Tag
Xiaokang Chen shares two prompts, 'Think with Grounding' and 'Think with Pointing', to improve model performance in domains like counting in Thinking mode. These prompts use bounding boxes and points to make the MLLM's reasoning more human-like.
MemGUI-Agent introduces proactive context management for long-horizon mobile GUI tasks, using Context-as-Action (ConAct) to maintain critical information. It includes the MemGUI-3K dataset and achieves state-of-the-art performance on MemGUI-Bench and MobileWorld benchmarks with an 8B model.
A new benchmark called StylisticBias systematically evaluates attribute-level social bias in multimodal large language models, finding that a small set of visual cues like fashion style drive most biases.
This paper addresses the problem of spoken language adherence in multimodal LLMs for ASR, proposing a soft prompting approach and novel metric to quantify language violations. It evaluates three mitigation strategies—zero-shot prompting, supervised fine-tuning, and chain-of-thought reasoning—across multiple languages to improve transcription fidelity.
This paper introduces MODE, a modality-decomposed expert-level mixed-precision quantization framework for MoE multimodal LLMs that addresses biases in expert importance estimation by decomposing selection frequency by modality and filtering redundant vision tokens, achieving minimal performance loss under aggressive quantization.
This paper introduces the Forced Deferral Attack (FDA), an adversarial image attack that manipulates confidence scores in multimodal LLM cascades, causing queries to be unnecessarily routed to stronger (more expensive) models, thereby shifting compute costs to the provider without degrading answer correctness.
Visual-Seeker proposes a visual-native multimodal deep search agent that actively reasons over fine-grained visual details and synthesizes multimodal evidence, achieving state-of-the-art performance on five challenging multimodal search benchmarks.
RepFusion introduces a method to use pretrained multimodal LLMs as noisy representation encoders in diffusion transformers for text-to-image generation, outperforming baselines with similar compute.
SAGA framework uses frozen multimodal large language models to provide attribute-aware supervision for vision encoders via Group Relative Policy Optimization, improving zero-shot image retrieval by 3–6 points on fine-grained benchmarks.
The paper introduces UXBench, a multimodal benchmark for evaluating MLLMs on mobile UX reasoning tasks, and presents UI-UX, a fine-tuned MLLM based on Qwen3-VL-4B-Thinking that achieves state-of-the-art performance on this benchmark.
The paper proposes SVoT, a reinforcement learning framework that generates interleaved, verifiable intermediate states and visualizations for multi-hop spatial reasoning in MLLMs, achieving significant accuracy gains on new benchmarks involving multi-object interactions and numerical reasoning.
ART (Art-based Reinforcement Training) enables parameter-efficient fine-tuning of frozen multimodal LLMs by optimizing raw visual input via gradient backpropagation, achieving performance comparable to LoRA while supporting pre-compiled computational graphs for high-throughput engines like vLLM.
This paper introduces MGAP, a training-free decoding method that reduces hallucinations in Multimodal Large Language Models by adaptively suppressing only the harmful parts of language priors while preserving the model's semantic manifold. The method outperforms prior baselines on POPE and CHAIR benchmarks.
This paper studies how audio and visual information flow inside Audio-Visual Large Language Models (AVLLMs), revealing that AVLLMs follow sequential or parallel routing depending on input configuration, and that some tokens can be discarded after information transfer for efficiency.
This paper proposes a query-based cross-modal projector that compresses visual tokens via cross-attention to improve Mamba-based multimodal LLMs, boosting both performance and throughput on vision-language benchmarks while eliminating the need for manual 2D scan order design.
Researchers from Jilin University systematically evaluate positional bias in multi-video summarization using MLLMs, constructing a benchmark from ActivityNet and News videos and assessing nine models with metrics including Coverage, Directional Positional Bias, and Middle-Edge Gap. Results show positional effects are domain- and model-dependent, and increasing visual or generation budget does not uniformly resolve the imbalance.
VCIFBench is a new benchmark for evaluating complex instruction following in video understanding, featuring 306 test instructions with content, format, style, and structure constraints, plus a DPO preference dataset. Experiments on 10 MLLMs reveal that joint constraint satisfaction remains challenging, and DPO training on the benchmark data improves instruction-following performance.
BiNSGPS is a framework that introduces bidirectional interaction between a multimodal LLM adviser and a symbolic solver for geometry problem solving, allowing feedback from the solver to correct errors and generate auxiliary hypotheses. It achieves state-of-the-art performance of 90.5% on Geometry3K and 90.1% on PGPS9K benchmarks.
VAMPS is a new benchmark of 1,168 multimodal bilingual math problems designed to evaluate whether LLMs can benefit from constructing and reasoning over graphs/visualizations. Key finding: direct analytical solving surprisingly outperforms tool-enabled visual solving even on problems where plotting is a natural strategy.
Introduces Future-L1, an interleaved latent visual reasoning framework that improves video event prediction by maintaining visual semantics in latent space. Achieves state-of-the-art results on FutureBench and TwiFF-Bench benchmarks.