MiniMax-01: Scaling Foundation Models with Lightning Attention

We introduce MiniMax-01 series, including MiniMax-Text-01 and MiniMax-VL-01, which are comparable to top-tier models while offering superior capabilities in processing longer contexts. The core lies in lightning attention and its efficient scaling. To maximize computational capacity, we integrate it with Mixture of Experts (MoE), creating a model with 32 experts and 456 billion total parameters, of which 45.9 billion are activated for each token. We develop an optimized parallel strategy and highly efficient computation-communication overlap techniques for MoE and lightning attention. This approach enables us to conduct efficient training and inference on models with hundreds of billions of parameters across contexts spanning millions of tokens. The context window of MiniMax-Text-01 can reach up to 1 million tokens during training and extrapolate to 4 million tokens during inference at an affordable cost. Our vision-language model, MiniMax-VL-01 is built through continued training with 512 billion vision-language tokens. Experiments on both standard and in-house benchmarks show that our models match the performance of state-of-the-art models like GPT-4o and Claude-3.5-Sonnet while offering 20-32 times longer context window. We publicly release MiniMax-01 at https://github.com/MiniMax-AI.

MoVE-KD: Knowledge Distillation for VLMs with Mixture of Visual Encoders

Visual encoders are fundamental components in vision-language models (VLMs), each showcasing unique strengths derived from various pre-trained visual foundation models. To leverage the various capabilities of these encoders, recent studies incorporate multiple encoders within a single VLM, leading to a considerable increase in computational cost. In this paper, we present Mixture-of-Visual-Encoder Knowledge Distillation (MoVE-KD), a novel framework that distills the unique proficiencies of multiple vision encoders into a single, efficient encoder model. Specifically, to mitigate conflicts and retain the unique characteristics of each teacher encoder, we employ low-rank adaptation (LoRA) and mixture-of-experts (MoEs) to selectively activate specialized knowledge based on input features, enhancing both adaptability and efficiency. To regularize the KD process and enhance performance, we propose an attention-based distillation strategy that adaptively weighs the different visual encoders and emphasizes valuable visual tokens, reducing the burden of replicating comprehensive but distinct features from multiple teachers. Comprehensive experiments on popular VLMs, such as LLaVA and LLaVA-NeXT, validate the effectiveness of our method. The code will be released.

Cross-Self KV Cache Pruning for Efficient Vision-Language Inference

KV cache pruning has emerged as a promising technique for reducing memory and computation costs in long-context auto-regressive generation. Existing methods for vision-language models (VLMs) typically rely on self-attention scores from large language models (LLMs) to identify and prune irrelevant tokens. However, these approaches overlook the inherent distributional discrepancies between modalities, often leading to inaccurate token importance estimation and the over-pruning of critical visual tokens. To address this, we propose decomposing attention scores into intra-modality attention (within the same modality) and inter-modality attention (across modalities), enabling more precise KV cache pruning by independently managing these distinct attention types. Additionally, we introduce an n-softmax function to counteract distribution shifts caused by pruning, preserving the original smoothness of attention scores and ensuring stable performance. Our final training-free method, \textbf{C}ross-\textbf{S}elf \textbf{P}runing (CSP), achieves competitive performance compared to models with full KV caches while significantly outperforming previous pruning methods. Extensive evaluations on MileBench, a benchmark encompassing 29 multimodal datasets, demonstrate CSP's effectiveness, achieving up to a 41\% performance improvement on challenging tasks like conversational embodied dialogue while reducing the KV cache budget by 13.6\%. The code is available at https://github.com/TerryPei/CSP

Intermediate Outputs Are More Sensitive Than You Think

The increasing reliance on deep computer vision models that process sensitive data has raised significant privacy concerns, particularly regarding the exposure of intermediate results in hidden layers. While traditional privacy risk assessment techniques focus on protecting overall model outputs, they often overlook vulnerabilities within these intermediate representations. Current privacy risk assessment techniques typically rely on specific attack simulations to assess risk, which can be computationally expensive and incomplete. This paper introduces a novel approach to measuring privacy risks in deep computer vision models based on the Degrees of Freedom (DoF) and sensitivity of intermediate outputs, without requiring adversarial attack simulations. We propose a framework that leverages DoF to evaluate the amount of information retained in each layer and combines this with the rank of the Jacobian matrix to assess sensitivity to input variations. This dual analysis enables systematic measurement of privacy risks at various model layers. Our experimental validation on real-world datasets demonstrates the effectiveness of this approach in providing deeper insights into privacy risks associated with intermediate representations.

Learning Humanoid Locomotion with Perceptive Internal Model

In contrast to quadruped robots that can navigate diverse terrains using a "blind" policy, humanoid robots require accurate perception for stable locomotion due to their high degrees of freedom and inherently unstable morphology. However, incorporating perceptual signals often introduces additional disturbances to the system, potentially reducing its robustness, generalizability, and efficiency. This paper presents the Perceptive Internal Model (PIM), which relies on onboard, continuously updated elevation maps centered around the robot to perceive its surroundings. We train the policy using ground-truth obstacle heights surrounding the robot in simulation, optimizing it based on the Hybrid Internal Model (HIM), and perform inference with heights sampled from the constructed elevation map. Unlike previous methods that directly encode depth maps or raw point clouds, our approach allows the robot to perceive the terrain beneath its feet clearly and is less affected by camera movement or noise. Furthermore, since depth map rendering is not required in simulation, our method introduces minimal additional computational costs and can train the policy in 3 hours on an RTX 4090 GPU. We verify the effectiveness of our method across various humanoid robots, various indoor and outdoor terrains, stairs, and various sensor configurations. Our method can enable a humanoid robot to continuously climb stairs and has the potential to serve as a foundational algorithm for the development of future humanoid control methods.

Enhancing DP-SGD through Non-monotonous Adaptive Scaling Gradient Weight

In the domain of deep learning, the challenge of protecting sensitive data while maintaining model utility is significant. Traditional Differential Privacy (DP) techniques such as Differentially Private Stochastic Gradient Descent (DP-SGD) typically employ strategies like direct or per-sample adaptive gradient clipping. These methods, however, compromise model accuracy due to their critical influence on gradient handling, particularly neglecting the significant contribution of small gradients during later training stages. In this paper, we introduce an enhanced version of DP-SGD, named Differentially Private Per-sample Adaptive Scaling Clipping (DP-PSASC). This approach replaces traditional clipping with non-monotonous adaptive gradient scaling, which alleviates the need for intensive threshold setting and rectifies the disproportionate weighting of smaller gradients. Our contribution is twofold. First, we develop a novel gradient scaling technique that effectively assigns proper weights to gradients, particularly small ones, thus improving learning under differential privacy. Second, we integrate a momentum-based method into DP-PSASC to reduce bias from stochastic sampling, enhancing convergence rates. Our theoretical and empirical analyses confirm that DP-PSASC preserves privacy and delivers superior performance across diverse datasets, setting new standards for privacy-sensitive applications.

Gradient-Guided Conditional Diffusion Models for Private Image Reconstruction: Analyzing Adversarial Impacts of Differential Privacy and Denoising

We investigate the construction of gradient-guided conditional diffusion models for reconstructing private images, focusing on the adversarial interplay between differential privacy noise and the denoising capabilities of diffusion models. While current gradient-based reconstruction methods struggle with high-resolution images due to computational complexity and prior knowledge requirements, we propose two novel methods that require minimal modifications to the diffusion model's generation process and eliminate the need for prior knowledge. Our approach leverages the strong image generation capabilities of diffusion models to reconstruct private images starting from randomly generated noise, even when a small amount of differentially private noise has been added to the gradients. We also conduct a comprehensive theoretical analysis of the impact of differential privacy noise on the quality of reconstructed images, revealing the relationship among noise magnitude, the architecture of attacked models, and the attacker's reconstruction capability. Additionally, extensive experiments validate the effectiveness of our proposed methods and the accuracy of our theoretical findings, suggesting new directions for privacy risk auditing using conditional diffusion models.

Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Dataset

The pre-training of visual representations has enhanced the efficiency of robot learning. Due to the lack of large-scale in-domain robotic datasets, prior works utilize in-the-wild human videos to pre-train robotic visual representation. Despite their promising results, representations from human videos are inevitably subject to distribution shifts and lack the dynamics information crucial for task completion. We first evaluate various pre-trained representations in terms of their correlation to the downstream robotic manipulation tasks (i.e., manipulation centricity). Interestingly, we find that the "manipulation centricity" is a strong indicator of success rates when applied to downstream tasks. Drawing from these findings, we propose Manipulation Centric Representation (MCR), a foundation representation learning framework capturing both visual features and the dynamics information such as actions and proprioceptions of manipulation tasks to improve manipulation centricity. Specifically, we pre-train a visual encoder on the DROID robotic dataset and leverage motion-relevant data such as robot proprioceptive states and actions. We introduce a novel contrastive loss that aligns visual observations with the robot's proprioceptive state-action dynamics, combined with a behavior cloning (BC)-like actor loss to predict actions during pre-training, along with a time contrastive loss. Empirical results across 4 simulation domains with 20 tasks verify that MCR outperforms the strongest baseline method by 14.8%. Moreover, MCR boosts the performance of data-efficient learning with a UR5e arm on 3 real-world tasks by 76.9%. Project website: https://robots-pretrain-robots.github.io/.

SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference

In vision-language models (VLMs), visual tokens usually consume a significant amount of computational overhead, despite their sparser information density compared to text tokens. To address this, most existing methods learn a network to prune redundant visual tokens and require additional training data. Differently, we propose an efficient training-free token optimization mechanism dubbed SparseVLM without extra parameters or fine-tuning costs. Concretely, given that visual tokens complement text tokens in VLMs for linguistic reasoning, we select visual-relevant text tokens to rate the significance of vision tokens within the self-attention matrix extracted from the VLMs. Then we progressively prune irrelevant tokens. To maximize sparsity while retaining essential information, we introduce a rank-based strategy to adaptively determine the sparsification ratio for each layer, alongside a token recycling method that compresses pruned tokens into more compact representations. Experimental results show that our SparseVLM improves the efficiency of various VLMs across a range of image and video understanding tasks. In particular, LLaVA equipped with SparseVLM reduces 61% to 67% FLOPs with a compression ratio of 78% while maintaining 93% of the accuracy. Our code is available at https://github.com/Gumpest/SparseVLMs.

Multi-robot Task Allocation and Path Planning with Maximum Range Constraints

This letter presents a novel multi-robot task allocation and path planning method that considers robots' maximum range constraints in large-sized workspaces, enabling robots to complete the assigned tasks within their range limits. Firstly, we developed a fast path planner to solve global paths efficiently. Subsequently, we propose an innovative auction-based approach that integrates our path planner into the auction phase for reward computation while considering the robots' range limits. This method accounts for extra obstacle-avoiding travel distances rather than ideal straight-line distances, resolving the coupling between task allocation and path planning. Additionally, to avoid redundant computations during iterations, we implemented a lazy auction strategy to speed up the convergence of the task allocation. Finally, we validated the proposed method's effectiveness and application potential through extensive simulation and real-world experiments. The implementation code for our method will be available at https://github.com/wuuya1/RangeTAP.