Multimodal efficiency/long-video/speech (Tuna-2, World-R1, IAM)

CoPD unifies; Audio AKB +10%; Visual +18%; Persona>4o; MATHNET; UniVidX; Persistent Visual Memory LVLMs; OceanPile corpus. New: Pantheon360 3D-aware 360° video diffusion; Native Multimodal Modeling roadmap; WBench multi-turn world model eval; adversarial flow distillation for video; ParaVT parallel tool use in video RL; EvalVerse pipeline-aware benchmark for cinematic video generation; SpatialBench comprehensive spatial foundation model benchmark (41 models, 6 paradigms). Also: Gemini Embedding 2 native multimodal embedding model (unifies video, audio, image, text, SOTA on MSCOCO, Vatex, MTEB). New: NEO-ov native one-vision model (end-to-end pixel-word learning, 2B/8B, SOTA on Video-MME and spatial tasks). New: NAVA (native audio-visual alignment for generation, 6.3B params, Align-then-Fuse MMDiT, strong sync and timbre control). New: SmartDirector (keyframe-conditioned cinematic video generation with narrative pacing control, two-stage pipeline). Also: Gamma-World multi-agent world model (Simplex Rotary Encoding, Sparse Hub Attention, 24 FPS); GEM generative supervision for embodied AI (depth map generation, SOTA on benchmarks); FGO frequency-guided action diffusion for robotic manipulation (improves smoothness/success). New: NeuROK (generative 4D dynamics, CVPR 2026, latent kinematic space); PhyGenHOI (physically-aware 4D HOI generation, windowed attraction loss); OSP-Next (efficient video generation, 2x speedup via sparse sequence parallelism, HiF8, RL); AdaState (self-evolving anchors for streaming video generation, improved dynamics). New: minWM (full-stack open-source framework for real-time interactive video world models, camera control, distillation for low-latency rollout). New: YoCausal (causality benchmark for video diffusion, reveals arrow-of-time perception ≠ causal understanding, challenges world model claims). New: LocateAnything (parallel box decoding for visual grounding, 138M dataset, May 2026). New: GPIC (28 trillion pixel permissive image dataset for generative models). New today: ChildVox (speech/audio benchmark across childhood, 17 datasets). Also: FlatSounds benchmark reveals video-to-audio models cheat by relying on text cues rather than true physical understanding, challenging world model claims. New: SANA-Streaming (real-time streaming video editing at 24 FPS on RTX 5090); Light Interaction (training-free inference acceleration for interactive video world models, 2.59x speedup); Linear Scaling Video VLMs (StateKV, linear scaling for long video without fine-tuning). Also: Representation Forcing (removes VAE bottleneck in unified multimodal models, matches VAE-based generation). New: SpatialUncertain — controlled framework for VLM abstention on spatial questions, adds to multimodal reliability evaluation. New articles: VideoMLA (low-rank latent KV cache for minute-scale autoregressive video diffusion, 92.7% reduction, 1.23x throughput), VLMs are Good Teachers (test-time optimization for video reasoning, 16.7-point gain over baselines), StreamChar (long-horizon streaming character audio-video generation, real-time with decoupled orchestration). New: World Models Meet Language Models (PF-OPSD, privileged future context, +10.6% on VRQABench). New: TRON (online generator-verifier for visual reasoning RL, 520 environments, curriculum control). New: Stable-Layers (VLM-guided image layer separation using RL+LoRA, practical for graphic design). New: NVIDIA Cosmos 3 (omnimodal world models, two-tower MoT, open-source, unifies physical reasoning, world generation, action generation). New: AAD-1 (asymmetric adversarial distillation for one-step autoregressive video generation, SOTA on VBench, ICML 2026). New: DCRL (wide-baseline matching for spatial reasoning in MLLMs, RL without CoT, human 84.0 F1 vs best baseline 37.2). New: Gemma 4 12B (encoder-free multimodal, laptop-ready, near 26B MoE performance). New: World-Language-Action Model (WLA, unified world modeling, language reasoning, action synthesis, SOTA on RoboTwin2.0 and RMBench, 2B params, 40ms inference). New: Imagine Before You Predict (interleaved latent visual reasoning for video event prediction, +24.4 points on FutureBench). New: Discrete-WAM (unified discrete vision-action token editing for world-policy learning in autonomous driving). New: OVO-S-Bench (streaming spatial benchmark for MLLMs, hierarchical taxonomy, prefix-only protocol). Also: ChartNet (synthetic chart data pipeline, small open-source models beat commercial giants on chart understanding). New: Towards One-to-Many Temporal Grounding (ICML'26, new task exposing MLLM blind spots in multi-segment queries). New: AdaCodec (predictive visual code for video MLLMs, 1/7 token budget, 5.7x TTFT reduction). New: Flash-WAM (modality-aware distillation for world action models, 23x speedup). New: Video2LoRA (parametric video internalization, 1500x token reduction). New: Stateful Encoders (VLMs with visual memory via cross-attention and stop-gradient, consistent gains across backbones, practical for multi-image reasoning). New: WorldBench (visually diverse multimodal reasoning benchmark, top model 64%). New: MMAE benchmark for audio editing (<5% exact match, 0% on complex mixed-modality tasks). New: Stream3D-VLM (online 3D spatial understanding from streaming video, incremental geometry priors). New: MemDreamer (hierarchical graph memory for long video, 12.5 point gain, 2% context). New: Mirage (latent spatial memory for 3D scenes, Microsoft Research). New: Los Alamos PAS hallucination detector for VLMs (CVPR 2026) — real-time detection, relevant to multimodal safety. New: world model paper using 2D stick-figure skeletons for conditioning (ex-19447889) — texture-free trick for robot policies, MMRV 0.57 vs 1.43/0.71 baselines. New: world models as weak link in home-robot pipelines (tweet, ex-e8639108). New: Kairos native world model stack (ex-e4011b3d). New: DreamX-World 1.0 general-purpose interactive world model (ex-64362cda). New: BadWorld adversarial attack on visual world models (ex-ec0cb91c); MVEB video embedding benchmark (ex-80dd0135). New: PAIWorld (3D-consistent world foundation model for robotic manipulation) — adds to world model influx. New: Reinforcing Dual-Path Reasoning in Spatial VLMs — advances spatial reasoning in VLMs. New: RNG-Bench (ex-a50875b5) — testing memory in MLLMs via non-Markov games, Memory Gap metric, top models struggle on spatial memory. New: Mind-Studio executable world models with lookahead evaluation for partially observable games (podcast, ex-e42dfb62). New: Current World Models Lack a Persistent State Core (critique, ex-7e087bac). New: Human videos to 4D robot hand-object trajectories (tweet, ex-5f291c9f) — advance for imitation learning. New: TriViewBench (controlled benchmark for multi-view structural reasoning, reveals occlusion blindness and cross-view identity confusion, CoT useless). New: Hallucination in World Models is Predictable and Preventable (three failure modes, MMBench2, coverage-aware sampling) adds diagnostic reliability improvement. New: Valdi (value diffusion world models, single-step MPC, trade-off between multimodality and control) adds new world model approach. New: Perceive-to-Reason (decoupling perception and reasoning for fine-grained visual reasoning, PRA-GRPO, strong on V-Star/HR-Bench) improves VLM reasoning. New: Multimodal Continuous Reasoning via AMVL (asymmetric mutual variational learning, +10.83 on BLINK) addresses train-inference mismatch in continuous latent reasoning. New today: WorldDirector (controllable world simulators with persistent dynamic memory, decouples semantic motion from visual generation), AnyGroundBench (video grounding benchmark for specialized domains, VLMs fail on rare concepts). New: DataComp-VLM (improved open datasets for VLMs, data mixing > filtering, +5.4pp over FineVision). New: GigaWorld-1 (systematic study of world models for robot policy evaluation, 324k rollouts, long-horizon action-faithful rollout consistency > short-term visual realism) adds to world model evaluation infrastructure. New today: Flex-Forcing (unified bidirectional/autoregressive video diffusion, better quality and faster inference), Vision as Unified Multimodal Generation (reformulates visual tasks as multimodal generation, matches specialized systems), MuseBench (intent-level artistic understanding benchmark, best model 48% vs human 87%), Parallelized Autoregressive Decoding for dense video captioning (lossless parallel generation), Light-Omni (reflexive video agent, 12.1x speedup, 2.6x memory improvement).

Status: Climaxing — world models and multimodal efficiency continue to advance. New diagnostic papers (YoCausal, FlatSounds, TriViewBench) challenge world model claims. The GigaWorld-1 study provides systematic evaluation. New benchmarks (MuseBench, AnyGroundBench) reveal gaps. The energy cost paper (H8) adds system-level perspective.