GPU kernels are critical for ML performance but difficult to optimize across diverse accelerators. We present KForge, a platform-agnostic framework built on two collaborative LLM-based agents: a generation agent that produces and iteratively refines programs through compilation and correctness feedback, and a performance analysis agent that interprets profiling data to guide optimization. This agent-based architecture requires only a single-shot example to target new platforms. We make three key contributions: (1) introducing an iterative refinement system where the generation agent and performance analysis agent collaborate through functional and optimization passes, interpreting diverse profiling data (from programmatic APIs to GUI-based tools) to generate actionable recommendations that guide program synthesis for arbitrary accelerators; (2) demonstrating that the generation agent effectively leverages cross-platform knowledge transfer, where a reference implementation from one architecture substantially improves generation quality for different hardware targets; and (3) validating the platform-agnostic nature of our approach by demonstrating effective program synthesis across fundamentally different parallel computing platforms: NVIDIA CUDA and Apple Metal.

Abstract--The rise of Generative AI (GenAI) tools like Chat-GPT has created new opportunities and challenges for computing education. Existing research has primarily focused on GenAI's ability to complete educational tasks and its impact on student performance, often overlooking its effects on knowledge gains. In this study, we investigate how GenAI assistance compares to conventional online resources in supporting knowledge gains across different proficiency levels. We conducted a controlled user experiment with 24 undergraduate students of two different levels of programming experience (beginner, intermediate) to examine how students interact with ChatGPT while solving programming tasks. We analyzed task performance, conceptual understanding, and interaction behaviors. Our findings reveal that generating complete solutions with GenAI significantly improves task performance, especially for beginners, but does not consistently result in knowledge gains. Importantly, usage strategies differ by experience: beginners tend to rely heavily on GenAI toward task completion often without knowledge gain in the process, while intermediates adopt more selective approaches. We find that both over-reliance and minimal use result in weaker knowledge gains overall. Based on our results, we call on students and educators to adopt GenAI as a learning rather than a problem solving tool. Our study highlights the urgent need for guidance when integrating GenAI into programming education to foster deeper understanding. The rapid development of Generative Artificial Intelligence (GenAI) has led to its widespread adoption across various domains to boost productivity and streamline workflows. Large Language Models (LLMs), such as OpenAI's ChatGPT and Codex, Google Gemini, and GitHub Copilot, have been integrated into domains including software engineering [1], [2], healthcare [3], education [4], creative writing [5], [6], and digital music [7], offering capabilities such as code generation, question answering, and image generation. These authors contributed equally to this work. Some studies evaluated GenAI's performance on programming tasks [8], user interface design education [9], and computer vision coursework [10]. Others focused on assessing the accuracy and usability of GenAIgenerated responses [11], [12].

Large multimodal models (LMMs) have demonstrated remarkable capabilities across a wide range of tasks, however their knowledge and abilities in the cross-view geo-localization and pose estimation domains remain unexplored, despite potential benefits for navigation, autonomous driving, outdoor robotics, \textit{etc}. To bridge this gap, we introduce \textbf{GeoX-Bench}, a comprehensive \underline{Bench}mark designed to explore and evaluate the capabilities of LMMs in \underline{cross}-view \underline{Geo}-localization and pose estimation. Specifically, GeoX-Bench contains 10,859 panoramic-satellite image pairs spanning 128 cities in 49 countries, along with corresponding 755,976 question-answering (QA) pairs. Among these, 42,900 QA pairs are designated for benchmarking, while the remaining are intended to enhance the capabilities of LMMs. Based on GeoX-Bench, we evaluate the capabilities of 25 state-of-the-art LMMs on cross-view geo-localization and pose estimation tasks, and further explore the empowered capabilities of instruction-tuning. Our benchmark demonstrate that while current LMMs achieve impressive performance in geo-localization tasks, their effectiveness declines significantly on the more complex pose estimation tasks, highlighting a critical area for future improvement, and instruction-tuning LMMs on the training data of GeoX-Bench can significantly improve the cross-view geo-sense abilities. The GeoX-Bench is available at \textcolor{magenta}{https://github.com/IntMeGroup/GeoX-Bench}.

Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse domains, but their training remains resource- and time-intensive, requiring massive compute power and careful orchestration of training procedures. Model souping-the practice of averaging weights from multiple models of the same architecture-has emerged as a promising pre- and post-training technique that can enhance performance without expensive retraining. In this paper, we introduce Soup Of Category Experts (SoCE), a principled approach for model souping that utilizes benchmark composition to identify optimal model candidates and applies non-uniform weighted averaging to maximize performance. Contrary to previous uniform-averaging approaches, our method leverages the observation that benchmark categories often exhibit low inter-correlations in model performance. SoCE identifies "expert" models for each weakly-correlated category cluster and combines them using optimized weighted averaging rather than uniform weights. We demonstrate that the proposed method improves performance and robustness across multiple domains, including multilingual capabilities, tool calling, and math and achieves state-of-the-art results on the Berkeley Function Calling Leaderboard.

Multimodal Model Editing (MMED) aims to correct erroneous knowledge in multimodal models. Existing evaluation methods, adapted from textual model editing, overstate success by relying on low-similarity or random inputs, obscure overfitting. We propose a comprehensive locality evaluation framework, covering three key dimensions: random-image locality, no-image locality, and consistent-image locality, op-erationalized through seven distinct data types, enabling a detailed and structured analysis of multimodal edits. We introduce De-VQA, a dynamic evaluation for visual question answering, uncovering a phenomenon we term transient blindness, overfitting to edit-similar text while ignoring visuals. Token analysis shows edits disproportionately affect textual tokens. We propose locality-aware adversarial losses to balance cross-modal representations. Empirical results demonstrate that our approach consistently outperforms existing baselines, reducing transient blindness and improving locality by 17% on average.

With the rise of Large Language Models (LLMs) and their vision-enabled counterparts (VLMs), numerous works have investigated their capabilities in tasks that fuse the modalities of vision and language. In this work, we benchmark the extent to which VLMs are able to act as highly-trained phoneticians, interpreting spectrograms and waveforms of speech. To do this, we synthesise a novel dataset containing 4k+ English words spoken in isolation alongside stylistically consistent spectrogram and waveform figures. We test the ability of VLMs to understand these representations of speech through a multiple-choice task whereby models must predict the correct phonemic or graphemic transcription of a spoken word when presented amongst 3 distractor transcriptions that have been selected based on their phonemic edit distance to the ground truth. We observe that both zero-shot and finetuned models rarely perform above chance, demonstrating the requirement for specific parametric knowledge of how to interpret such figures, rather than paired samples alone.

Emerging reasoning LLMs such as OpenAI-o1 and DeepSeek-R1 have achieved strong performance on complex reasoning tasks by generating long chain-of-thought (CoT) traces. However, these long CoTs result in increased token usage, leading to higher inference latency and memory consumption. As a result, balancing accuracy and reasoning efficiency has become essential for deploying reasoning LLMs in practical applications. Existing long-to-short (Long2Short) methods aim to reduce inference length but often sacrifice accuracy, revealing a need for an approach that maintains performance while lowering token costs. To address this efficiency-accuracy tradeoff, we propose TokenSqueeze, a novel Long2Short method that condenses reasoning paths while preserving performance and relying exclusively on self-generated data. First, to prevent performance degradation caused by excessive compression of reasoning depth, we propose to select self-generated samples whose reasoning depth is adaptively matched to the complexity of the problem. To further optimize the linguistic expression without altering the underlying reasoning paths, we introduce a distribution-aligned linguistic refinement method that enhances the clarity and conciseness of the reasoning path while preserving its logical integrity. Comprehensive experimental results demonstrate the effectiveness of TokenSqueeze in reducing token usage while maintaining accuracy. Notably, DeepSeek-R1-Distill-Qwen-7B fine-tuned using our proposed method achieved a 50\% average token reduction while preserving accuracy on the MATH500 benchmark. TokenSqueeze exclusively utilizes the model's self-generated data, enabling efficient and high-fidelity reasoning without relying on manually curated short-answer datasets across diverse applications. Our code is available at https://github.com/zhangyx1122/TokenSqueeze.

Navigating to a specified object in an unknown environment is a fundamental yet challenging capability of embodied intelligence. However, current methods struggle to balance decision frequency with intelligence, resulting in decisions lacking foresight or discontinuous actions. In this work, we propose PIGEON: Point of Interest Guided Exploration for Object Navigation with VLM, maintaining a lightweight and semantically aligned snapshot memory during exploration as semantic input for the exploration strategy. We use a large Visual-Language Model (VLM), named PIGEON-VL, to select Points of Interest (PoI) formed during exploration and then employ a lower-level planner for action output, increasing the decision frequency. Additionally, this PoI-based decision-making enables the generation of Reinforcement Learning with V erifiable Reward (RL VR) data suitable for simulators. Experiments on classic object navigation benchmarks demonstrate that our zero-shot transfer method achieves state-of-the-art performance, while RL VR further enhances the model's semantic guidance capabilities, enabling deep reasoning during real-time navigation.

Dynamic sequences with varying lengths have been widely used in the training of Transformer-based large language models (LLMs). However, current training frameworks adopt a pre-defined static parallel strategy for these sequences, causing neither communication-parallelization cancellation on short sequences nor out-of-memory on long sequences. To mitigate these issues, we propose ParaDySe, a novel adaptive Parallel strategy switching framework for Dynamic Sequences. ParaDySe enables on-the-fly optimal strategy adoption according to the immediate input sequence. It first implements the modular function libraries for parallel strategies with unified tensor layout specifications, and then builds sequence-aware memory and time cost models with hybrid methods. Guided by cost models, ParaDySe selects optimal layer-wise strategies for dynamic sequences via an efficient heuristic algorithm. By integrating these techniques together, ParaDySe achieves seamless hot-switching of optimal strategies through its well-designed function libraries. We compare ParaDySe with baselines on representative LLMs under datasets with sequence lengths up to 624K. Experimental results indicate that ParaDySe addresses OOM and CPC bottlenecks in LLM training by systematically integrating long-sequence optimizations with existing frameworks.

Large language models (LLMs) have demonstrated exceptional capabilities in general domains, yet their application in highly specialized and culturally-rich fields like Traditional Chinese Medicine (TCM) requires rigorous and nuanced evaluation. Building upon prior foundational work such as TCM-3CEval, which highlighted systemic knowledge gaps and the importance of cultural-contextual alignment, we introduce TCM-5CEval, a more granular and comprehensive benchmark. TCM-5CEval is designed to assess LLMs across five critical dimensions: (1) Core Knowledge (TCM-Exam), (2) Classical Literacy (TCM-LitQA), (3) Clinical Decision-making (TCM-MRCD), (4) Chinese Materia Medica (TCM-CMM), and (5) Clinical Non-pharmacological Therapy (TCM-ClinNPT). We conducted a thorough evaluation of fifteen prominent LLMs, revealing significant performance disparities and identifying top-performing models like deepseek\_r1 and gemini\_2\_5\_pro. Our findings show that while models exhibit proficiency in recalling foundational knowledge, they struggle with the interpretative complexities of classical texts. Critically, permutation-based consistency testing reveals widespread fragilities in model inference. All evaluated models, including the highest-scoring ones, displayed a substantial performance degradation when faced with varied question option ordering, indicating a pervasive sensitivity to positional bias and a lack of robust understanding. TCM-5CEval not only provides a more detailed diagnostic tool for LLM capabilities in TCM but aldso exposes fundamental weaknesses in their reasoning stability. To promote further research and standardized comparison, TCM-5CEval has been uploaded to the Medbench platform, joining its predecessor in the "In-depth Challenge for Comprehensive TCM Abilities" special track.