Tagged articles
multimodal LLM
100 articles · Page 1 of 1
Researchers from Northwestern, CMU, UNC, and Amazon introduce WOVEN, a framework that defines world modeling as state-transition prediction, creates a controllable benchmark with 36K questions, shows current models lack geometric reasoning, demonstrates that 2K training samples can teach world modeling efficiently, and proves it transfers to 22 downstream benchmarks, revealing that reasoning operation type matters more than action similarity for transfer.
BenchmarkWOVENcross-task transfer
0 likes · 17 min read

The author tests Tencent's Hunyuan Hy4 model across six demanding agent workflows in WorkBuddy — video generation from docs, lecture-to-webpage conversion, game cloning, 3D teardown simulators, industry report generation, and templated PPT creation — finding Hy4 handles 30-40 minute long-horizon tasks with solid code, vision, and planning at $0.55/task, 62% cheaper than Kimi K3.1 Pro.
AI AgentsHunyuan Hy4WorkBuddy
0 likes · 8 min read

DeepSeek releases V4.1-Flash-1017, a flash-tier model that outperforms flagship competitors on key agent benchmarks like DeepSWE and Automation-Bench while reducing KV cache to 1/4 of its predecessor via CSA2 sparse attention and FP4 quantization, and adds native multimodal support, adjustable reasoning intensity, and a new Search & Fetch API for real-time web access.
Agent BenchmarksCSA2 sparse attentionDeepSeek
0 likes · 6 min read

TerraVis introduces a systematic framework for evaluating world consistency in AI-generated images, covering 18 issue types across object, interaction, and scene levels, validated on 13,500 human annotations across five models.
AI-generated imagesBenchmarkEvaluation Metrics
0 likes · 15 min read


Amap TechSep 23, 2026 · Artificial Intelligence
Amap details its system for extracting navigation landmarks from street-view imagery using multimodal large models, covering data annotation standards, a three-stage training pipeline (SFT, DPO, GRPO with a risk-aware reward), bounding-box-free evaluation, offline production, online integration, and A/B tests showing 23.5% faster user departure.
DPOGRPOSFT
0 likes · 26 min read

Based on testing 17 industrial multimodal AI systems, this article exposes five critical misconceptions — confusing input coverage with semantic alignment, relying on static benchmarks, misattributing fusion-layer errors, ignoring temporal asynchrony, and using single metrics — that cause models to pass tests but fail in production.
AI Testingbenchmark datasetsfusion layer
0 likes · 11 min read

The article analyzes how eliminating the Vision Encoder from multimodal large language models simplifies the computation graph, reduces load imbalance, avoids conflicting parallel configurations, and makes scheduling and performance modeling more predictable, especially when visual features can be pre‑computed offline.
Cost ModelingParallelismTraining Infrastructure
0 likes · 9 min read

Octopus introduces a history‑free gradient orthogonalization (HiFGO) constraint and a two‑stage finetuning strategy that together eliminate catastrophic forgetting, achieve positive backward transfer, and set new SOTA results on the UCIT benchmark, improving Avg and Last metrics by 2.14% and 6.82% without extra inference cost.
Continual LearningUCIT benchmarkgradient orthogonalization
0 likes · 11 min read

Huawei Cloud adapts and optimizes the 280B‑parameter dots3‑note preview model released by Xiaohongshu, detailing its mixed sparse attention, MoE enhancements, multimodal pre‑training pipelines, unified reward training, and a suite of inference‑time optimizations—including FlashComm, FUSED_MC2, and MTP speculative decoding—delivered via the vLLM Ascend framework for high‑throughput, low‑latency deployment.
Huawei CloudMTP Speculative DecodingMoE optimization
0 likes · 7 min read

The paper introduces CDL Solver, a two‑stage decoupled framework that translates plane‑geometry diagrams into a concise symbolic language (CDL), reducing training data by 43× and achieving 85.7% accuracy on FormalGeo—surpassing Gemini 2.5 Pro, GPT‑4o and prior specialized models—while also demonstrating strong out‑of‑domain generalisation.
CVPR 2026concise description languagefew-shot training
0 likes · 9 min read

The paper introduces AgentHOI, a framework that eliminates the need for HOI‑specific annotations by decomposing detection into context‑aware semantic reasoning and multifaceted spatial localization, leveraging GPT‑4o and GroundingDINO, and demonstrates strong zero‑shot and robustness performance on HICO‑DET and SWIG‑HOI benchmarks.
AgentHOIContext-aware ReasoningHuman-Object Interaction
0 likes · 11 min read

Vision-OPD introduces an online self‑distillation framework that lets a 9B multimodal LLM internalize fine‑grained visual evidence from whole images, achieving state‑of‑the‑art results on six detailed visual‑understanding benchmarks without extra inference tools.
BenchmarkVision-OPDfine-grained visual understanding
0 likes · 13 min read

VideoChat3 is a 4‑billion‑parameter multimodal large model that unifies short‑video, long‑video, and streaming video understanding through a native spatiotemporal encoder, adaptive resolution budgeting, and four‑stage training, achieving competitive accuracy while dramatically reducing visual token count and inference cost.
Adaptive ResolutionI3D-ViTStreaming video
0 likes · 15 min read

BigMac introduces a dependency‑safe nested pipeline that keeps the LLM schedule intact while inserting encoder and generator work, achieving 1.08‑1.9× speedup and stable GPU memory usage across large‑scale multimodal LLM training workloads.
BigMacPipeline Parallelismmemory efficiency
0 likes · 17 min read

BigMac introduces a dependency‑safe nested pipeline that preserves LLM compute efficiency while reducing encoder and generator activation memory to O(1), delivering 1.08‑1.9× speedup and stable memory usage for large‑scale multimodal model training.
Nested PipelinePipeline Parallelismcompute efficiency
0 likes · 15 min read

VideoChat3 is a 4‑billion‑parameter multimodal model that introduces an Inflated 3D Vision Transformer and an adaptive frame‑resolution mechanism to efficiently handle short clips, hour‑long videos, and live streams, achieving competitive benchmarks while being fully open‑source.
Adaptive Frame ResolutionI3D-ViTStreaming video
0 likes · 13 min read

ProLaViT introduces a progressive latent visual reasoning framework that replaces costly explicit image generation and unstable one‑shot latent inference with a structured Locate→Focus→Isolate chain, endogenous self‑distillation, and a distance‑weighted diversity loss, achieving state‑of‑the‑art accuracy on multiple visual reasoning benchmarks.
Diversity LossProLaViTlatent space
0 likes · 10 min read

CoLT replaces the verbose textual chain‑of‑thought with just three latent vectors, delivering up to 22.6× faster generation and 10.1× end‑to‑end inference speedups while achieving a 79.1% average accuracy across eight multimodal benchmarks, all without any auxiliary visual annotations.
Chain of Latent ThoughtsEfficiencybenchmark evaluation
0 likes · 14 min read

The article benchmarks three Chinese multimodal large models—Step 3.7 Flash, MiniMax M3, and Qwen 3.6‑flash—across two real‑world tasks, measuring output quality, API latency, and token cost, and concludes that Step 3.7 Flash consistently offers the best speed‑cost trade‑off for production use.
API latencyBenchmarkMiniMax M3
0 likes · 10 min read

The open‑source 4B UI‑UX multimodal model achieves a 0.7963 SOTA score on the UXBench benchmark, surpassing much larger models such as Claude‑4.5‑Sonnet and Qwen3‑VL‑Thinking, thanks to reward routing, asymmetric rewards, and data‑alchemy techniques, and it can be installed and deployed with a few pip commands.
AI for UXModel EvaluationUI/UX
0 likes · 11 min read

The paper identifies a feature‑space mismatch that makes visual latent reasoning unstable, proposes the Granular Alignment Paradigm (GAP) with data, feature, and model‑capacity alignment, and demonstrates through extensive experiments that GAP improves both visual perception and multimodal reasoning performance.
Granular Alignment ParadigmPCA alignmentfeature alignment
0 likes · 19 min read

A new study from Westlake University shows that when harmful text is rendered as low‑resolution, blurry, or noisy images, multimodal large language models become significantly easier to jailbreak despite still recognizing the text, revealing a U‑shaped risk curve and a simple mitigation that decouples OCR from safety checks.
OCRjailbreakmultimodal LLM
0 likes · 10 min read

Westlake University's AGI Lab shows that when harmful text is rendered as low‑resolution, blurry or noisy images, multimodal large language models can still read the content but their safety filters fail, creating an 'attack comfort zone' that dramatically raises jailbreak success rates across several models.
OCRjailbreakmultimodal LLM
0 likes · 9 min read

Google DeepMind’s Gemma 4 12B model, released under Apache 2.0, runs fully offline on a 16 GB laptop, uses a novel no‑encoder unified architecture, delivers 80 token/s with only 9 GB VRAM, and matches the quality of the 26 B predecessor while powering advanced agentic and multimodal demos.
Apache 2.0BenchmarkGemma 4
0 likes · 13 min read

Google’s newly released Gemma 4 12B model delivers reasoning power comparable to the larger 26B MoE model while fitting within 16 GB of memory, thanks to a unified architecture, native audio support, and draft‑model acceleration, and it can run locally on consumer laptops.
12B modelBenchmarkGemma 4
0 likes · 6 min read


AntTechJun 1, 2026 · Artificial Intelligence
Venus‑DeFakerOne introduces a unified foundation model that combines global semantics and fine‑grained pixel analysis to detect and localize diverse image forgeries, leveraging a 12.5 M mixed dataset, MLLM‑SAM2 architecture, and achieving top scores on 39 detection and 9 localization benchmarks with 95.8% accuracy on the GPT‑Image‑2 attack suite.
SAM2Venus-DeFakerOnedeepfake
0 likes · 7 min read


SuanNiJun 1, 2026 · Artificial Intelligence
MiniMax M3, a domestically developed LLM, combines a new sparse‑attention MSA architecture, native multimodal support, and million‑token context to match or surpass top closed‑source models in programming and agent benchmarks, while achieving a 9.4× speedup on FP8 GEMM and preparing for open‑source release.
AIFP8 GEMMMiniMax M3
0 likes · 12 min read

MedScope proposes a "Think with Videos" paradigm that lets AI models actively locate and verify evidence in long clinical videos, using coarse‑to‑fine tool calling, evidence‑centric training data (ClinVideoSuite) and a grounding‑aware reinforcement learning objective, achieving superior performance on multiple video‑understanding benchmarks.
Evidence-based QAMedical Video AITool Calling
0 likes · 10 min read


AIWalkerMay 16, 2026 · Artificial Intelligence
Qwen3-VL-Seg, the latest open‑source multimodal LLM from Alibaba, extends bounding‑box predictions to pixel‑accurate masks using a lightweight box‑guided decoder, achieving strong performance on both closed‑set and open‑world segmentation tasks with only 0.4% extra parameters.
Qwen3-VL-SegSA1B-ORS datasetbox‑guided decoder
0 likes · 6 min read

APO (Autonomous Preference Optimization) converts the drift and conflict among multiple teacher multimodal LLMs into dynamic negative constraints while treating consensus as a positive preference, enabling robust concept alignment and superior diagnostic accuracy on the CXR‑MAX benchmark, as demonstrated by extensive ICML‑2026 experiments.
APOICML 2026concept drift
0 likes · 11 min read

The paper introduces APO, a novel autonomous preference optimization framework that converts concept drift among multiple teacher multimodal LLMs into dynamic negative constraints and treats consensus as a positive preference, achieving robust concept alignment and surpassing strong teachers on a high‑risk medical X‑ray benchmark.
APOCXR-MAXICML 2026
0 likes · 11 min read

MiniCPM‑V 4.6, a 1.3 B‑parameter multimodal LLM, outperforms larger rivals such as Qwen3.5‑0.8B and Gemma 4 on both accuracy and speed, thanks to early ViT token compression and 4×/16× visual token reduction, delivering sub‑100 ms latency and over 2.6 k token/s throughput on a single RTX 4090 while also running offline on mobile devices.
MiniCPM-VRTX 4090edge AI
0 likes · 16 min read

The authors demonstrate that visual perception, not reasoning, is the primary bottleneck for STEM multimodal large language models, introduce the CodePercept paradigm and the ICC-1M dataset, and show that code‑driven perception dramatically improves performance, surpassing much larger models on new benchmarks.
BenchmarkCVPR2026CodePercept
0 likes · 9 min read

The MME-Emotion benchmark, introduced by researchers from CUHK and Alibaba Tongyi and accepted at ICLR 2026, provides a large‑scale, multimodal evaluation of emotional intelligence in large language models, revealing current models’ limited emotion recognition and reasoning abilities across diverse real‑world scenarios.
AIBenchmarkMME-Emotion
0 likes · 10 min read

This article examines the visual token redundancy in decoder-only multimodal large language models and introduces a training-free dynamic computation reduction framework—featuring Probe-Activated Dynamic FFN, Hollow Attention, and a Layer Ranking Algorithm—that significantly lowers inference cost while preserving performance.
decoder-only architecturedynamic computationefficient inference
0 likes · 12 min read

Google DeepMind released the open‑source Gemma 4 family—four model sizes ranging from 2 B to 31 B parameters, supporting text, images, video and audio, with up to 256 k token context, Apache 2.0 licensing, and benchmark results that place it on par with the 397 B Qwen 3.5 despite being far smaller.
Apache 2.0BenchmarkGemma 4
0 likes · 11 min read

This article dissects Step‑Audio2, an industrial‑grade multimodal large language model that unifies speech understanding, translation, dialogue and audio generation in a single causal LM, detailing its inference pipeline, key implementation tricks, deployment modes, strengths, limitations, and suitable application scenarios.
PythonStep-Audio2Token2Wav
0 likes · 10 min read

This article presents MME-Emotion, a large‑scale multimodal benchmark that evaluates both emotion recognition and reasoning abilities of multimodal large language models across 27 real‑world scenarios, revealing current models’ significant gaps in emotional intelligence and outlining future research directions.
AIBenchmarkDataset
0 likes · 9 min read

A new multimodal browsing benchmark, BrowseComp‑V3, reveals that human experts achieve a 68.03% success rate while the strongest closed‑source model, GPT‑5.2, manages just 36.17%, highlighting current limitations in deep web‑scale visual‑text reasoning and the critical role of tool‑augmented agents.
GPT-5.2OmniSeekerProcess Evaluation
0 likes · 12 min read

LongHorizonUI tackles the steep success‑rate drop of GUI agents on tasks longer than 10‑15 steps by introducing three tightly coupled modules—enhanced perception, deep reflective decision, and compensatory execution—and validates the approach on the new LongGUIBench benchmark with consistent performance gains across both app and game scenarios.
BenchmarkGUI automationICLR 2026
0 likes · 12 min read

Lalamove tackled the scalability and accuracy challenges of worldwide driver onboarding by designing a layered edge‑cloud AI architecture that combines lightweight mobile models, cloud‑based large‑language and computer‑vision models, OCR, and multimodal LLMs to filter low‑quality inputs, automate identity checks, and reduce manual effort while maintaining data compliance.
AIDriver VerificationOCR
0 likes · 12 min read


AIWalkerMar 3, 2026 · Artificial Intelligence
RetouchIQ introduces an instruction‑driven AI retouching system that uses a general reward model to interpret abstract user commands, delivering precise image adjustments with higher semantic consistency and visual naturalness than existing multimodal large language models, thereby lowering the technical barrier for cinematic‑style edits.
AI Image EditingRetouchIQcreative automation
0 likes · 3 min read

This article analyzes why multimodal large language models often fail to detect small objects, identifies three core bottlenecks, and presents a four‑tiered optimization roadmap—from zero‑cost inference tricks to data augmentation, model fine‑tuning, and engineering safeguards—backed by three real‑world case studies and actionable guidelines.
Data AugmentationInference Optimizationmodel fine-tuning
0 likes · 20 min read

This guide details how to deploy the open‑source Qwen 3.5‑397B‑A17B multimodal LLM on Alibaba Cloud ACK using the RoleBasedGroup (RBG) engine, covering model preparation, Kubernetes resources, role‑based orchestration, performance tuning, and benchmark testing.
Cloud-Native AIKubernetesQwen3.5
0 likes · 24 min read

Alibaba’s Qwen3.5-397B-A17B, a 397‑billion‑parameter open‑source multimodal LLM, combines mixed linear attention with a sparse MoE architecture to achieve 8.6‑19× higher decoding throughput than Qwen3‑Max, supports 201 languages, and can be deployed via vLLM, Docker, Transformers, or SGLang with various optimization presets.
Inference OptimizationQwen3.5large language model
0 likes · 8 min read


JD TechJan 27, 2026 · Artificial Intelligence
Uni-Layout introduces a unified framework that integrates a universal layout generator, a human‑feedback‑simulating evaluator, and a dynamic margin preference optimization technique to align generation and evaluation across diverse e‑commerce design tasks, backed by a new 100k human‑annotated dataset.
Human Feedbackdynamic margin optimizatione-commerce design
0 likes · 11 min read

Uni-Layout introduces a unified framework that combines a multimodal large language model‑based generator, a human‑like evaluator trained on the large Layout‑HF100k dataset, and a Dynamic Margin Preference Optimization (DMPO) method to align generation and evaluation, achieving state‑of‑the‑art results across diverse layout tasks.
DMPOHuman Feedbackevaluation
0 likes · 11 min read

Uni-Layout introduces a unified framework that generates layouts across diverse tasks, simulates human evaluation with a novel feedback dataset, and aligns generation and assessment through dynamic margin preference optimization, achieving state‑of‑the‑art performance on multiple benchmarks.
AI designHuman Feedbackevaluation
0 likes · 11 min read

Uni-Layout introduces a unified layout generation framework that consolidates diverse design tasks, leverages multimodal large language models for flexible generation, and aligns outputs with human perception through a novel human‑feedback dataset (Layout‑HF100k) and a dynamic margin preference optimization (DMPO) evaluator.
ACM MultimediaHuman Feedbackdynamic margin optimization
0 likes · 11 min read

RAVEN is a reinforcement‑reasoning framework that combines curriculum learning with hierarchical rewards to enable multimodal large language models to accurately locate and classify violation segments in advertisement videos, even under noisy, large‑scale industrial data.
Advertisingcurriculum learningmultimodal LLM
0 likes · 17 min read

CrossVid is an open‑source benchmark that evaluates multimodal large language models on cross‑video reasoning tasks, providing 5,331 videos, 9,015 QA pairs, four high‑level dimensions and ten specific tasks, and exposing significant performance gaps between current models and humans.
AI-evaluationcross-video reasoningmultimodal LLM
0 likes · 9 min read

This weekly roundup highlights five recent AI papers—including HumanSense for multimodal LLM evaluation, JAM‑2 for de novo antibody design, the open‑source Olmo 3 language models, the Lumine generalist 3D agent, and the lightweight HunyuanOCR vision‑language model—summarizing their core contributions, results, and links.
OCROpen Source LLMgeneralist agents
0 likes · 6 min read

UNO‑Bench, an open‑source benchmark from Meituan’s LongCat team, provides the first high‑quality, low‑redundancy unified evaluation framework for omni‑modal large language models, featuring 1,250 manually annotated cross‑modal samples and 2,480 enhanced single‑modal samples covering 44 fine‑grained tasks and five modality combinations.
AI Scaling LawBenchmarkdata pipeline
0 likes · 15 min read


JavaEdgeNov 10, 2025 · Artificial Intelligence
This report outlines emerging AI trends for the coming year, highlighting Physical AI, AI agents, multimodal models, the MCP standard, and the growing role of AI‑driven DevOps, while also discussing security implications and adoption stages across the technology lifecycle.
AI AgentsAI DevOpsAI Trends
0 likes · 12 min read

iDetex, the champion solution of the ICCV 2025 MIPI Detailed Image Quality Assessment Challenge, introduces a novel multimodal LLM-driven framework that precisely locates, describes, and grades image distortions, outperforming traditional IQA models and enabling practical deployments across video, live streaming, e‑commerce, and image‑processing pipelines.
AIComputer VisionICCV 2025
0 likes · 18 min read

This week’s AI roundup highlights five recent papers: DeepSeek‑OCR’s context‑compression model for large‑scale data generation, Rex‑Omni’s 3‑billion‑parameter multimodal LLM achieving state‑of‑the‑art object perception, Alpha‑Service’s proactive AI‑glass framework, a bias‑variance approach to narrowing cross‑lingual gaps, and GATK’s MapReduce‑based toolkit for next‑generation DNA sequencing.
AI glassesCross-lingual NLPDNA Sequencing
0 likes · 6 min read

This article reviews the MLLM4TS framework, which fuses visual representations of multivariate time series with large language models to address complex temporal dependencies, cross‑channel interactions, and task generalization, and demonstrates superior performance on classification, anomaly detection, forecasting, and few‑shot scenarios across multiple benchmarks.
Few-Shot LearningVision-Language Fusionablation study
0 likes · 11 min read

This article details how Bilibili’s multimedia lab leveraged a multimodal training pipeline combining data‑compressed SFT and the GRPO reinforcement‑learning algorithm to achieve a 13.5% metric boost and secure second place in the ICCV MIPI Detailed Image Quality Assessment competition.
GRPOMIPI competitionSFT
0 likes · 15 min read


Amap TechOct 4, 2025 · Artificial Intelligence
JanusVLN presents a groundbreaking Vision‑and‑Language Navigation framework that decouples semantic understanding from spatial geometry using dual implicit memory, eliminates explicit memory overhead, achieves state‑of‑the‑art performance with only RGB video input, and dramatically improves efficiency and generalization across VLN benchmarks.
3D spatial reasoningDual Implicit MemoryVision-Language Navigation
0 likes · 10 min read

FinZero is a multimodal large‑model that leverages a 30‑billion‑parameter Qwen2.5‑VL backbone fine‑tuned with the UARPO strategy on the FVLDB dataset, enabling accurate financial time‑series prediction, uncertainty quantification, and outperforming larger models such as GPT‑4o by about 13.5% in high‑confidence groups.
FinZeroGPT-4o comparisonQwen2.5-VL-3B
0 likes · 10 min read

Keye‑VL‑1.5, an 8‑billion‑parameter multimodal large language model, introduces a Slow‑Fast video encoding strategy, a four‑stage progressive pre‑training pipeline with 128K context, and a sophisticated post‑training regime that together achieve state‑of‑the‑art performance on video and vision‑language benchmarks while maintaining strong general capabilities.
BenchmarkPretraininglarge language model
0 likes · 21 min read

This article surveys five recent AI papers, covering reinforcement learning for large reasoning models, a tree‑structured table QA framework (ST‑Raptor), visual representation alignment for multimodal LLMs, GraphRAG‑based generation, and an LLM‑driven cryptographic vulnerability detector, each with key insights and links.
cryptographic vulnerability detectiongraph retrievallarge language models
0 likes · 5 min read

Kuaishou's Kling-Avatar leverages a multimodal large‑language‑model‑driven two‑stage generation framework to produce minute‑long digital‑human videos that synchronize lip movements, facial expressions, and body gestures with audio, achieving high visual quality, identity consistency, and controllable storytelling across diverse scenarios.
AI AvatarDigital Humanlong video synthesis
0 likes · 9 min read

This article presents a training‑free, architecture‑agnostic method that leverages LaTeX‑style structured inputs to preserve document hierarchy and spatial relationships, thereby improving multimodal large language model performance on document question answering tasks across multiple benchmarks.
AIDocQADocument Understanding
0 likes · 8 min read

Kwai's newly released Keye-VL-1.5-8B multimodal large language model dramatically improves visual, reasoning, and temporal understanding, achieving top scores on public video benchmarks and surpassing closed‑source models like GPT‑4o, while offering an open‑source release and detailed technical documentation.
benchmark performancemultimodal LLMopen-source
0 likes · 11 min read

A retrospective of an AI‑based insurance claim audit project shows that a well‑designed workflow, precise prompt engineering, and rule‑based pre‑filtering can achieve stable, high‑accuracy results, while overly complex agent architectures often become fragile patchwork solutions.
AI auditWorkflow Designinsurance claim
0 likes · 24 min read

Perception‑R1, a post‑training framework that applies rule‑based reinforcement learning to existing multimodal LLMs, dramatically improves visual perception tasks such as grounding, OCR, counting and object detection, as demonstrated by extensive benchmarks and ablation studies.
GRPOPerception PolicyReward Modeling
0 likes · 10 min read

DiMo-GUI is a plug‑and‑play framework that dramatically improves multimodal large language models' ability to locate GUI elements by using a hierarchical dynamic visual reasoning loop and modality‑aware optimization, achieving up to double the performance on high‑resolution GUI benchmarks without any additional training data.
GUI groundingTest-Time Scalingdynamic visual reasoning
0 likes · 7 min read


AIWalkerAug 13, 2025 · Artificial Intelligence
Look-Back is an implicit training paradigm that enables the Qwen‑2.5‑VL‑7B multimodal LLM to autonomously re‑focus on visual inputs during reasoning, achieving a 6.3 % boost in perception tasks, outperforming prior baselines while requiring no extra image tokens or model architecture changes.
Look-BackQwen-2.5-VLimplicit training
0 likes · 26 min read

This article compares two ways to add vision to large language models—training a native multimodal model from scratch or attaching a visual module to a pretrained LLM—then details the VoRA approach that uses LoRA adapters to inject visual knowledge without extra inference cost.
ChameleonLLaVALoRA
0 likes · 7 min read

This article introduces a language‑centric tree reasoning (LTR) framework that recursively decomposes VideoQA queries into perceptual sub‑questions and performs bottom‑up logical inference with video assistance, achieving significantly higher accuracy and explainability across eleven benchmark datasets.
Tree ReasoningVideoQAartificial intelligence
0 likes · 17 min read


AIWalkerAug 5, 2025 · Artificial Intelligence
The paper introduces Perception‑R1, a rule‑based reinforcement‑learning framework that trains multimodal large language models for visual perception tasks without relying on chain‑of‑thought reasoning, and demonstrates up to 17.9% performance gains on RefCOCO+, PixMo‑Count, PageOCR and COCO2017, while analyzing the key roles of perception confusion and reward design.
BenchmarkRLHFReward Design
0 likes · 24 min read


AIWalkerAug 4, 2025 · Artificial Intelligence
CAIG leverages a multimodal large language model, a novel reward model, and product‑centered preference optimization to generate ad images that maximize click‑through rate, achieving state‑of‑the‑art performance in both online and offline evaluations.
CTRad image generationmultimodal LLM
0 likes · 7 min read


AIWalkerAug 3, 2025 · Artificial Intelligence
DeQA-Score introduces a soft‑label discretization that lets multimodal large language models regress continuous image‑quality scores as Gaussian distributions, achieving 30× lower mean error and preserving variance and inter‑image relationships, with KL‑divergence and fidelity losses driving state‑of‑the‑art performance.
CVPR2025DeQA-Scoreimage quality assessment
0 likes · 8 min read

Meituan’s research team showcases eight ACL 2025 papers spanning generative retrieval, multi‑objective preference alignment, rich‑text image understanding, cross‑language transfer, multimodal math reasoning, and more, offering insights and breakthroughs that can inspire and aid fellow researchers.
ACL 2025Code-Switchinggenerative retrieval
0 likes · 15 min read


Amap TechJul 24, 2025 · Artificial Intelligence
The paper introduces FingER, an entity-level evaluation framework and the FingER-Instruct-60k dataset for assessing AI-generated video quality with fine-grained reasoning, and demonstrates state-of-the-art zero-shot performance on multiple benchmarks using novel training strategies.
AI-generated videoDatasetfine-grained evaluation
0 likes · 9 min read


AntTechJul 2, 2025 · Artificial Intelligence
This article details how Alibaba's Ant Group leverages multimodal large‑language models and multi‑agent architectures to create a low‑code, AI‑driven UI automation testing framework that improves test coverage, reduces manual effort, and scales across diverse mobile mini‑program scenarios.
AI TestingUI automationagent-based testing
0 likes · 9 min read


AIWalkerJun 30, 2025 · Artificial Intelligence
The ICCV MIPI workshop introduces the ViDA-UGC competition, presenting a richly annotated UGC image quality dataset, a benchmark suite covering degradation detection, region perception, and quality description, detailed evaluation metrics, submission formats, prize information, and open participation for researchers worldwide.
BenchmarkDatasetICCV
0 likes · 15 min read

ChatTS is a time‑series‑native multimodal large language model trained on purely synthetic data, offering superior understanding and reasoning over both real and synthetic time‑series datasets, and outperforming existing LLM baselines across alignment and inference tasks.
AILLM AlignmentTS‑MLLM
0 likes · 18 min read


Amap TechApr 21, 2025 · Artificial Intelligence
At ICASSP 2025, Gaode’s two accepted papers present Lenna, a language‑enhanced reasoning detection assistant that adds a DET token to multimodal LLMs and achieves state‑of‑the‑art accuracy on RefCOCO benchmarks, and a chain‑of‑thought image‑editing framework that converts complex prompts into segmented masks and repair prompts for diffusion‑based inpainting, surpassing existing methods.
AIComputer VisionICASSP
0 likes · 10 min read


AIWalkerApr 11, 2025 · Artificial Intelligence
DeQA-Score, a CVPR 2025 work, shows how to train multimodal large language models to regress continuous image quality scores by discretizing scores into soft-label level tokens, preserving Gaussian distribution statistics and achieving state‑of‑the‑art performance without any installation.
CVPR2025DeQA-Scoreimage quality assessment
0 likes · 8 min read

This guide introduces Hugging Face's TRL library, explains how to install it alongside Transformers, and walks through modifying LLaVA's trainer, dataset, and data collator to apply the DPO reinforcement‑learning algorithm for multimodal model fine‑tuning.
DPOHugging FaceLLaVA
0 likes · 4 min read

ChartMoE, an oral paper at ICLR 2025, introduces a multi‑stage alignment training pipeline and a diversified MoE Connector that dramatically improves chart comprehension while maintaining performance on general multimodal tasks, backed by extensive data construction, training recipes, and thorough evaluations.
ChartMoEMixture of Expertschart understanding
0 likes · 10 min read

This article explores how multimodal large language models (MM‑LLMs) combined with structured prompt engineering can transform frontend regression testing by enabling natural‑language‑driven test case generation, visual verification, and script self‑healing, thereby reducing maintenance costs and improving coverage across dynamic UI scenarios.
AI automationNL2Testmultimodal LLM
0 likes · 17 min read


JD TechMar 26, 2025 · Artificial Intelligence
The JD advertising team proposes a CTR‑driven advertising image generation framework (CAIG) that leverages multimodal large language models, a novel reward model, and product‑centric preference optimization to produce ad images with superior click‑through performance, validated by extensive offline and online experiments.
CTR optimizationadvertising image generationmultimodal LLM
0 likes · 10 min read

This tutorial shows how to build a LangChain pipeline that uses GPT‑4 Vision to read vehicle images, defines a Pydantic model for type, license, make, model and color, and returns the results as structured JSON for both single and batch inference.
GPT-4 VisionLangChainPydantic
0 likes · 11 min read

This article provides an in‑depth review of alignment algorithms for multimodal large language models, covering application scenarios, dataset construction methods, evaluation benchmarks, current challenges, and future research directions, while summarizing contributions from leading academic institutions.
AI researchalignment algorithmsdataset construction
0 likes · 22 min read

This article surveys the latest multimodal large language model research, dissecting the design, training strategies, and performance trade‑offs of models such as Llama 3.2, Molmo, NVLM, Qwen2‑VL, Pixtral, MM1.5, Emu3, and Janus, and highlights the challenges of fair cross‑model evaluation.
AI researchModel Training Strategiescross-attention
0 likes · 16 min read

This article explains the fundamentals of multimodal large language models, covering their definition, typical applications, two main architectural approaches—unified embedding decoder and cross‑modal attention—along with detailed component breakdowns, a PyTorch implementation of image‑patch projection, and training considerations, ending with a discussion of trade‑offs between the methods.
Image EncoderLinear ProjectionPyTorch
0 likes · 14 min read


AntTechMar 14, 2025 · Artificial Intelligence
The CVPR 2025 paper "MP-GUI: Modality Perception with MLLMs for GUI Understanding" presents a novel algorithm that enhances multimodal large language models' ability to perceive and reason about graphical user interfaces by integrating text, visual, and spatial signals through specialized perception modules and a dynamic fusion gate, achieving state‑of‑the‑art performance on multiple GUI benchmarks.
CVPR2025Computer VisionGUI Understanding
0 likes · 5 min read

The paper presents CAIG, a CTR‑driven advertising image generation pipeline that pre‑trains a multimodal LLM on e‑commerce data, trains a reward model on CTR‑labeled image pairs, and fine‑tunes generation via product‑centric preference optimization, achieving state‑of‑the‑art online and offline performance.
AICTRad image generation
0 likes · 11 min read

This article provides a comprehensive overview of AI agents, focusing on GUI‑based agents, their definitions, classifications, core capabilities, recent research such as OpenAI's ComputerUse, SpiritSight and MobileFlow, practical applications, technical and security challenges, and future development directions.
AI AgentsComputerUseGUI automation
0 likes · 16 min read

This paper presents a CTR‑driven advertising image generation framework that leverages multimodal large language models, reward modeling, and reinforcement learning to produce product‑centric ad visuals with higher click‑through performance, validated by extensive offline and online experiments.
CTR optimizationadvertising image generationmultimodal LLM
0 likes · 13 min read

Xiaohongshu’s team unveiled a self‑developed RLHF system that trains multimodal large language models using heterogeneous and homogeneous network architectures, extensive PPO optimizations, and Medusa speculative sampling, achieving over 50% throughput gains, reduced hardware needs, and 5‑20% performance improvements on zero‑shot benchmarks.
PPOPRMRLHF
0 likes · 21 min read

This guide shows how to set up the open‑source Ollama‑OCR tool, which leverages the Llama 3.2‑Vision multimodal model to perform high‑quality OCR, covering installation of Ollama, the vision model, the OCR package, and example code for plain‑text and Markdown outputs.
Llama 3.2-VisionNode.jsOCR
0 likes · 6 min read


NewBeeNLPNov 11, 2024 · Artificial Intelligence
This article surveys ten recent multimodal large language model papers, covering vision representation laws, a stricter instruction benchmark, safety impacts of visual adaptation, the Mini‑Gemini architecture, automatic pruning, vision capability boosting, long‑context transfer, efficient token sparsification, math reasoning, and hallucination mitigation.
BenchmarkEfficiencyTraining Strategies
0 likes · 18 min read

This article reviews recent advances in multimodal large language models, covering their background, architectural components, training strategies, application scenarios, evaluation benchmarks, team research on hallucination mitigation and long‑video understanding, and outlines promising future research directions.
evaluation benchmarksfuture researchmodel architecture
0 likes · 15 min read

NoteLLM-2 introduces multimodal In-Content Learning and Late Fusion to overcome visual‑modality bias in end‑to‑end fine‑tuned large representation models, delivering significant gains over baseline multimodal LLMs and traditional retrieval methods in recommendation tasks.
AI researchContrastive Learningin‑content learning
0 likes · 11 min read

Meta’s newly released 92‑page Movie Gen paper introduces a multimodal LLM that unifies text‑to‑image, text‑to‑video, personalized video, precise video editing, and audio generation, detailing its dual‑model architecture, training pipeline, temporal auto‑encoder design, scaling strategies, evaluation benchmark, and ablation studies.
Model Scalingdeep learningevaluation
0 likes · 34 min read

The article reviews why visual ability is essential for artificial general intelligence, compares native multimodal and expert‑stitching integration approaches, details the architectures of models such as KOSMOS‑1, PALM‑E, Flamingo, BLIP‑2, LLAVA, miniGPT‑4, and introduces the SEEChat project that fuses CLIP vision encoders with chatGLM6B via a projection layer, presenting its training pipeline, experimental results, and future directions.
AGIImage CaptioningSEEChat
0 likes · 13 min read
