Research Papers Kimi K3 is an open Mixture-of-Experts model with 2.8T parameters (104B activated per token), native vision input and a 1-million-token context window, built on Kimi Delta Attention and post-trained with reinforcement learning across general, agentic and coding tasks.
Symbolic music models plan composition explicitly but never sound like a finished recording, while audio models produce full songs that keep the composition hidden. YuE2 combines both in a single AR-NAR Mixture-of-Transformers that plans symbolically, expands the plan into semantic music tokens, and renders the full song, using two new representation models (MERT2, SheetSage2) to learn from recordings that lack aligned scores. The authors report a SongBench Global Avg of 6.73 on WildSongBench (6.96 at best-of-8), and in expert listening the best-of-8 samples were preferred over Suno v4.5 and roughly tied with Suno v5; the readable score also makes the output editable by external language-model agents.
VoiceMem gives duplex speech models a working memory that keeps facts and emotional context across a conversation, so a voice assistant can remember both what you said and how you said it, in real time.
Frontis-MA1 ships with OpenMLE, an open full-stack system (task gyms, RL operator learning, evolutionary search) for studying recursive self-improvement. Its 35B meta-evolution agent, post-trained around Draft, Improve, Debug and Crossover program-evolution operators, claims MLE-Bench Lite results the authors say exceed GPT-5.5 plus Codex, with components transferring to held-out NatureBench Lite.
Edge0 is an open-source inference engine that serves a 35B mixture-of-experts model from a single 24GB GPU by loading expert weights from SSD on demand; a trained per-layer prerouter predicts the next layer's expert choices one token ahead so SSD reads overlap with compute, and a recovery LoRA pays back the quality lost to int4 quantization. The authors report 20 tok/s at about 3GiB of peak active memory, within a few points of the fp16 teacher on five public benchmarks.
HexaShell is a deliberately thin robot-control harness: it only converts a coding agent's commands into motion, giving the agent three calibrated camera views, one command line, failure logs, and the ability to write reusable position-estimation tools based on camera-ray triangulation and ray-plane intersections. Trained on 10 layouts with one seed per task and then frozen, the tool library is tested on 10 unseen layouts per task across all 42 RoboDojo tasks using Codex with GPT-6-Astra at medium reasoning effort. The authors report 39.7% success and a 48.0 progress score, which they call a new state of the art on RoboDojo, with pushing at 92.5%, memory tasks at 65.0%, but tight insertion only 8.3%. Their analysis says the bottleneck is physical contact rather than reasoning: the agent often knows what to do but cannot execute the motion reliably. The tools are open-sourced.
Fine-grained perception usually means stitching together separate specialist models for detection, segmentation, counting, and OCR. WeVisOne-VL (from the WeChat Vision team) unifies object-level and pixel-level tasks in one VLM: a lightweight pixel decoder plugged in via a special <OBJ> token lets the same model output boxes, points, masks, and even alpha mattes at whatever granularity the user asks for. Trained with three-stage curriculum SFT plus mixed-task RL, the authors report leading results across 30+ benchmarks while still refusing queries where the requested object is absent, a failure mode they explicitly target for real-world deployment.
Every decision-model call today pays the full cost of a long reasoning trace even when the answer is obvious. StartLux-Decision trains a family of 0.8B–27B models with an Instant variant that answers in a single pass using bidirectional attention in full-attention layers, and an AutoThink variant that routes uncertain cases to explicit reasoning and answers confident ones directly. Its RLCD² stage freezes the model and rewards improvements in decision-probability quality with proper scoring rules, plus a soft penalty on overlong traces; the authors report this raises the share of reasoning traces that finish within the generation limit by up to 48 percentage points and cuts average reasoning length by up to 65% while largely preserving accuracy. They also release DecisionPivot, a programmatically verifiable benchmark of contrast pairs that differ in a single answer-changing detail.
OmniCompass evaluates video understanding along three complementary tasks: temporal grounding (when a target appears), spatial grounding (where it is), and question answering (what happens), in two tracks. OmniCompass-VL targets vision-language models with 820 videos and 2,460 instances, while an audio-rich subset forms OmniCompass-AV with 240 videos and 720 instances for models that must combine audio and visual evidence. Every instance is tested under paired text-only and text+image settings, and all data is manually annotated with multi-round verification. Evaluating 50 MLLMs, the authors report open-source models still trail proprietary ones, especially on audio-visual understanding, and that precise temporal and spatial grounding remains hard even for leading models.
Video understanding agents usually need slow chain-of-thought decoding, which is too late for tasks that demand timely choices from a video stream. VideoJev instead casts temporal grounding and multiple-choice video QA as structured decision tasks: it swaps the MLLM's autoregressive answer head for a shared scoring head that rates candidate time spans and text options against the video directly, trained with GRPO. A cached- and joint-inference scheme reuses visual computation across sequential queries on the same video. The authors report mean-mIoU gains of 6.7 and 10.8 points on 4B and 35B-A3B backbones with 2.8-3.1× grounding speedups, and over 100× speedups on video QA while matching Thinking-mode accuracy.
GroundingPI is a 4B foundation model that outputs points and boxes as quantized coordinates in a shared vocabulary, trained through multimodal and spatial pretraining, supervised fine-tuning, and GRPO reinforcement learning. The authors report state-of-the-art results across 34 grounding benchmarks covering 11 perceptual capabilities, averaging 73.68% against 44 baselines and edging out the larger GPT-6 Astra (71.54%). Used as the visual backbone for downstream tasks, they report it improves robotic manipulation, reaching the best results among evaluated backbones on RoboTwin 2.0 in all four out-of-distribution settings, and lowers open-loop L2 error to 0.296 m on nuScenes for autonomous driving.
Computer-use models can already read a screen and click, but games—with their visually rich worlds and diverse mechanics—remain hard, and pipelines built on large-scale demonstrations cover only a slice of dynamic game states. Player-1 post-trains such models with synthetic experience instead: it extends the computer-use action space with relative pointer motion, sustained inputs, and composite actions; fine-tunes on 6,044 curated gameplay trajectories paired with task instructions and visual reasoning; and then runs PPO in Player-1-Gym, sixteen synthetic 2D and 3D environments built around control demands shared across games, with programmatically verifiable rewards and randomized observations. The authors report that Player-1-9B and Player-1-35B-A3B beat their own SFT policies on MineDojo-300 by 27.8 and 17.7 percentage points, with Player-1-9B reaching 38.33% success versus 12.63% for the strongest baseline, plus gains on ViZDoom-10 and GameWorld—and the improvements transfer beyond the RL training suite.
Ropedia W0 is billed as the first egocentric loco-manipulation world foundation model spanning humans and robots with diverse embodiments. To handle heterogeneous action spaces, the authors encode locomotion through camera poses and manipulation through projected skeletons, with robot skeletons built from embodiment-specific forward kinematics plus manual keypoint annotation. They pair it with Xp-Data (~10k hours of egocentric human video plus 3k hours of robot data) and a new long-horizon Xp-Teleop benchmark, and report state-of-the-art results on five cross-embodiment benchmarks and held-out egocentric video in zero-shot, few-shot, and full-data settings. They also show a distilled version running real-time, minute-long rollouts for VR interaction, policy evaluation, and model-based planning.
Standard VLMs treat video as a stack of uniformly sampled frames, which is wasteful for streaming perception. Mage-VL instead uses a custom tokenizer, Mage-ViT, that encodes only entropy-rich regions guided by motion vectors and residual energy across I/P frames, reducing visual token consumption by over 75% while keeping spatiotemporal context. Trained from scratch on roughly 560M unlabeled images and 100M unlabeled video frames, Mage-VL adds a bio-inspired dual system: a lightweight System 1 event gate decides when to look, and a causal System 2 decoder handles reasoning. The authors report Mage-VL-4B matches Qwen3-VL-4B on static tasks, gains on video understanding and 2D/3D spatial reasoning, runs up to 3.5x faster in wall-clock terms, and surpasses the 15B Phi-4-reasoning-vision baseline.
Reinforcement learning on diffusion models only rewards the final image, leaving intermediate denoising steps unsupervised. This framework converts the endpoint reward into explicit bounded targets for each sampled query state, training the model on-policy through self-distillation.