📰
Meta ProgramBench 測試 AI 無網重建程式
Meta Superintelligence Lab 推出 ProgramBench 基準測試，挑戰 SOTA AI 模型無需網路存取，從零重建如 ffmpeg、SQLite 和 ripgrep 等複雜可執行程式。Reddit 貼文強調此 AI 程式合成能力評估。質疑當前頂尖模型能否達成。
Search
Tag: #program-synthesis8 results
AlgoEvolve：用於演算法交易的 LLM 驅動元演化框架
AlgoEvolve 是一個利用 LLM 生成並迭代改進可執行 Python 交易策略的新框架。它具備一個元演化迴圈，能優化基於提示詞的搜尋啟發式演算法，以適應嘈雜且非平穩的市場環境。
SymLang：從噪聲數據發現支配方程
SymLang 是一個新框架，整合對稱約束文法、語言模型引導的程式合成，以及貝氏模型選擇，從噪聲和部分觀測中發現支配方程。它在 10% 噪聲下，對 133 個動力系統達到 83.7% 精確結構恢復率，優於基準 22.4 個百分點。此開源工具將外推誤差降低 61%，並最小化守恆律違反。
🔬
將 LLM 軌跡轉化為確定性流程
一項研究討論探討如何將重複性的 LLM 工作負載，自動替換為由正規表示式、解析器、傳統 ML 與 NLP 運算子組成的具型別 DAG。提議中的系統會合成並驗證流程、最佳化品質、成本與延遲，並在不確定性較高時回退至前沿模型。
Cadmus: Low-Cost Program Synthesis System
Apple ML introduces Cadmus, a small-scale system for autoregressive program synthesis. It features an integer virtual machine, a dataset of diverse true programs, and a transformer model trained for under $200 compute. This setup enables controlled experiments bypassing issues with large LLMs like OOD challenges and high resource demands.
Cadmus Enables Cheap Program Synthesis Experiments
Apple Machine Learning introduces Cadmus, a small-scale system for autoregressive program synthesis. It features an integer virtual machine, a dataset of diverse true programs, and a transformer model trained for under $200 compute. This setup allows controlled experimentation without the complexities of large LLMs.
Cadmus: Cheap Program Synthesis System
Apple unveils Cadmus, a small-scale system for autoregressive program synthesis. It features an integer VM, diverse program dataset, and transformer model trained under $200 compute. Enables controlled experiments bypassing LLM challenges like OOD and tokenization.
Cadmus: Affordable Autoregressive Program Synthesis
Apple ML introduces Cadmus, a small-scale system for autoregressive program synthesis. It features an integer virtual machine, a dataset of diverse true programs, and a transformer model trained for under $200 compute. This setup enables controlled experiments avoiding LLM pitfalls like OOD issues and high compute demands.