Every FLOP Counts: Scaling a 300B Mixture-of-Experts LING LLM without Premium GPUs

In this technical report, we tackle the challenges of training large-scale Mixture of Experts (MoE) models, focusing on overcoming cost inefficiency and resource limitations prevalent in such systems. To address these issues, we present two differently sized MoE large language models (LLMs), namely Ling-Lite and Ling-Plus (referred to as "Bailing" in Chinese, spelled B\v{a}il\'ing in Pinyin). Ling-Lite contains 16.8 billion parameters with 2.75 billion activated parameters, while Ling-Plus boasts 290 billion parameters with 28.8 billion activated parameters. Both models exhibit comparable performance to leading industry benchmarks. This report offers actionable insights to improve the efficiency and accessibility of AI development in resource-constrained settings, promoting more scalable and sustainable technologies. Specifically, to reduce training costs for large-scale MoE models, we propose innovative methods for (1) optimization of model architecture and training processes, (2) refinement of training anomaly handling, and (3) enhancement of model evaluation efficiency. Additionally, leveraging high-quality data generated from knowledge graphs, our models demonstrate superior capabilities in tool use compared to other models. Ultimately, our experimental findings demonstrate that a 300B MoE LLM can be effectively trained on lower-performance devices while achieving comparable performance to models of a similar scale, including dense and MoE models. Compared to high-performance devices, utilizing a lower-specification hardware system during the pre-training phase demonstrates significant cost savings, reducing computing costs by approximately 20%. The models can be accessed at https://huggingface.co/inclusionAI.

Practical Continual Forgetting for Pre-trained Vision Models

For privacy and security concerns, the need to erase unwanted information from pre-trained vision models is becoming evident nowadays. In real-world scenarios, erasure requests originate at any time from both users and model owners, and these requests usually form a sequence. Therefore, under such a setting, selective information is expected to be continuously removed from a pre-trained model while maintaining the rest. We define this problem as continual forgetting and identify three key challenges. (i) For unwanted knowledge, efficient and effective deleting is crucial. (ii) For remaining knowledge, the impact brought by the forgetting procedure should be minimal. (iii) In real-world scenarios, the training samples may be scarce or partially missing during the process of forgetting. To address them, we first propose Group Sparse LoRA (GS-LoRA). Specifically, towards (i), we introduce LoRA modules to fine-tune the FFN layers in Transformer blocks for each forgetting task independently, and towards (ii), a simple group sparse regularization is adopted, enabling automatic selection of specific LoRA groups and zeroing out the others. To further extend GS-LoRA to more practical scenarios, we incorporate prototype information as additional supervision and introduce a more practical approach, GS-LoRA++. For each forgotten class, we move the logits away from its original prototype. For the remaining classes, we pull the logits closer to their respective prototypes. We conduct extensive experiments on face recognition, object detection and image classification and demonstrate that our method manages to forget specific classes with minimal impact on other classes. Codes have been released on https://github.com/bjzhb666/GS-LoRA.

PO-GVINS: Tightly Coupled GNSS-Visual-Inertial Integration with Pose-Only Representation

Accurate and reliable positioning is crucial for perception, decision-making, and other high-level applications in autonomous driving, unmanned aerial vehicles, and intelligent robots. Given the inherent limitations of standalone sensors, integrating heterogeneous sensors with complementary capabilities is one of the most effective approaches to achieving this goal. In this paper, we propose a filtering-based, tightly coupled global navigation satellite system (GNSS)-visual-inertial positioning framework with a pose-only formulation applied to the visual-inertial system (VINS), termed PO-GVINS. Specifically, multiple-view imaging used in current VINS requires a priori of 3D feature, then jointly estimate camera poses and 3D feature position, which inevitably introduces linearization error of the feature as well as facing dimensional explosion. However, the pose-only (PO) formulation, which is demonstrated to be equivalent to the multiple-view imaging and has been applied in visual reconstruction, represent feature depth using two camera poses and thus 3D feature position is removed from state vector avoiding aforementioned difficulties. Inspired by this, we first apply PO formulation in our VINS, i.e., PO-VINS. GNSS raw measurements are then incorporated with integer ambiguity resolved to achieve accurate and drift-free estimation. Extensive experiments demonstrate that the proposed PO-VINS significantly outperforms the multi-state constrained Kalman filter (MSCKF). By incorporating GNSS measurements, PO-GVINS achieves accurate, drift-free state estimation, making it a robust solution for positioning in challenging environments.

RoboHanger: Learning Generalizable Robotic Hanger Insertion for Diverse Garments

For the task of hanging clothes, learning how to insert a hanger into a garment is crucial but has been seldom explored in robotics. In this work, we address the problem of inserting a hanger into various unseen garments that are initially laid out flat on a table. This task is challenging due to its long-horizon nature, the high degrees of freedom of the garments, and the lack of data. To simplify the learning process, we first propose breaking the task into several stages. Then, we formulate each stage as a policy learning problem and propose low-dimensional action parameterization. To overcome the challenge of limited data, we build our own simulator and create 144 synthetic clothing assets to effectively collect high-quality training data. Our approach uses single-view depth images and object masks as input, which mitigates the Sim2Real appearance gap and achieves high generalization capabilities for new garments. Extensive experiments in both simulation and the real world validate our proposed method. By training on various garments in the simulator, our method achieves a 75\% success rate with 8 different unseen garments in the real world.

ScissorBot: Learning Generalizable Scissor Skill for Paper Cutting via Simulation, Imitation, and Sim2Real

This paper tackles the challenging robotic task of generalizable paper cutting using scissors. In this task, scissors attached to a robot arm are driven to accurately cut curves drawn on the paper, which is hung with the top edge fixed. Due to the frequent paper-scissor contact and consequent fracture, the paper features continual deformation and changing topology, which is diffult for accurate modeling. To ensure effective execution, we customize an action primitive sequence for imitation learning to constrain its action space, thus alleviating potential compounding errors. Finally, by integrating sim-to-real techniques to bridge the gap between simulation and reality, our policy can be effectively deployed on the real robot. Experimental results demonstrate that our method surpasses all baselines in both simulation and real-world benchmarks and achieves performance comparable to human operation with a single hand under the same conditions.

Towards Fast Rates for Federated and Multi-Task Reinforcement Learning

We consider a setting involving $N$ agents, where each agent interacts with an environment modeled as a Markov Decision Process (MDP). The agents' MDPs differ in their reward functions, capturing heterogeneous objectives/tasks. The collective goal of the agents is to communicate intermittently via a central server to find a policy that maximizes the average of long-term cumulative rewards across environments. The limited existing work on this topic either only provide asymptotic rates, or generate biased policies, or fail to establish any benefits of collaboration. In response, we propose Fast-FedPG - a novel federated policy gradient algorithm with a carefully designed bias-correction mechanism. Under a gradient-domination condition, we prove that our algorithm guarantees (i) fast linear convergence with exact gradients, and (ii) sub-linear rates that enjoy a linear speedup w.r.t. the number of agents with noisy, truncated policy gradients. Notably, in each case, the convergence is to a globally optimal policy with no heterogeneity-induced bias. In the absence of gradient-domination, we establish convergence to a first-order stationary point at a rate that continues to benefit from collaboration.

Text-Augmented Multimodal LLMs for Chemical Reaction Condition Recommendation

High-throughput reaction condition (RC) screening is fundamental to chemical synthesis. However, current RC screening suffers from laborious and costly trial-and-error workflows. Traditional computer-aided synthesis planning (CASP) tools fail to find suitable RCs due to data sparsity and inadequate reaction representations. Nowadays, large language models (LLMs) are capable of tackling chemistry-related problems, such as molecule design, and chemical logic Q\&A tasks. However, LLMs have not yet achieved accurate predictions of chemical reaction conditions. Here, we present MM-RCR, a text-augmented multimodal LLM that learns a unified reaction representation from SMILES, reaction graphs, and textual corpus for chemical reaction recommendation (RCR). To train MM-RCR, we construct 1.2 million pair-wised Q\&A instruction datasets. Our experimental results demonstrate that MM-RCR achieves state-of-the-art performance on two open benchmark datasets and exhibits strong generalization capabilities on out-of-domain (OOD) and High-Throughput Experimentation (HTE) datasets. MM-RCR has the potential to accelerate high-throughput condition screening in chemical synthesis.

Instruct-ReID++: Towards Universal Purpose Instruction-Guided Person Re-identification

Human intelligence can retrieve any person according to both visual and language descriptions. However, the current computer vision community studies specific person re-identification (ReID) tasks in different scenarios separately, which limits the applications in the real world. This paper strives to resolve this problem by proposing a novel instruct-ReID task that requires the model to retrieve images according to the given image or language instructions. Instruct-ReID is the first exploration of a general ReID setting, where existing 6 ReID tasks can be viewed as special cases by assigning different instructions. To facilitate research in this new instruct-ReID task, we propose a large-scale OmniReID++ benchmark equipped with diverse data and comprehensive evaluation methods e.g., task specific and task-free evaluation settings. In the task-specific evaluation setting, gallery sets are categorized according to specific ReID tasks. We propose a novel baseline model, IRM, with an adaptive triplet loss to handle various retrieval tasks within a unified framework. For task-free evaluation setting, where target person images are retrieved from task-agnostic gallery sets, we further propose a new method called IRM++ with novel memory bank-assisted learning. Extensive evaluations of IRM and IRM++ on OmniReID++ benchmark demonstrate the superiority of our proposed methods, achieving state-of-the-art performance on 10 test sets. The datasets, the model, and the code will be available at https://github.com/hwz-zju/Instruct-ReID

Causal Deconfounding via Confounder Disentanglement for Dual-Target Cross-Domain Recommendation

In recent years, dual-target Cross-Domain Recommendation (CDR) has been proposed to capture comprehensive user preferences in order to ultimately enhance the recommendation accuracy in both data-richer and data-sparser domains simultaneously. However, in addition to users' true preferences, the user-item interactions might also be affected by confounders (e.g., free shipping, sales promotion). As a result, dual-target CDR has to meet two challenges: (1) how to effectively decouple observed confounders, including single-domain confounders and cross-domain confounders, and (2) how to preserve the positive effects of observed confounders on predicted interactions, while eliminating their negative effects on capturing comprehensive user preferences. To address the above two challenges, we propose a Causal Deconfounding framework via Confounder Disentanglement for dual-target Cross-Domain Recommendation, called CD2CDR. In CD2CDR, we first propose a confounder disentanglement module to effectively decouple observed single-domain and cross-domain confounders. We then propose a causal deconfounding module to preserve the positive effects of such observed confounders and eliminate their negative effects via backdoor adjustment, thereby enhancing the recommendation accuracy in each domain. Extensive experiments conducted on five real-world datasets demonstrate that CD2CDR significantly outperforms the state-of-the-art methods.

3D Geometry-aware Deformable Gaussian Splatting for Dynamic View Synthesis

In this paper, we propose a 3D geometry-aware deformable Gaussian Splatting method for dynamic view synthesis. Existing neural radiance fields (NeRF) based solutions learn the deformation in an implicit manner, which cannot incorporate 3D scene geometry. Therefore, the learned deformation is not necessarily geometrically coherent, which results in unsatisfactory dynamic view synthesis and 3D dynamic reconstruction. Recently, 3D Gaussian Splatting provides a new representation of the 3D scene, building upon which the 3D geometry could be exploited in learning the complex 3D deformation. Specifically, the scenes are represented as a collection of 3D Gaussian, where each 3D Gaussian is optimized to move and rotate over time to model the deformation. To enforce the 3D scene geometry constraint during deformation, we explicitly extract 3D geometry features and integrate them in learning the 3D deformation. In this way, our solution achieves 3D geometry-aware deformation modeling, which enables improved dynamic view synthesis and 3D dynamic reconstruction. Extensive experimental results on both synthetic and real datasets prove the superiority of our solution, which achieves new state-of-the-art performance. The project is available at https://npucvr.github.io/GaGS/