Structured, downloadable JSONL dataset of Seedance 2.0 video-generation prompts with matching MP4 previews and cover images; includes English/Chinese texts and standardized metadata (duration, resolution, safety) and is released under CC BY 4.0 for reuse.
Delivers image and video generation, editing, and understanding inside a single 3B-parameter multimodal model trained from scratch with a multi-task recipe. Notable for strong unified benchmarks at 3B scale; inference requires large GPU memory (≈40GB+ VRAM).
Provides the renderer weights and inference code for Bernini’s video renderer, enabling text→video, image→video and video editing inference. Offers a ready diffusers-format bundle or safetensors checkpoints under Apache‑2.0; intended for multi‑GPU/Hopper inference and reproducible research.
Measures how autonomous AI agents learn via long-horizon, feedback-rich executable tasks; publishes 51 public tasks from a 134-task suite and provides SForge, a two-container evaluation harness for iterative 12+ hour runs to track learning trajectories.
Treats hybrid layer selection as a budget-constrained subset optimization and introduces FlashMorph: a pipeline that equips each transformer layer with a linear-attention branch, jointly optimizes layerwise gates on synthetic long-context retrieval data, then discretizes, distills, and finetunes—achieving strong long-context recall using only 20M selection tokens.
Converts posed indoor RGB(-D) video into editable, simulation-ready 3D scene graphs by parsing multi-view evidence into per-object bundles, generating complete object assets from that evidence, and placing them with GizmoAct, a VLM policy that refines 9-DoF poses through closed-loop GUI actions.
Defines recursive self-improvement (RSI) with an evaluation metric (Headroom-Closed Index) and a five-level autonomy roadmap—from executing human-specified improvements to recursively improving improvement mechanisms—while analyzing scenarios, verification challenges, and governance risks.
Demonstrates that additive weighted aggregation in rubric-based reinforcement learning enables reward hacking—models improve rubric coverage but produce worse answers—and proposes ProRubric, an offline regrouping that aggregates checklist criteria into conjunctive protocol-level dimensions to raise appropriateness without losing coverage.
Predicts identity-preserving dense pixel correspondences between image pairs that violate spatio-temporal priors (e.g., edits and reference-guided generation). Fuses generative (FLUX2) and semantic (DINOv3) foundation representations with heterogeneous supervision and teacher-guided iterative refinement to generalize beyond classical optical-flow assumptions.
Trains agent models with two supervised fine-tuning stages followed by reinforcement learning, and releases the data-generation, verification, and training pipeline. The recipe skips dedicated mid-training, DPO, and on-policy distillation.
Provides scientific-reasoning and tool-use trajectories for training agent foundation models, alongside reinforcement-learning tasks. Its staged data makes the Axion 1.0 training process easier to inspect and reproduce.
Turns LLMs’ physical knowledge into video training data using simulator constraints, physical verification, and iterative correction. It also repairs physical errors in generated videos and releases 100K videos spanning 2,500 physical events.