Clinical oncology generates vast, unstructured data that often contain inconsistencies, missing information, and ambiguities, making it difficult to extract reliable insights for data-driven decision-making. General-purpose large language models (LLMs) struggle with these challenges due to their lack of domain-specific reasoning, including specialized clinical terminology, context-dependent interpretations, and multi-modal data integration. We address these issues with an oncology-specialized, efficient, and adaptable NLP framework that combines instruction tuning, retrieval-augmented generation (RAG), and graph-based knowledge integration. Our lightweight models prove effective at oncology-specific tasks, such as named entity recognition (e.g., identifying cancer diagnoses), entity linking (e.g., linking entities to standardized ontologies), TNM staging, document classification (e.g., cancer subtype classification from pathology reports), and treatment response prediction. Our framework emphasizes adaptability and resource efficiency. We include minimal German instructions, collected at the University Hospital Zurich (USZ), to test whether small amounts of non-English language data can effectively transfer knowledge across languages. This approach mirrors our motivation for lightweight models, which balance strong performance with reduced computational costs, making them suitable for resource-limited healthcare settings. We validated our models on oncology datasets, demonstrating strong results in named entity recognition, relation extraction, and document classification.

Large language models have been widely adopted across different tasks, but their auto-regressive generation nature often leads to inefficient resource utilization during inference. While batching is commonly used to increase throughput, performance gains plateau beyond a certain batch size, especially with smaller models, a phenomenon that existing literature typically explains as a shift to the compute-bound regime. In this paper, through an in-depth GPU-level analysis, we reveal that large-batch inference remains memory-bound, with most GPU compute capabilities underutilized due to DRAM bandwidth saturation as the primary bottleneck. To address this, we propose a Batching Configuration Advisor (BCA) that optimizes memory allocation, reducing GPU memory requirements with minimal impact on throughput. The freed memory and underutilized GPU compute capabilities can then be leveraged by concurrent workloads. Specifically, we use model replication to improve serving throughput and GPU utilization. Our findings challenge conventional assumptions about LLM inference, offering new insights and practical strategies for improving resource utilization, particularly for smaller language models.

Multimodal large language models (MLLMs) show promise in tasks like visual question answering (VQA) but still face challenges in multimodal reasoning. Recent works adapt agentic frameworks or chain-of-thought (CoT) reasoning to improve performance. However, CoT-based multimodal reasoning often demands costly data annotation and fine-tuning, while agentic approaches relying on external tools risk introducing unreliable output from these tools. In this paper, we propose Seeing and Reasoning with Confidence (SRICE), a training-free multimodal reasoning framework that integrates external vision models with uncertainty quantification (UQ) into an MLLM to address these challenges. Specifically, SRICE guides the inference process by allowing MLLM to autonomously select regions of interest through multi-stage interactions with the help of external tools. We propose to use a conformal prediction-based approach to calibrate the output of external tools and select the optimal tool by estimating the uncertainty of an MLLM's output. Our experiment shows that the average improvement of SRICE over the base MLLM is 4.6% on five datasets and the performance on some datasets even outperforms fine-tuning-based methods, revealing the significance of ensuring reliable tool use in an MLLM agent.

The emergence of large language models (LLMs) opens new frontiers for unmanned aerial vehicle (UAVs), yet existing systems remain confined to predefined tasks due to hardware-software co-design challenges. This paper presents the first aerial intelligent agent capable of open-world task execution through tight integration of LLM-based reasoning and robotic autonomy. Our hardware-software co-designed system addresses two fundamental limitations: (1) Onboard LLM operation via an edge-optimized computing platform, achieving 5-6 tokens/sec inference for 14B-parameter models at 220W peak power; (2) A bidirectional cognitive architecture that synergizes slow deliberative planning (LLM task planning) with fast reactive control (state estimation, mapping, obstacle avoidance, and motion planning). Validated through preliminary results using our prototype, the system demonstrates reliable task planning and scene understanding in communication-constrained environments, such as sugarcane monitoring, power grid inspection, mine tunnel exploration, and biological observation applications. This work establishes a novel framework for embodied aerial artificial intelligence, bridging the gap between task planning and robotic autonomy in open environments.

Long-form writing agents require flexible integration and interaction across information retrieval, reasoning, and composition. Current approaches rely on predetermined workflows and rigid thinking patterns to generate outlines before writing, resulting in constrained adaptability during writing. In this paper we propose a general agent framework that achieves human-like adaptive writing through recursive task decomposition and dynamic integration of three fundamental task types, i.e. retrieval, reasoning, and composition. Our methodology features: 1) a planning mechanism that interleaves recursive task decomposition and execution, eliminating artificial restrictions on writing workflow; and 2) integration of task types that facilitates heterogeneous task decomposition. Evaluations on both fiction writing and technical report generation show that our method consistently outperforms state-of-the-art approaches across all automatic evaluation metrics, which demonstrate the effectiveness and broad applicability of our proposed framework.

Recent research has shown an increasing interest in utilizing pre-trained large language models (LLMs) for a variety of time series applications. However, there are three main challenges when using LLMs as foundational models for time series forecasting: (1) Cross-domain generalization. (2) Cross-modality alignment. (3) Error accumulation in autoregressive frameworks. To address these challenges, we proposed LangTime, a language-guided unified model for time series forecasting that incorporates cross-domain pre-training with reinforcement learning-based fine-tuning. Specifically, LangTime constructs Temporal Comprehension Prompts (TCPs), which include dataset-wise and channel-wise instructions, to facilitate domain adaptation and condense time series into a single token, enabling LLMs to understand better and align temporal data. To improve autoregressive forecasting, we introduce TimePPO, a reinforcement learning-based fine-tuning algorithm. TimePPO mitigates error accumulation by leveraging a multidimensional rewards function tailored for time series and a repeat-based value estimation strategy. Extensive experiments demonstrate that LangTime achieves state-of-the-art cross-domain forecasting performance, while TimePPO fine-tuning effectively enhances the stability and accuracy of autoregressive forecasting.

Code optimization is the process of enhancing code efficiency, while preserving its intended functionality. This process often requires a deep understanding of the code execution behavior at run-time to identify and address inefficiencies effectively. Recent studies have shown that language models can play a significant role in automating code optimization. However, these models may have insufficient knowledge of how code execute at run-time. To address this limitation, researchers have developed strategies that integrate code execution information into language models. These strategies have shown promise, enhancing the effectiveness of language models in various software engineering tasks. However, despite the close relationship between code execution behavior and efficiency, the specific impact of these strategies on code optimization remains largely unexplored. This study investigates how incorporating code execution information into language models affects their ability to optimize code. Specifically, we apply three different training strategies to incorporate four code execution aspects -- line executions, line coverage, branch coverage, and variable states -- into CodeT5+, a well-known language model for code. Our results indicate that execution-aware models provide limited benefits compared to the standard CodeT5+ model in optimizing code.

We present EgoBlind, the first egocentric VideoQA dataset collected from blind individuals to evaluate the assistive capabilities of contemporary multimodal large language models (MLLMs). EgoBlind comprises 1,210 videos that record the daily lives of real blind users from a first-person perspective. It also features 4,927 questions directly posed or generated and verified by blind individuals to reflect their needs for visual assistance under various scenarios. We provide each question with an average of 3 reference answers to alleviate subjective evaluation. Using EgoBlind, we comprehensively evaluate 15 leading MLLMs and find that all models struggle, with the best performers achieving accuracy around 56\%, far behind human performance of 87.4\%. To guide future advancements, we identify and summarize major limitations of existing MLLMs in egocentric visual assistance for the blind and provide heuristic suggestions for improvement. With these efforts, we hope EgoBlind can serve as a valuable foundation for developing more effective AI assistants to enhance the independence of the blind individuals' lives.

In this paper, I present our work on DeepRAG, a specialized embedding model we built specifically for Hindi language in RAG systems. While LLMs have gotten really good at generating text, their performance in retrieval tasks still depends heavily on having quality embeddings - something that's been lacking for Hindi despite being one of the world's most spoken languages. We tackled this by creating embeddings from the ground up rather than just fine-tuning existing models. Our process involved collecting diverse Hindi texts (over 2.7M samples), training a custom SentencePiece tokenizer that actually understands Hindi morphology, designing transformer architecture with Hindi-specific attention mechanisms, and optimizing with contrastive learning. Results were honestly better than I expected - we saw a 23% improvement in retrieval precision compared to the multilingual models everyone's been using. The paper details our methodology, which I think could help others working with low-resource languages where the one-size-fits-all multilingual models fall short. We've also integrated our embeddings with LangChain to build complete Hindi RAG systems, which might be useful for practitioners. While there's still tons more to explore, I believe this work addresses a critical gap for Hindi NLP and demonstrates why language-specific approaches matter.

Mechanistic interpretability of large language models (LLMs) aims to uncover the internal processes of information propagation and reasoning. Sparse autoencoders (SAEs) have demonstrated promise in this domain by extracting interpretable and monosemantic features. However, prior works primarily focus on feature extraction from a single layer, failing to effectively capture activations that span multiple layers. In this paper, we introduce Route Sparse Autoencoder (RouteSAE), a new framework that integrates a routing mechanism with a shared SAE to efficiently extract features from multiple layers. It dynamically assigns weights to activations from different layers, incurring minimal parameter overhead while achieving high interpretability and flexibility for targeted feature manipulation. We evaluate RouteSAE through extensive experiments on Llama-3.2-1B-Instruct. Specifically, under the same sparsity constraint of 64, RouteSAE extracts 22.5% more features than baseline SAEs while achieving a 22.3% higher interpretability score. These results underscore the potential of RouteSAE as a scalable and effective method for LLM interpretability, with applications in feature discovery and model intervention. Our codes are available at https://github.com/swei2001/RouteSAEs.