#1 ·
Confucius4-R2T2: A Stable-Prefix Streaming ASR Model for Real Real-Time Transcription
Daniel Huang, Qi He, Shaoxiong Duan, Yingjie Li
Speech RecognitionStreaming ASRPost-Training
Why it matters
Voice agents, live translators, and any machine consuming streaming transcripts act on tokens as they appear, and every revision downstream is wasted work or a wrong action. A streaming ASR model that commits to text permanently removes an entire failure class from those pipelines. This is the rare case where the guarantee (no rewrites) comes with near-full-context accuracy rather than a quality tax.
The authors post-train Qwen3-ASR-1.7B on 1,500 hours of Chinese and English audio, freezing the pretrained audio encoder and projector, to produce a streaming model that never modifies earlier output. The key tool is a new label, the Longest Stable Prefix: for each prefix of the audio it is the longest label that equals a prefix of the full-audio transcription. They prove it is monotone under the prefix order and consistent with the full-context output, so learning from LSP labels converts a full-context model into a stable-prefix one with no architectural change. Evaluations show streaming accuracy close to the base model's full-context quality with low emission lag, and gains that extend to languages outside the post-training set. Weights and code are open-sourced.
arXiv | HuggingFace
#2 ·
SplatAnything: Gaussian Spacing Normalization for Feed-forward 3DGS across Arbitrary Scenes
Hao Yu, Zhenwei Wang, Yongzhen Hu, Yuning Gong
3D Gaussian SplattingFeed-Forward Reconstruction3D Vision
Why it matters
Feed-forward 3D Gaussian splatting is the path to instant scene reconstruction without per-scene optimization, but existing methods are calibrated around radial distance and fall apart when scene scale or camera configuration changes. For robotics, AR, and any workflow that ingests arbitrary captures, scale robustness is the difference between a demo and a usable system. The authors also ship the systematic benchmark the field lacked.
The paper identifies a normalization mismatch in feed-forward 3DGS: Gaussian attributes are predicted in a radial-distance-normalized space, yet the actual spacing between neighboring Gaussian centers varies substantially across scene scales and camera setups, shifting the Gaussian-scale supervision targets. SplatAnything introduces Gaussian Spacing Normalization, canonicalizing a robust scene-level statistic of neighbor spacing to give Gaussian prediction a consistent spatial reference. The authors build a benchmark spanning datasets, input sparsity levels, and extrapolated viewpoints, and report state-of-the-art results with the largest gains under large variations in scene scale and viewpoint.
arXiv | HuggingFace
#3 ·
P^2-MOPD: Accelerating Multi-Teacher On-Policy Distillation via Parameter Priors
Sinuo Wang, Jiacheng Lu, Zeyu Zhang, Maojia Song
DistillationModel MergingPost-Training
Why it matters
Multi-teacher on-policy distillation is a practical way to consolidate several expert models, but it is training-heavy; parameter merging like task arithmetic is cheap but permanently confined to the span of the frozen parameters. If you operate a fleet of specialized models, this result means the consolidation step gets both faster and stronger, not a trade-off between the two.
The authors treat the linear combination of student and expert parameters as a single differentiable model and jointly optimize the mixing coefficients and the student weights under the standard on-policy distillation objective. Because the student stays trainable, the experts act as parameter priors rather than confining the model to the span of frozen student and expert parameters. Across four domains at 1.7B and 4B scales, P2-MOPD surpasses the final average recovery of MOPD from the same initialization in half the training steps, and the 1.7B model ends at 94.7% average recovery versus 80.9% for MOPD.
arXiv | HuggingFace
Outlook
The through-line is commitment: systems that lock in early decisions rather than revising them, whether that is a transcribed token, a Gaussian scale target, or a student model's weights. Expect follow-ups on stable-prefix decoding for other autoregressive pipelines, and on parameter-prior distillation scaling past the 4B regime tested here.