optimization

该推文总结了优化 LLM 推理的 12 种 KV 缓存技术，包括滑动窗口、跨层共享、MLA、量化、前缀复用、分页、驱逐和 GQA/MQA 等，帮助开发者根据具体瓶颈选择合适的组合。

A deep-dive debugging writeup showing how antcc's -O2 pessimization made a Lua benchmark 1.4x slower than the unoptimized build, tracing the 2.4x code-size blowup in luaV_execute past inlining heuristics and IR expansion into instruction selection and register allocation.

该论文研究了在结果依赖截断和删失条件下的生存分析模型高效学习问题，证明了在已知基线风险时 Cox 回归可通过一阶优化高效学习，同时构造了存在严格次优局部极小值的可识别模型，并在实验中展示了相比期望最大化方法的数量级加速。

DuoDiscover 是一个自主发现搜索系统，通过让编码代理协同发现优化目标与快速代理评估器（yardstick），并用保真度测试防止代理指标漂移，来突破自主发现中的评估瓶颈。在 SWEfficiency、SOL-Exec、vLLM、SQLite 及 CPU 微架构优化等多个基准上取得显著成果，部分任务上超越 AlphaEvolve 系统与人类专家。

A hobbyist details running Next Flash (IQ3_XXS) at ~70 tok/s with 100k context and two Qwen 3.6 35B A3B instances at ~145 tok/s with 256k context on a 4-board $500 AMD BC-250 mining cluster, using llama.cpp with Vulkan RPC and aggressive optimizations like MTP speculative decoding, recorded Vulkan command replay, and improved context handling.

一篇技术博客，介绍如何利用浮点数指数相当于对数这一特性，用 AVX2 浮点指令实现可向量化的 clz（前导零计数）和 ctz（尾随零计数）多播（polyfill），并在 Haswell/Alder Lake 上给出优于标量版本的性能数据。

MaRN (Mapping Networks) is a new open-source PyTorch library that trains neural networks by optimizing a compact latent representation instead of all parameters directly, showing up to 57.7× parameter reduction on MNIST with modest accuracy trade-offs. The author presents exploratory benchmarks and invites community feedback on the approach.

一篇技术博客深入解释了投机解码为什么快：并非因为总FLOPs减少（实际上反而增加），而是因为它改变了GPU的工作类型——在低批量场景下把原本空闲的算力从内存瓶颈（memory-bound）转为计算瓶颈（compute-bound），用单序列的'深度'换取了批量解码的'广度'。

该论文研究多智能体bandit学习中探索信号泄露带来的安全隐患，提出deceptive bandit learning（DBL）动力学，证明对手可通过耦合自身探索注入外部性使系统收敛到新的deceptive Nash equilibrium（DNE），并保持最优收敛速率且放宽了二阶光滑性假设。

FedRSPO+：一种面向决策聚焦联邦学习的异构感知框架，基于 RSPO+ 正则化预测-优化替代损失，通过平滑决策映射在客户端目标与约束异构时保持稳定优化，并给出随异构程度缩放、在同构时消失的跨客户端误差界。

The paper introduces ORACLE, an optimizer-relative constrained learning method that assesses constraint compatibility on the post-optimizer update rather than the objective or gradient, showing improvements over native optimizers in 94% of configurations across eight PDE benchmarks.

rauchg argues that any program can be endlessly hardened and optimized, but with real costs in time and attention—even in a world of infinite tokens, engineers must decide when to stop, since AI agents will happily drill no matter what.

Quail，一个开源 AI-SQL 引擎，新增了前缀共享（prefix sharing）支持，在对 1,772 条长 agent 轨迹使用 qwen3-4b 打标签的示例中，计算 token 减少 3 倍，查询速度提升 2.6 倍。

A tweet claims that combining Opus 5.5 with Jev is the most powerful AI model setup, allegedly 200x faster and 400x cheaper than running Opus alone, promising to reveal the setup details.

Meta is open-sourcing Rebalancer, an assignment-problem solver used internally for nine years to handle resource allocation across its infrastructure, including hardware placement, service/task placement, and traffic routing. It offers both an optimal MIP solver and a parallelized local search solver, handling ~40 million problems daily at Meta.

The Dust paper introduces the first zeroth-order pretraining method competitive with backpropagation for transformers, using per-token activation perturbations that make each token a virtual population member evaluated in a single forward pass. It reports up to 10^3–10^4x greater efficiency than weight-space ES methods like EGGROLL and finds that larger models are more population-efficient, suggesting zeroth-order methods could scale in compute-rich regimes.

MuLoRA introduces a spectrally balanced low-rank adaptation method for continual learning that identifies 'spectral plasticity collapse' in LoRA updates and addresses it via historical whitening and momentum-based orthogonalization, achieving the best mean accuracy in 15 of 16 reported metrics across class-incremental benchmarks.

An essay arguing that human intelligence is a rare evolutionary accident rather than an inevitable endpoint, and that AI's ability to optimize directly against explicit objectives makes defining worthwhile goals—not building capability—the harder problem.

Strata 0.1.39 has been released with multi-GPU optimization. According to user feedback, inference speeds in some scenarios have doubled — benchmarks on a 4090 48G running the hundred-billion-parameter Qwen3.8-Flash-Next showed a significant boost in tokens/s.

MetaKernelBench introduces a benchmark measuring whether experience distilled by GPU kernel optimization agents in one DSL (CuTeDSL) transfers to another (TIRx) on the same problem, finding mixed results with lift ranging from -19% to +29% across six models.