MiMo: Unlocking the Reasoning Potential of Language Model -- From Pretraining to Posttraining

We present MiMo-7B, a large language model born for reasoning tasks, with optimization across both pre-training and post-training stages. During pre-training, we enhance the data preprocessing pipeline and employ a three-stage data mixing strategy to strengthen the base model's reasoning potential. MiMo-7B-Base is pre-trained on 25 trillion tokens, with additional Multi-Token Prediction objective for enhanced performance and accelerated inference speed. During post-training, we curate a dataset of 130K verifiable mathematics and programming problems for reinforcement learning, integrating a test-difficulty-driven code-reward scheme to alleviate sparse-reward issues and employing strategic data resampling to stabilize training. Extensive evaluations show that MiMo-7B-Base possesses exceptional reasoning potential, outperforming even much larger 32B models. The final RL-tuned model, MiMo-7B-RL, achieves superior performance on mathematics, code and general reasoning tasks, surpassing the performance of OpenAI o1-mini. The model checkpoints are available at https://github.com/xiaomimimo/MiMo.

Bidirectional Prototype-Reward co-Evolution for Test-Time Adaptation of Vision-Language Models

Test-time adaptation (TTA) is crucial in maintaining Vision-Language Models (VLMs) performance when facing real-world distribution shifts, particularly when the source data or target labels are inaccessible. Existing TTA methods rely on CLIP's output probability distribution for feature evaluation, which can introduce biases under domain shifts. This misalignment may cause features to be misclassified due to text priors or incorrect textual associations. To address these limitations, we propose Bidirectional Prototype-Reward co-Evolution (BPRE), a novel TTA framework for VLMs that integrates feature quality assessment with prototype evolution through a synergistic feedback loop. BPRE first employs a Multi-Dimensional Quality-Aware Reward Module to evaluate feature quality and guide prototype refinement precisely. The continuous refinement of prototype quality through Prototype-Reward Interactive Evolution will subsequently enhance the computation of more robust Multi-Dimensional Quality-Aware Reward Scores. Through the bidirectional interaction, the precision of rewards and the evolution of prototypes mutually reinforce each other, forming a self-evolving cycle. Extensive experiments are conducted across 15 diverse recognition datasets encompassing natural distribution shifts and cross-dataset generalization scenarios. Results demonstrate that BPRE consistently achieves superior average performance compared to state-of-the-art methods across different model architectures, such as ResNet-50 and ViT-B/16. By emphasizing comprehensive feature evaluation and bidirectional knowledge refinement, BPRE advances VLM generalization capabilities, offering a new perspective on TTA.

Prediction Is All MoE Needs: Expert Load Distribution Goes from Fluctuating to Stabilizing

MoE facilitates the development of large models by making the computational complexity of the model no longer scale linearly with increasing parameters. The learning sparse gating network selects a set of experts for each token to be processed; however, this may lead to differences in the number of tokens processed by each expert over several successive iterations, i.e., the expert load fluctuations, which reduces computational parallelization and resource utilization. To this end, we traced and analyzed loads of each expert in the training iterations for several large language models in this work, and defined the transient state with "obvious load fluctuation" and the stable state with "temporal locality". Moreover, given the characteristics of these two states and the computational overhead, we deployed three classical prediction algorithms that achieve accurate expert load prediction results. For the GPT3 350M model, the average error rates for predicting the expert load proportion over the next 1,000 and 2,000 steps are approximately 1.3% and 1.8%, respectively. This work can provide valuable guidance for expert placement or resource allocation for MoE model training. Based on this work, we will propose an expert placement scheme for transient and stable states in our coming work.