Accelerating LLM Inference Throughput via Asynchronous KV Cache Prefetching

Large Language Models (LLMs) exhibit pronounced memory-bound characteristics during inference due to High Bandwidth Memory (HBM) bandwidth constraints. In this paper, we propose an L2 Cache-oriented asynchronous KV Cache prefetching method to break through the memory bandwidth bottleneck in LLM inference through computation-load overlap. By strategically scheduling idle memory bandwidth during active computation windows, our method proactively prefetches required KV Cache into GPU L2 cache, enabling high-speed L2 cache hits for subsequent accesses and effectively hiding HBM access latency within computational cycles. Extensive experiments on NVIDIA H20 GPUs demonstrate that the proposed method achieves 2.15x improvement in attention kernel efficiency and up to 1.97x end-to-end throughput enhancement, surpassing state-of-the-art baseline FlashAttention-3. Notably, our solution maintains orthogonality to existing optimization techniques and can be integrated with current inference frameworks, providing a scalable latency-hiding solution for next-generation LLM inference engines.

StarFlow: Generating Structured Workflow Outputs From Sketch Images

Workflows are a fundamental component of automation in enterprise platforms, enabling the orchestration of tasks, data processing, and system integrations. Despite being widely used, building workflows can be complex, often requiring manual configuration through low-code platforms or visual programming tools. To simplify this process, we explore the use of generative foundation models, particularly vision-language models (VLMs), to automatically generate structured workflows from visual inputs. Translating hand-drawn sketches or computer-generated diagrams into executable workflows is challenging due to the ambiguity of free-form drawings, variations in diagram styles, and the difficulty of inferring execution logic from visual elements. To address this, we introduce StarFlow, a framework for generating structured workflow outputs from sketches using vision-language models. We curate a diverse dataset of workflow diagrams -- including synthetic, manually annotated, and real-world samples -- to enable robust training and evaluation. We finetune and benchmark multiple vision-language models, conducting a series of ablation studies to analyze the strengths and limitations of our approach. Our results show that finetuning significantly enhances structured workflow generation, outperforming large vision-language models on this task.

Surg-3M: A Dataset and Foundation Model for Perception in Surgical Settings

Advancements in computer-assisted surgical procedures heavily rely on accurate visual data interpretation from camera systems used during surgeries. Traditional open-access datasets focusing on surgical procedures are often limited by their small size, typically consisting of fewer than 100 videos with less than 100K images. To address these constraints, a new dataset called Surg-3M has been compiled using a novel aggregation pipeline that collects high-resolution videos from online sources. Featuring an extensive collection of over 4K surgical videos and more than 3 million high-quality images from multiple procedure types, Surg-3M offers a comprehensive resource surpassing existing alternatives in size and scope, including two novel tasks. To demonstrate the effectiveness of this dataset, we present SurgFM, a self-supervised foundation model pretrained on Surg-3M that achieves impressive results in downstream tasks such as surgical phase recognition, action recognition, and tool presence detection. Combining key components from ConvNeXt, DINO, and an innovative augmented distillation method, SurgFM exhibits exceptional performance compared to specialist architectures across various benchmarks. Our experimental results show that SurgFM outperforms state-of-the-art models in multiple downstream tasks, including significant gains in surgical phase recognition (+8.9pp, +4.7pp, and +3.9pp of Jaccard in AutoLaparo, M2CAI16, and Cholec80), action recognition (+3.1pp of mAP in CholecT50) and tool presence detection (+4.6pp of mAP in Cholec80). Moreover, even when using only half of the data, SurgFM outperforms state-of-the-art models in AutoLaparo and achieves state-of-the-art performance in Cholec80. Both Surg-3M and SurgFM have significant potential to accelerate progress towards developing autonomous robotic surgery systems.

RFMI: Estimating Mutual Information on Rectified Flow for Text-to-Image Alignment

Rectified Flow (RF) models trained with a Flow matching framework have achieved state-of-the-art performance on Text-to-Image (T2I) conditional generation. Yet, multiple benchmarks show that synthetic images can still suffer from poor alignment with the prompt, i.e., images show wrong attribute binding, subject positioning, numeracy, etc. While the literature offers many methods to improve T2I alignment, they all consider only Diffusion Models, and require auxiliary datasets, scoring models, and linguistic analysis of the prompt. In this paper we aim to address these gaps. First, we introduce RFMI, a novel Mutual Information (MI) estimator for RF models that uses the pre-trained model itself for the MI estimation. Then, we investigate a self-supervised fine-tuning approach for T2I alignment based on RFMI that does not require auxiliary information other than the pre-trained model itself. Specifically, a fine-tuning set is constructed by selecting synthetic images generated from the pre-trained RF model and having high point-wise MI between images and prompts. Our experiments on MI estimation benchmarks demonstrate the validity of RFMI, and empirical fine-tuning on SD3.5-Medium confirms the effectiveness of RFMI for improving T2I alignment while maintaining image quality.

DPC: Dual-Prompt Collaboration for Tuning Vision-Language Models

The Base-New Trade-off (BNT) problem universally exists during the optimization of CLIP-based prompt tuning, where continuous fine-tuning on base (target) classes leads to a simultaneous decrease of generalization ability on new (unseen) classes. Existing approaches attempt to regulate the prompt tuning process to balance BNT by appending constraints. However, imposed on the same target prompt, these constraints fail to fully avert the mutual exclusivity between the optimization directions for base and new. As a novel solution to this challenge, we propose the plug-and-play Dual-Prompt Collaboration (DPC) framework, the first that decoupling the optimization processes of base and new tasks at the prompt level. Specifically, we clone a learnable parallel prompt based on the backbone prompt, and introduce a variable Weighting-Decoupling framework to independently control the optimization directions of dual prompts specific to base or new tasks, thus avoiding the conflict in generalization. Meanwhile, we propose a Dynamic Hard Negative Optimizer, utilizing dual prompts to construct a more challenging optimization task on base classes for enhancement. For interpretability, we prove the feature channel invariance of the prompt vector during the optimization process, providing theoretical support for the Weighting-Decoupling of DPC. Extensive experiments on multiple backbones demonstrate that DPC can significantly improve base performance without introducing any external knowledge beyond the base classes, while maintaining generalization to new classes. Code is available at: https://github.com/JREion/DPC.

Can LLMs Formally Reason as Abstract Interpreters for Program Analysis?

LLMs have demonstrated impressive capabilities in code generation and comprehension, but their potential in being able to perform program analysis in a formal, automatic manner remains under-explored. To that end, we systematically investigate whether LLMs can reason about programs using a program analysis framework called abstract interpretation. We prompt LLMs to follow two different strategies, denoted as Compositional and Fixed Point Equation, to formally reason in the style of abstract interpretation, which has never been done before to the best of our knowledge. We validate our approach using state-of-the-art LLMs on 22 challenging benchmark programs from the Software Verification Competition (SV-COMP) 2019 dataset, widely used in program analysis. Our results show that our strategies are able to elicit abstract interpretation-based reasoning in the tested models, but LLMs are susceptible to logical errors, especially while interpreting complex program structures, as well as general hallucinations. This highlights key areas for improvement in the formal reasoning capabilities of LLMs.

FinTMMBench: Benchmarking Temporal-Aware Multi-Modal RAG in Finance

Finance decision-making often relies on in-depth data analysis across various data sources, including financial tables, news articles, stock prices, etc. In this work, we introduce FinTMMBench, the first comprehensive benchmark for evaluating temporal-aware multi-modal Retrieval-Augmented Generation (RAG) systems in finance. Built from heterologous data of NASDAQ 100 companies, FinTMMBench offers three significant advantages. 1) Multi-modal Corpus: It encompasses a hybrid of financial tables, news articles, daily stock prices, and visual technical charts as the corpus. 2) Temporal-aware Questions: Each question requires the retrieval and interpretation of its relevant data over a specific time period, including daily, weekly, monthly, quarterly, and annual periods. 3) Diverse Financial Analysis Tasks: The questions involve 10 different tasks, including information extraction, trend analysis, sentiment analysis and event detection, etc. We further propose a novel TMMHybridRAG method, which first leverages LLMs to convert data from other modalities (e.g., tabular, visual and time-series data) into textual format and then incorporates temporal information in each node when constructing graphs and dense indexes. Its effectiveness has been validated in extensive experiments, but notable gaps remain, highlighting the challenges presented by our FinTMMBench.

TPC: Cross-Temporal Prediction Connection for Vision-Language Model Hallucination Reduction

Vision-language models (VLMs) have achieved remarkable advancements, capitalizing on the impressive capabilities of large language models (LLMs) across diverse tasks. Despite this, a critical challenge known as hallucination occurs when models overconfidently describe objects or attributes absent from the image, a problem exacerbated by the tendency of VLMs to rely on linguistic priors. This limitation reduces model reliability in high-stakes applications. In this work, we have observed the characteristic of logits' continuity consistency enhancement and introduced a straightforward and efficient method, Cross-Temporal Prediction Connection (TPC), designed to enhance the semantic consistency of logits by connecting them temporally across timesteps. TPC amplifies information flow and improves coherence, effectively reducing hallucination. Extensive experiments show that TPC surpasses existing representatives, delivering superior performance in both accuracy and efficiency while maintaining robustness in open-ended text generation tasks.

Advancing vision-language models in front-end development via data synthesis

Modern front-end (FE) development, especially when leveraging the unique features of frameworks like React and Vue, presents distinctive challenges. These include managing modular architectures, ensuring synchronization between data and visual outputs for declarative rendering, and adapting reusable components to various scenarios. Such complexities make it particularly difficult for state-of-the-art large vision-language models (VLMs) to generate accurate and functional code directly from design images. To address these challenges, we propose a reflective agentic workflow that synthesizes high-quality image-text data to capture the diverse characteristics of FE development. This workflow automates the extraction of self-contained\footnote{A \textbf{self-contained} code snippet is one that encapsulates all necessary logic, styling, and dependencies, ensuring it functions independently without requiring external imports or context.} code snippets from real-world projects, renders the corresponding visual outputs, and generates detailed descriptions that link design elements to functional code. To further expand the scope and utility of the synthesis, we introduce three data synthesis strategies: Evolution-based synthesis, which enables scalable and diverse dataset expansion; Waterfall-Model-based synthesis, which generates logically coherent code derived from system requirements; and Additive Development synthesis, which iteratively increases the complexity of human-authored components. We build a large vision-language model, Flame, trained on the synthesized datasets and demonstrate its effectiveness in generating React code via the $\text{pass}@k$ metric. Our results suggest that a code VLM trained to interpret images before code generation may achieve better performance.

Can Large Language Models Unveil the Mysteries? An Exploration of Their Ability to Unlock Information in Complex Scenarios

Combining multiple perceptual inputs and performing combinatorial reasoning in complex scenarios is a sophisticated cognitive function in humans. With advancements in multi-modal large language models, recent benchmarks tend to evaluate visual understanding across multiple images. However, they often overlook the necessity of combinatorial reasoning across multiple perceptual information. To explore the ability of advanced models to integrate multiple perceptual inputs for combinatorial reasoning in complex scenarios, we introduce two benchmarks: Clue-Visual Question Answering (CVQA), with three task types to assess visual comprehension and synthesis, and Clue of Password-Visual Question Answering (CPVQA), with two task types focused on accurate interpretation and application of visual data. For our benchmarks, we present three plug-and-play approaches: utilizing model input for reasoning, enhancing reasoning through minimum margin decoding with randomness generation, and retrieving semantically relevant visual information for effective data integration. The combined results reveal current models' poor performance on combinatorial reasoning benchmarks, even the state-of-the-art (SOTA) closed-source model achieves only 33.04% accuracy on CVQA, and drops to 7.38% on CPVQA. Notably, our approach improves the performance of models on combinatorial reasoning, with a 22.17% boost on CVQA and 9.40% on CPVQA over the SOTA closed-source model, demonstrating its effectiveness in enhancing combinatorial reasoning with multiple perceptual inputs in complex scenarios. The code will be publicly available.