy-preserving alternatives for sensitive applications raises a fundamental question about their inherent understanding capabilities compared to Large Language Models (LLMs). This paper investigates the mental health understanding capabilities of current SLMs through systematic evaluation across diverse classification tasks. Employing zero-shot and few-shot learning paradigms, we benchmark their performance against established LLM baselines to elucidate their relative strengths and limitations in this critical domain. We assess five state-of-the-art SLMs (Phi-3, Phi-3.5, Qwen2.5, Llama-3.2, Gemma2) against three LLMs (GPT-4, FLAN-T5-XXL, Alpaca-7B) on six mental health understanding tasks. Our findings reveal that SLMs achieve mean performance within 2\% of LLMs on binary classification tasks (F1 scores of 0.64 vs 0.66 in zero-shot settings), demonstrating notable competence despite orders of magnitude fewer parameters. Both model categories experience similar degradation on multi-class severity tasks (a drop of over 30\%), suggesting that nuanced clinical understanding challenges transcend model scale. Few-shot prompting provides substantial improvements for SLMs (up to 14.6\%), while LLM gains are more variable. Our work highlights the potential of SLMs in mental health understanding, showing they can be effective privacy-preserving tools for analyzing sensitive online text data. In particular, their ability to quickly adapt and specialize with minimal data through few-shot learning positions them as promising candidates for scalable mental health screening tools.

Discrete diffusion language models learn to reconstruct text from randomly masked inputs, yet under mild assumptions their denoiser already implements the exact Bayesian posterior over the original tokens. We prove that the expected denoiser output under the forward corruption distribution recovers the true posterior, and that a simple Monte Carlo estimator converges to this posterior at rate O(1/sqrt(K)) with finite-sample concentration bounds. Building on this insight, we introduce an inference-time ensemble that runs K independent denoising passes and aggregates both posterior means and variances without any extra training. On WikiText-2, our MC-marginal sampler recovers the analytic lambda-DCE zero-shot perplexity (approximately 39) to within a few points at K=128, and its per-token variance shows a strong rank correlation with reconstruction error (Spearman rho = 0.996). This cost-proportional procedure yields calibrated uncertainty estimates and a direct trade-off between compute and posterior fidelity in discrete diffusion LMs.