Pengbo Li

Pengbo Li, Feiyuan Zhang, Guangming Sheng et al.

GRPO- and PPO-style LLM post-training commonly sample multiple trajectories from the same prompt and then train on the resulting group. In long-context RL workloads, this shared prompt-side prefix can contain retrieved passages, visual tokens, tool schemas, system instructions, or task context, while the full rollout group is still too large to pack into one training microbatch. Standard dense trainers therefore recompute the same prefix forward and backward for every trajectory. We present a schedule-level reuse mechanism that decouples prefix and suffix computation. The schedule runs prefix forward once, executes suffixes as ordinary microbatches while reading prefix K/V and accumulating prefix-side gK/gV , and then runs prefix backward once on the accumulated gradient cache. This reordered schedule is equivalent to baseline training over real arithmetic and aligns numerically within finite-precision tolerance. Because only K/V and gK/gV are hot during suffix computation, the approach offloads dormant prefix activations, integrates with TP/EP/CP/PP and DP-style placement at the execution level, and preserves aux-loss-based MoE router semantics through logical prefix-token accounting. On dense Llama3-8B, Qwen3-8B, and MoE Qwen3-MoE-30B-A3B configurations, the schedule matches optimizer updates across TP/CP/PP/EP combinations, aligns on a 100-step real RL trace replay, reaches up to 4.395x speedup (2.930x under a conservative compile-on comparison) as prefix ratio and rollout group size grow, and reduces Phase-B peak HBM by up to 59.1%, extending the Llama3-8B capacity frontier from 17,920 to 29,696 total tokens.

Mengying Ge, Mingyang Li, Dongkai Tang et al.

In this paper, we present our solutions for emotion recognition in the sub-challenges of Multimodal Emotion Recognition Challenge (MER2024). To mitigate the modal competition issue between audio and text, we adopt an early fusion strategy based on a large language model, where joint training of audio and text is conducted initially. And the joint Audio-Text modal feature will be late-fused with other unimodal features. In order to solve the problems of data insufficiency and class imbalance, We use multiple turns of multi-model voting for data mining. Moreover, to enhance the quality of audio features, we employ speech source separation to preprocess audios. Our model ranks \textbf{2nd} in both MER2024-SEMI and MER2024-NOISE, validating our method's effectiveness.

Longfei Guo, Pengbo Li, Ting Gao et al.

With the rapid advancement of AI technology, we have seen more and more concerns on data privacy, leading to some cutting-edge research on machine learning with encrypted computation. Fully Homomorphic Encryption (FHE) is a crucial technology for privacy-preserving computation, while it struggles with continuous non-polynomial functions, as it operates on discrete integers and supports only addition and multiplication. Spiking Neural Networks (SNNs), which use discrete spike signals, naturally complement FHE's characteristics. In this paper, we introduce FHE-DiCSNN, a framework built on the TFHE scheme, utilizing the discrete nature of SNNs for secure and efficient computations. By leveraging bootstrapping techniques, we successfully implement Leaky Integrate-and-Fire (LIF) neuron models on ciphertexts, allowing SNNs of arbitrary depth. Our framework is adaptable to other spiking neuron models, offering a novel approach to homomorphic evaluation of SNNs. Additionally, we integrate convolutional methods inspired by CNNs to enhance accuracy and reduce the simulation time associated with random encoding. Parallel computation techniques further accelerate bootstrapping operations. Experimental results on the MNIST and FashionMNIST datasets validate the effectiveness of FHE-DiCSNN, with a loss of less than 3\% compared to plaintext, respectively, and computation times of under 1 second per prediction. We also apply the model into real medical image classification problems and analyze the parameter optimization and selection.

Yankai Wang, Yiding Sun, Qirui Wang et al.

Understanding spatial dynamics and semantics in point cloud is fundamental for comprehensive 3D comprehension. While reinforcement learning algorithms such as Group Relative Policy Optimization (GRPO) have recently achieved remarkable breakthroughs in large language models by incentivizing reasoning capabilities through strategic reward design, their potential remains largely unexplored in the 3D perception domain. This naturally raises a pivotal question: Can RL-based methods effectively empower 3D point cloud fine-tuning? In this paper, we propose PointRFT, the first reinforcement fine-tuning paradigm tailored specifically for point cloud representation learning. We select three prevalent 3D foundation models and devise specialized accuracy reward and dispersion reward functions to stabilize training and mitigate distribution shifts. Through comprehensive few-shot classification experiments comparing distinct training paradigms, we demonstrate that PointRFT consistently outperforms vanilla supervised fine-tuning (SFT) across diverse benchmarks. Furthermore, when organically integrated into a hybrid Pretraining-SFT-RFT paradigm, the representational capacity of point cloud foundation models is substantially unleashed, achieving state-of-the-art performance particularly under data-scarce scenarios.

Pengbo Li, Yiding Sun, Haozhe Cheng

Self-supervised representation learning has shown significant improvement in Natural Language Processing and 2D Computer Vision. However, existing methods face difficulties in representing 3D data because of its unordered and uneven density. Through an in-depth analysis of mainstream contrastive and generative approaches, we find that contrastive models tend to suffer from overfitting, while 3D Mask Autoencoders struggle to handle unordered point clouds. This motivates us to learn 3D representations by sharing the merits of diffusion and contrast models, which is non-trivial due to the pattern difference between the two paradigms. In this paper, we propose \textit{PointDico}, a novel model that seamlessly integrates these methods. \textit{PointDico} learns from both denoising generative modeling and cross-modal contrastive learning through knowledge distillation, where the diffusion model serves as a guide for the contrastive model. We introduce a hierarchical pyramid conditional generator for multi-scale geometric feature extraction and employ a dual-channel design to effectively integrate local and global contextual information. \textit{PointDico} achieves a new state-of-the-art in 3D representation learning, \textit{e.g.}, \textbf{94.32\%} accuracy on ScanObjectNN, \textbf{86.5\%} Inst. mIoU on ShapeNetPart.

Marina Mao, Alexandr Liu, Pengbo Li et al.

Memory systems are widely adopted to enhance LLMs for long-horizon tasks, and are commonly organized as multi-agent pipelines with memory building, summarizing, and retrieval agents. To empower this system, existing RL-based methods either apply final downstream task rewards (e.g., QA accuracy) for all agents uniformly, which are coarse and ambiguous, or design task-specific rewards for agents on different subtasks, which require costly annotations (e.g., key evidence) and are difficult to define reliably. To address these limitations, we propose Tree-based Credit Assignment for Multi-Agent Memory Systems (TreeMem), which derives agent-specific credit from the final reward without task-specific annotations. Specifically, TreeMem extends the multi-agent pipeline (builder--summarizer--retrieval) into a tree structure, where each agent's outputs are expanded into multiple subsequent branches. The contribution of each agent is estimated via Monte Carlo averaging over its subsequent branches, capturing how intermediate agent actions may influence the final reward. This converts the coarse final reward into agent-specific optimization signals. These signals are then used to update all agent policies simultaneously, helping heterogeneous agents specialize effectively. Experiments on long-horizon benchmarks show that TreeMem improves memory system performance over strong baselines, validating the effectiveness of tree-structured credit assignment for the multi-agent memory system.

Chaoyi Lu, Yiding Sun, Pengbo Li et al.

As an emerging paradigm of federated learning, asynchronous federated learning offers significant speed advantages over traditional synchronous federated learning. Unlike synchronous federated learning, which requires waiting for all clients to complete updates before aggregation, asynchronous federated learning aggregates the models that have arrived in realtime, greatly improving training speed. However, this mechanism also introduces the issue of client model version inconsistency. When the differences between models of different versions during aggregation become too large, it may lead to conflicts, thereby reducing the models accuracy. To address this issue, this paper proposes an asynchronous federated learning version correction algorithm based on knowledge distillation, named FedADT. FedADT applies knowledge distillation before aggregating gradients, using the latest global model to correct outdated information, thus effectively reducing the negative impact of outdated gradients on the training process. Additionally, FedADT introduces an adaptive weighting function that adjusts the knowledge distillation weight according to different stages of training, helps mitigate the misleading effects caused by the poorer performance of the global model in the early stages of training. This method significantly improves the overall performance of asynchronous federated learning without adding excessive computational overhead. We conducted experimental comparisons with several classical algorithms, and the results demonstrate that FedADT achieves significant improvements over other asynchronous methods and outperforms all methods in terms of convergence speed.

Di Chai, Pengbo Li, Feiyuan Zhang et al.

Token filtering has been proposed to enhance utility of large language models (LLMs) by eliminating inconsequential tokens during training. While using fewer tokens should reduce computational workloads, existing studies have not succeeded in achieving higher efficiency. This is primarily due to the insufficient sparsity caused by filtering tokens only in the output layers, as well as inefficient sparse GEMM (General Matrix Multiplication), even when having sufficient sparsity. This paper presents Collider, a system unleashing the full efficiency of token filtering in LLM training. At its core, Collider filters activations of inconsequential tokens across all layers to maintain sparsity. Additionally, it features an automatic workflow that transforms sparse GEMM into dimension-reduced dense GEMM for optimized efficiency. Evaluations on three LLMs-TinyLlama-1.1B, Qwen2.5-1.5B, and Phi1.5-1.4B-demonstrate that Collider reduces backpropagation time by up to 35.1% and end-to-end training time by up to 22.0% when filtering 40% of tokens. Utility assessments of training TinyLlama on 15B tokens indicate that Collider sustains the utility advancements of token filtering by relatively improving model utility by 16.3% comparing to regular training, and reduces training time from 4.7 days to 3.5 days using 8 GPUs. Collider is designed for easy integration into existing LLM training frameworks, allowing systems already using token filtering to accelerate training with just one line of code.