multi-stage-training

VLX-Go 提出一种视觉-语言航点策略，通过多阶段多模态训练（视觉-语言对齐、多模态推理、追踪专项化）和在线策略优化，显著提升机器人在干扰物环境中跟踪目标人物的成功率，达到 58.22%，并在对比方法中取得最佳的干扰物与模糊场景追踪成绩。

Oyster-II proposes a reinforcement learning framework for constructive safety alignment in LLMs, overcoming limitations of prior SFT-based methods via a multi-stage Zero-RL paradigm, achieving state-of-the-art safety performance while preserving general capabilities.

LLMZero uses LLM agents to search over training trajectories via tree search, discovering adaptive multi-parameter transitions for RL post-training that outperform fixed schedules and grid search across diverse tasks.

This paper investigates the long-term effects of data selection strategies in multi-stage LLM fine-tuning, revealing that myopic selection can harm future adaptability. It introduces a Long-Horizon Aware Selection (LHAS) objective to mitigate these issues.

This paper proposes a multi-stage training pipeline using language-based preprocessing and an ensemble of models to detect abusive comments in Indic languages, aiming to minimize false positives while preserving freedom of expression.

SmartPhotoCrafter introduces an automatic photographic image editing pipeline that unifies quality comprehension and enhancement without explicit human instructions, outperforming existing generative models on photo-realistic enhancement tasks.