Multi-agent debate has been shown to improve reasoning in large language models (LLMs).However, it is compute-intensive, requiring generation of long transcripts before answering questions.To address this inefficiency, we develop a framework that distills multi-agent debate into a single LLM through a two-stage fine-tuning pipeline combining debate structure learning with internalization via dynamic reward scheduling and length clipping.Across multiple models and benchmarks, our internalized models match or exceed explicit multiagent debate performance using up to 93% fewer tokens.We then investigate the mechanistic basis of this capability through activation steering, finding that internalization creates agent-specific subspaces: interpretable directions in activation space corresponding to different agent perspectives.We further demonstrate a practical application: by instilling malicious agents into the LLM through internalized debate, then applying negative steering to suppress them, we show that distillation makes harmful behaviors easier to localize and control with smaller reductions in general performance compared to steering base models.Our findings offer a new perspective for understanding multi-agent capabilities in distilled models and provide practical guidelines for controlling internalized reasoning behaviors.1

We present a highly scalable self-training framework for incrementally adapting vision-based end-to-end autonomous driving policies in a semi-supervised manner, i.e., over a continual stream of incoming video data. To facilitate large-scale model training (e.g., open web or unlabeled data), we do not assume access to ground-truth labels and instead estimate pseudo-label policy targets for each video. Our framework comprises three key components: knowledge distillation, a sample purification module, and an exploration and knowledge retention mechanism. First, given sequential image frames, we pseudo-label the data and estimate uncertainty using an ensemble of inverse dynamics models. The uncertainty is used to select the most informative samples to add to an experience replay buffer. We specifically select high-uncertainty pseudo-labels to facilitate the exploration and learning of new and diverse driving skills. However, in contrast to prior work in continual learning that assumes ground-truth labeled samples, the uncertain pseudo-labels can introduce significant noise. Thus, we also pair the exploration with a label refinement module, which makes use of consistency constraints to re-label the noisy exploratory samples and effectively learn from diverse data. Trained as a complete never-ending learning system, we demonstrate state-of-the-art performance on training from domain-changing data as well as millions of images from the open web.

With the increasing popularity of masked autoencoders, self-supervised learning (SSL) in vision undertakes a similar trajectory as in NLP. Specifically, generative pretext tasks with the masked prediction have become a de facto standard SSL practice in NLP (e.g., BERT). By contrast, early attempts at generative methods in vision have been outperformed by their discriminative counterparts (like contrastive learning). However, the success of masked image modeling has revived the autoencoder-based visual pretraining method. As a milestone to bridge the gap with BERT in NLP, masked autoencoder in vision has attracted unprecedented attention. This work conducts a survey on masked autoencoders for visual SSL.