https://xuquant.com/en/posts/Recent content in Posts on Xu'Bloghttps://xuquant.com/og-default.pnghttps://xuquant.com/og-default.pngHugo -- 0.152.2enSat, 07 Mar 2026 14:00:00 +0800https://xuquant.com/en/posts/foundation-models/qwen3-vs-qwen3-5-architecture/Sat, 07 Mar 2026 14:00:00 +0800https://xuquant.com/en/posts/foundation-models/qwen3-vs-qwen3-5-architecture/A deep architectural comparison of Qwen3.5 versus Qwen3, examining hybrid attention, joint multimodal training, high-sparsity MoE, and partial RoPE across attention, vision, and MoE dimensionshttps://xuquant.com/en/posts/foundation-models/coral-autonomous-multi-agent-evolution/Sat, 22 Nov 2025 10:00:00 +0800https://xuquant.com/en/posts/foundation-models/coral-autonomous-multi-agent-evolution/How delegating evolutionary search decisions to autonomous agents—rather than relying on fixed heuristics—enables faster convergence and stronger results across mathematical and systems optimization tasks.https://xuquant.com/en/posts/foundation-models/inspatio-world-4d-simulator/Sat, 25 Oct 2025 10:00:00 +0800https://xuquant.com/en/posts/foundation-models/inspatio-world-4d-simulator/A deep technical analysis of InSpatio-World: a 1.3B-parameter real-time 4D world simulator that combines implicit spatiotemporal caching with explicit geometric constraints, achieving 24 FPS novel-view synthesis from monocular video.https://xuquant.com/en/posts/autonomous-driving/basic_rl/Sat, 20 Sep 2025 10:00:00 +0800https://xuquant.com/en/posts/autonomous-driving/basic_rl/Comprehensive analysis of applying reinforcement learning to end-to-end autonomous driving, covering metric caching, Direct Preference Optimization (DPO) across action representations, and strategies for breaking sampling ceilings in iterative self-improvement.https://xuquant.com/en/posts/foundation-models/deepseek_series1_mla/Sat, 13 Sep 2025 10:00:00 +0800https://xuquant.com/en/posts/foundation-models/deepseek_series1_mla/MLA from the deployment perspective of DeepSeek V2/V3: KV cache compression ratio, inference throughput, engineering comparison with GQA/MQA, and real-world gains under long context. The mathematical derivation is in the companion article.https://xuquant.com/en/posts/autonomous-driving/nvidia_vla/Sat, 30 Aug 2025 10:00:00 +0800https://xuquant.com/en/posts/autonomous-driving/nvidia_vla/Technical deep-dive into Nvidia's Alpamayo VLA system for autonomous driving, built on the Cosmos-Reason VLM backbone, covering tri-plane vision encoding, ego-shortcut avoidance, Cause-of-Change dataset paradigm, and reasoning-action alignment via reinforcement learning.https://xuquant.com/en/posts/autonomous-driving/rl-policy-optimization-e2e-driving/Sat, 09 Aug 2025 10:00:00 +0800https://xuquant.com/en/posts/autonomous-driving/rl-policy-optimization-e2e-driving/A systematic derivation of policy optimization methods for end-to-end autonomous driving: from REINFORCE through PPO to GRPO, covering advantage estimation, sampling differences between LLM and driving, multi-objective loss design, and the role of noise in diffusion-based exploration.https://xuquant.com/en/posts/autonomous-driving/e2e-autonomous-driving-evolution/Sat, 19 Jul 2025 10:00:00 +0800https://xuquant.com/en/posts/autonomous-driving/e2e-autonomous-driving-evolution/A technical survey on the architectural evolution of end-to-end autonomous driving, covering planner decoder selection (AR vs Diffusion vs Flow Matching), VLA integration strategies, and engineering best practices for data infrastructure, training optimization, and evaluation systems.https://xuquant.com/en/posts/autonomous-driving/ar-trajectory-tokenization/Sat, 28 Jun 2025 10:00:00 +0800https://xuquant.com/en/posts/autonomous-driving/ar-trajectory-tokenization/A deep dive into trajectory tokenization for autoregressive driving planners: from state-based discretization via k-means clustering, through token matching and reconstruction, to the AR+Diffusion paradigm and GRPO-based reinforcement learning post-training.https://xuquant.com/en/posts/autonomous-driving/generative-planning-nonconvex/Sat, 07 Jun 2025 10:00:00 +0800https://xuquant.com/en/posts/autonomous-driving/generative-planning-nonconvex/A first-principles analysis of why regression-based planners fail in autonomous driving: the feasible set is non-convex, MSE averages into obstacles, GMM is a patch not a solution, and generative approaches are necessary.