Google launches Gemini Decisions API and Gemini Decision 1
One request can score up to 1,000 closed-set questions for $0.10 per million input tokens, with output tokens free.
1
Stories per day · 30 days
19 RSS
Relatedbenchmark19open-weights16reinforcement-learning16benchmarks9agents8evaluation6mllm6vision-language6code-generation5dataset5huggingface5meta5mixture-of-experts5muse-spark-open5visual-reasoning5contrastive-learning4image-generation4lfm34
289 stories, newest first · one card per story
Stories tagged #multimodal
GARD cuts CLIP visual backdoor success below 1%
The post-hoc defense covers six visual backdoor attacks on two CLIP backbones and needs no extra inference compute.
Study rechecks where Platonic representation alignment appears in newer models
The capability-alignment trend holds when captions are clean, but same-modality alignment still far exceeds cross-modal scores.
dNaViT encodes images as discrete general-purpose visual tokens
Authors say it matches or beats task-specific baselines on 12 understanding, 4 generation, and one editing benchmark.
EmoHarness raises small multimodal emotion scores with audio evidence
On 5,507 MME-Emotion items, Recognition rose 7.6-13.4 points, mostly from transcripts and prosody rather than evolved edits.
SurgKnowBench and SurgAgentGym test generalist models on surgery scenes
Agentic tool use mainly helps perception-heavy tasks; localization and expert-style assessment stay weak.
Thinky Machines releases Inkling 2 Small, 594B MoE beating Muse Spark
Apache 2.0 weights ship next week after safety testing; ranks 31 on Artificial Analysis Index.
SkinMind uses graph probing for dermatology retrieval-augmented generation
Authors report average gains of 8.70 and 8.89 points with GPT-5.4-mini and Gemini-2.5-Flash on six VQA benchmarks.
NPCBench benchmark evaluates multimodal care-pathway reasoning for nasopharyngeal carcinoma
NPCBench includes 4,777 items and a 100-patient CarePath cohort with 417 scored tasks, revealing composition gaps where strong individual model performance does not ensure jointly correct diagnostic and care answers.
Micius-Flash, a 27B open engineering model, is released
It averages 64.2% on hardware code benchmarks and 62.4% on a semiconductor agent benchmark, above a same-size baseline.
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.
Kling multimodal understanding insights for video generation
Kling-MLLM unifies image video audio inputs; captioning and prompt enhancement act as semantic contracts that maximize description density in the generation chain.
YANAMI Dataset and Benchmark Released for Multimodal Stateful Interaction
YANAMI introduces the first large-scale multimodal dialogue dataset with 400K multi-turn interactions across four scenarios.
VidUEU-Agent Data-Curation Pipeline for Multimodal Tasks
VidUEU-Agent synthesizes 100K instruction samples from raw videos for multimodal understanding, editing, and unified tasks.
Factorized MLLM embeddings for composed image retrieval
FactEmbed decouples compositional query comprehension from discriminative embeddings using factorized latents and contrastive losses for improved CIR performance.
LongCat-2.5 model released with native multimodality for agentic tasks
LongCat-2.5 features 1.6 trillion parameters and native visual understanding for long-horizon agentic tasks.
1
Lab2Sim benchmark turns science videos into interactive experiments
Lab2Sim provides 218 videos and EduVideo2Code improves apparatus fidelity by 24.15 points.
Moonshot AI releases Kimi K3.1 Pro open-weight multimodal model
Kimi K3.1 Pro is an open-weight MoE model with 1M token context and 104B activated parameters.
1
LARC enables continuous latent reasoning for multimodal image generation
LARC uses curriculum SFT and information-gain RL to bridge discrete token and continuous image spaces, delivering 7% GenEval gains and test-time scaling.
LARC advances continuous-space reasoning in multimodal models
LARC reasons over continuous hidden states to fix token bottlenecks, boosting image generation quality and test-time scaling on GenEval.
Google releases MTG-AD benchmark for inline audio descriptions
The 36-hour set has more than 8,000 descriptions, and end-to-end scoring checks whether narration fits existing video silences.
Elorian launches Glimpse and the DecoyCount visual benchmark
The paper puts frontier models under 50% accuracy; CEO Dai says Glimpse scored 50.48% on one run.
6
Paper turns the Platonic Representation Hypothesis into a JEPA theorem
A 109M-parameter text JEPA beat language models up to 72B on alignment with DINOv2 and CLIP image encoders.
Youtu-Parsing-Omni parses documents, audio, and video
A single 5B checkpoint is reported at 96.96 on OmniDocBench v1.6 and 75.08 average on OmniParsingBench.