Muir improves linear attention with gating and readout
Muir reduces perplexity by 1.8-2.0% over GDN-2 and supports 10M-token extrapolation.
Stories per day · 30 days
5 RSS
Relatedbenchmark6multimodal6jailbreak3grpo2mobile-agents2qwen2robotics2video-captioning2activation-steering1active-exploration1active-perception1adversarial-images1agentic-ai1agent-security1attention1attention-heads1autoregression1backdoor1
55 stories, newest first · one card per story
Stories tagged #mllm
MAEP prunes MoE multimodal experts by modality in one pass
The training-free method cuts about 27 GB of BF16 expert weights so an about 58 GB model fits a 40 GB GPU.
Reviso prioritizes key visual evidence in video captioning
Reviso builds CFP observation plans and trains with GRPO to outperform baselines on four benchmarks and improve downstream video question answering.
Jailbreak Connectivity links image paths for transferable MLLM jailbreaks
On MM-SafetyBench, reported success reaches 97.95% on Gemma-3-12B-IT and beats FigStep on GPT-5.6 and Gemini 3.1 Pro.
VideoLVR adds latent visual evidence to MLLM video reasoning
Latent tokens reconstructed from the input video let models revisit local frames without large context growth.
CodeCap-Bench asks whether code captions preserve video content
On 323 structured 2D videos, code captions beat natural language on question answering and reconstruction, but models still miss complex motion.
GPT-6-Astra evaluated on EBench for direct and policy-harness robot control
Direct control leads on short tasks but lags on long-horizon work; calling a VLA or WAM matches top policies at lower token cost.
SATE-Bench tests MLLMs on active exploration of occluded scenes
The benchmark has 316 scenes and 2,525 multi-turn pairs; GRPO increases view-seeking but lowers accuracy.
InteractiveWorld benchmarks egocentric MLLM assistants on 120 tasks
Across 12 frontier models, Claude Fable 5.1 reaches at most 20% and 25% on long-horizon tasks.
Retrieval heads stay sparse and shared across MLLM modalities
Ranking training on those heads raises Qwen3-VL-8B page Recall@1 on MMDocIR by 6.03 points.
FAPL trains MLLM latents to control cinematic video generation
A diffusion transformer is grounded on frozen MLLM features, then the model is tuned for sparse prompts and single-factor edits.
TARSphere benchmarks long-horizon robots in changing physical and social scenes
The benchmark has 360 episodes across 30 indoor scenes, and tests of eight MLLMs found updates and human interaction still hard.
CausGround grounds action preconditions for embodied robot planning
Without parameter updates, reported VirtualHome success rises from 24.9% to 53.3% and RoboCasa from 43.0% to 57.0%.
Cross-modal triggers activate backdoors in multimodal LLMs
From a backdoor in one modality, the method builds a trigger in another without poisoning that modality during training.
MobileInject: Evaluating Reasoning-to-Action Security of MLLM-Based Mobile Agents under Cross-Modal Consistency Injection
MobileInject benchmarks MLLM mobile agents against multimodal injection across 108 risk instances, finding broad vulnerability.
MobileJailBench benchmarks jailbreaks against MLLM mobile agents
The 520-task suite spans 14 risk categories and reports 100% ASR on Qwen3.5-122B after dynamic evolution.
MARS steers multimodal safety from textual refusal directions
The training-free method was tested on five state-of-the-art MLLMs for safety, utility, and video jailbreaks.
Event-STVG benchmark tests long-form event reasoning in video
The strongest tested MLLM reached only 5.4% m-vIoU at a 10-minute horizon.
WinARDet cuts repeated-box loops in dense vision-language grounding
On RSDense, authors report F1 rising from 32.16 to 39.36 and redundancy falling from 35.48% to 13.00%.
When MLLMs miss what humans read: diagnosing and learning to read squint text
Squint OPSD lifts multimodal LLM reading of hard-to-see image text from 26.50 to 88.61 ANLS.
S-Space: Exploring Spatial Workspace in Multimodal Language Models
Researchers identify S-Space, a linear subspace where multimodal models encode object coordinates.
WOVEN: Weaving Visual World Modeling into Multimodal LLMs
WOVEN trains multimodal LLMs on visual transitions, lifting 22 of 26 external benchmarks by up to 27.3 points.
SuperNav: An Agentic Navigation System for Any Task in Any Scene
SuperNav uses an unfine-tuned MLLM plus navigation tools for general robot navigation.
Humanity's Sixth Sense: Benchmarking Intuitive Visual Reasoning in Multimodal Models
Researchers introduce HSS, where humans score 93.1% and GPT-6-astra only 53.6% on intuitive visual reasoning.