Tag
Wave Forcing introduces a wavefront-scheduling training and execution framework for streaming video generation, letting later video chunks begin denoising before earlier ones finish while overlapping communication with computation. On Wan2.1-T2V-1.3B it matches Rolling Forcing quality and reaches 147.6 DiT FPS on 8×H200, delivering 7.07× single-GPU end-to-end throughput.
ChronoReplay 是一个针对流式视觉-语言模型的空间推理训练框架，通过按时间顺序组织异构空间监督并进行族平衡训练，缓解接口锁定问题，在 OST-Bench 等基准上取得显著提升。
TorchTitan 发布重大性能与架构更新，新增对 DeepSeek-V4、Qwen3.5、Kimi K3、Muse Spark Open 等 SOTA 模型的完整 5D 并行支持，并推出统一后训练与智能体工作负载的 TitanRL 框架，FlexShard 与 torch_checkpointing 大幅提升大规模训练效率。
ProgAgent 提出一个完整的程序化工具调用（PTC）训练框架，通过可验证的长时程任务合成、轻量级 PTC 运行时和两阶段训练配方，显著优于同规模基线，在 LOCA-bench 64k token 状态下达 30.7% 通过率而基线为零。
The paper introduces SpatialASM, a framework that trains large vision-language models to actively plan task-relevant viewpoints and mentally imagine missing spatial evidence via novel-view synthesis scaffolding, improving spatial reasoning across four backbones and seven benchmarks without extra views at inference.
SALD 提出了一种自参考训练框架，通过在两个噪声水平下用同一模型评估图像-文本对来获取训练信号，无需外部教师网络或额外可训练参数，包含优势引导扩散（AGD）、时间优势记忆（TAM）和谱优势分解（SAD）三个组件，在多个架构和数据集上持续提升了生成质量。
The paper introduces Hermes, a family of configurable harnesses that let models control context allocation across inference-time compute, and Hermes-Learn, a two-stage framework that trains models to make these contextual reasoning decisions adaptively, closing the gap for smaller open-source models.
This paper proposes a rollout-based training framework (Anchor-Label Relabelling and In-Rollout Anchors) to recover full supervision for speculative decoding block drafters trained on off-policy corpora, boosting greedy accepted length by up to 36.5% over DFlash without modifying the training text.
The paper introduces GUI-SD-v2, a two-stage on-policy self-distillation framework for multi-turn GUI agents that enhances privilege following and guidance, achieving superior performance on AndroidWorld and MobileWorld benchmarks.
The paper proposes Abstract Token Curriculum (ATC), a curriculum learning framework that elicits effective continuous intermediate representations in LLMs without direct supervision, supported by theoretical and experimental evidence.
EnvCraft is an automated framework for synthesizing executable environments to address the scarcity in Agentic RL training, showing significant performance gains on claw-like and general tool-use benchmarks using Qwen models.
RAPTOR is a role-aware framework for differentially private fine-tuning of Mixture-of-Experts models, addressing failure modes like clipping interference and noise dilution, with experimental improvements over standard DP baselines.
This paper introduces TKFQA, a counterfactual benchmark of 10,130 QA pairs over tables, texts, and knowledge graphs for evaluating LLM factuality consistency and order-robust reasoning, and proposes ORLF, a training framework that improves reasoning-chain accuracy and reduces input-order sensitivity.
iFAN is a training framework that improves mask transformers for segmentation by aligning query ranking with mask quality and distilling intermediate predictions to the final layer, yielding consistent gains across benchmarks.
MiniWorld is a reproducible framework for training video world models from scratch using a block-causal Video Diffusion Transformer with Flow Matching, enabling efficient streaming generation and trainable in days on a single 8-GPU server.
Molt is a PyTorch-native training framework for agentic reinforcement learning designed to be compact and clean for easy modification, while achieving performance comparable to Megatron-based stacks.
Introduces a framework for training a harness to improve task LLM capabilities in a model-agnostic and task-environment-agnostic way, with results on Terminal Bench and SWE-Bench.
PRIME-RL is a framework for large-scale asynchronous reinforcement learning, designed to be hackable and scale to 1000+ GPUs with support for various models and environments.
Introduces Pyligent, a training framework that uses task validators to label failures and teaches LLMs to backtrack during reasoning, improving solve rates on hidden graphs, Sudoku, and Blocksworld.
DeepSeek open-sourced DeepSpec, a full-stack codebase for training and evaluating draft models for speculative decoding, enabling 60-85% faster generation. It includes data preparation, training, and evaluation scripts with support for multiple draft model algorithms (DSpark, DFlash, Eagle3).