AI Audio2026
Generates and edits speech from natural-language instructions plus optional reference audio, supporting zero-shot TTS, content/acoustic/paralinguistic edits, enhancement, and source separation. Open-source 1.5B-parameter base model with a 4-step distilled AuK‑Flash for faster inference.
AI Dataset2026
10,005 single-turn prompts and metadata derived from Ox Alpha chat traces, labeled by category, subcategory, and difficulty. Built for response-teacher generation and distillation workflows, with topic and difficulty distributions to aid dataset curation.
Converts image-level rewards into explicit intermediate targets for diffusion-model denoising via an on-policy self-distillation loop. Constructs bounded positive/negative targets around anchors from reward gradients, fits those targets with finite updates, and refreshes a behavior policy by EMA—improving aligned performance across backbones while reducing GPU hours.
Analyzes on-policy self-distillation for language-model reasoning, diagnosing “collapse” and framing it as controlled by three levers: where token-level signals apply, what privileged information the teacher sees, and how teacher dynamics evolve.
AI Video2026
Provides Parallel Decoding Distillation (PDD) LoRA adapters that accelerate MiniMax-H3 video generation into few inference steps. Includes official 8-step Acc LoRAs for FL2VA and Ref2VA (rank=64, network_alpha=64, BF16), demo comparison videos, and example scripts using Diffusers' MiniMax-H3 ModularPipeline.
AI Dataset2026
Contains 244 Cursor agent sessions recorded from Claude Fable‑5, formatted for training and research. Sessions include multi-turn assistant/tool interactions and are Teich-compatible; several rows exceed one million characters, so apply explicit tokenization and oversize policies before training.
AI Agent Papers2026
Synthesizes, repairs, and self-evolves task-adaptive agent harnesses on demand for off-the-shelf LLM agents, using a trainable harness-intelligence model that distills signals from past configurations. Demonstrates consistent performance gains across benchmarks and model families by producing four-module, composable harnesses.
Aozhe Wang, Zhengxi Lu +9Affiliation: Zhejiang University, Affiliation: Alibaba Group{waz,zhengxilu,syl}@zju.edu.cn [email protected]
A test-time method that adapts LLMs without labels by distilling rollouts that agree with majority pseudo-labels and penalizing disagreeing rollouts via grouped RL, improving robustness under frequent pseudo-label errors.
Turns a flow-matching image generator's self-exploration into dense, per-step supervision without a pretrained teacher; it branches the student's next-state into stochastic SDE candidates, scores them against a deterministic self-reference, and applies an advantage-weighted pull–push velocity regression with reward-level fusion for multi-objective alignment.
AI Video Papers2026
Generates synchronized native 2K audio-video from a single first frame and a text prompt using a compact 7B joint generator. Combines gated cross-modal attention, progressive joint training, audio-video reinforcement learning, and an Autoregressive 1-Step 2K Refinement; releases a 7B generator and 2K Refiner for research use.
AI Video2026
Replaces a character in a video using a single repainted frame from the same clip and propagates that edit across the shot while preserving motion, camera and lighting; requires no pose estimator, segmentation, face tracker or text prompt. Key facts: a 33.1B MiniMax-H3 finetune, DMD-distilled to three forward passes, 124 frames in ~26s on one B200 GPU.
Analyzes on-policy distillation for LLM fine-tuning, shows teacher token-level supervision is often noisy and not the main driver of gains, and introduces OPSA, a supervision-free, entropy-adaptive method that suppresses low-probability tokens to improve downstream accuracy.