arXiv subjects

Haoyu Feng

Publications and source records attributed to Haoyu Feng.

3 recordsLinked to original sources

arXiv2025

GPU-Accelerated Belief Propagation for Program Analysis

Belief Propagation (BP) is a widely used approximate inference algorithm in probabilistic graphical models (PGMs), but is computationally expensive when applied to large-scale program analysis. Existing GPU-based approaches are unable to support flexible update strategies and have yet to integrate logical constraints with GPU acceleration, leading to challenges in both generality and efficiency. We present FastLBP, a GPU-accelerated BP framework for program analysis. We propose a unified representation for specifying flexible update strategies required in program analysis, along with a dependency analysis algorithm to enable parallel execution. Furthermore, we implement BP with local structures on GPUs by assigning individual threads to message computations and utilizing a memory-efficient representation. Experiments on SmartFL and BINGO show that FastLBP achieves average speedups of $17.42\times$ and $2.82\times$ over CPU-based approaches on SmartFL and BINGO, respectively, and $6.14\times$ over GPU-based approach on SmartFL, while preserving accuracy. Moreover, FastLBP supports update strategies that existing GPU-based approaches cannot support, demonstrating its improved generality for real-world program analysis.

cs.SE↗

arXiv2025

Multipulse Soliton Attractors Facilitated by High-Birefringence Fibers

The stability of optical solitons is a crucial factor in various applications. This work reveals a novel stable multipulse soliton attractor in fiber lasers. The attractor represents a bound state of multiple solitons, pulling other dynamical states toward itself. By introducing a polarization-maintaining fiber (PMF), the enhanced birefringence in the cavity induces the differential group delay (DGD) between the polarization components of the pulse, leading to soliton splitting and, ultimately, the formation of multi-soliton states. These multi-solitons, affected by intracavity nonlinear and dissipative effects, form attractors consisting of one to seven soliton pulses. Through a combination of experiments and numerical simulations, we systematically investigate the robustness and dynamical behavior of the multipulse soliton attractors. In this context, we analyze disturbances introduced by "rogue" solitons with different frequencies and conduct a comprehensive study of their interaction with the multipulse soliton attractors. The results show that the multipulse soliton attractors are structurally stable and highly resistant to external disturbances, highlighting their potential for high-precision fiber lasers and advanced multipulse systems.

physics.optics↗

arXiv2025

MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention

We introduce MiniMax-M1, the world's first open-weight, large-scale hybrid-attention reasoning model. MiniMax-M1 is powered by a hybrid Mixture-of-Experts (MoE) architecture combined with a lightning attention mechanism. The model is developed based on our previous MiniMax-Text-01 model, which contains a total of 456 billion parameters with 45.9 billion parameters activated per token. The M1 model natively supports a context length of 1 million tokens, 8x the context size of DeepSeek R1. Furthermore, the lightning attention mechanism in MiniMax-M1 enables efficient scaling of test-time compute. These properties make M1 particularly suitable for complex tasks that require processing long inputs and thinking extensively. MiniMax-M1 is trained using large-scale reinforcement learning (RL) on diverse problems including sandbox-based, real-world software engineering environments. In addition to M1's inherent efficiency advantage for RL training, we propose CISPO, a novel RL algorithm to further enhance RL efficiency. CISPO clips importance sampling weights rather than token updates, outperforming other competitive RL variants. Combining hybrid-attention and CISPO enables MiniMax-M1's full RL training on 512 H800 GPUs to complete in only three weeks, with a rental cost of just $534,700. We release two versions of MiniMax-M1 models with 40K and 80K thinking budgets respectively, where the 40K model represents an intermediate phase of the 80K training. Experiments on standard benchmarks show that our models are comparable or superior to strong open-weight models such as the original DeepSeek-R1 and Qwen3-235B, with particular strengths in complex software engineering, tool utilization, and long-context tasks. We publicly release MiniMax-M1 at https://github.com/MiniMax-AI/MiniMax-M1.

cs.CL↗