SRFT团队 投稿 量子位 | 公众号 QbitAI 通过单阶段监督微调与强化微调结合，让大模型在训练时能同时利用专家演示和自我探索试错，有效提升大模型推理性能。 中国科学院自动化研究所深度强化学习团队 联合 美团 ，提出一种 单阶段监督-强化微调方法——SRFT (Supervised Reinforcement Fine-Tuning) 。该方法通过基于熵的动态加权机制，将两种训练范式结合。 在大语言模型（LLM）的推理能力提升上，监督微调（SFT） 和强化学习（RL，有时也称作强化微调，RFT）是两条核心技术路线。但它们 各自都存在瓶颈： SFT擅长模仿专家解题思路，类似"背书"，能快速为模型打下基础，但缺点是容易陷入死记硬背，缺乏在新问题上灵活应用和寻找最优解的能 力； RFT/RL通过不断试错来探索解题方法，类似"刷题"，能够发现更优解法，但其探索过程效率低下，容易面临模式崩溃风险。 因此，目前研究者通常采用两阶段 顺序 方法SFT→RFT/RL：先用SFT学习高质量数据集，再用RFT/RL进一步优化对齐LLM策略（即先"背 完书"再"去刷题"）。 然而，这种串行方式不仅影响学习效率，还常常导致模型 ...

美国纽约联储就常备回购便利工具（SRF）发布公告。 ...

美联储官员和投资者正密切关注银行存放在央行的现金规模，以判断何时应停止资产负债表缩减(即量 化紧缩，QT)。若手头现金增加，表明系统内流动性更充裕，美联储可延长QT进程。 最新数据显示，截至5月14日当周，美联储资产负债表规模为3.24万亿美元，高于前一周的3万亿美元， 略低于近三年前启动QT时的水平。华尔街策略师估计，为维持充足流动性并避免市场压力，美联储需 将资产负债表规模维持在3万亿至3.25万亿美元以上。 常备回购便利(SRF) 智通财经APP获悉，纽约联邦储备银行负责管理庞大证券投资组合的官员表示，美联储缩减资产负债表 的举措已开始对回购协议市场施加压力。 纽约联储系统公开市场账户经理罗伯托·佩尔利(Roberto Perli)周四在预先准备的讲话中指出，这种压力 可能意味着美联储控制短期利率的工具将愈发重要。他在纽约联储与哥伦比亚大学国际与公共事务学院 (SIPA)联合举办的活动中表示，随着美联储资产负债表规模持续缩减，银行储备从"充裕"转向"充足"水 平，货币市场利率面临的上行压力可能加剧。 佩尔利称，货币市场利率的变动"代表流动性状况的正常化，无需担忧"。但他补充道，这确实意味 着"未来常备回 ...

Core Viewpoint - The Federal Reserve is encouraging financial institutions to actively utilize the Standing Repo Facility (SRF) to address market liquidity challenges, despite current liquidity being ample [1] Group 1: SRF Tool Utilization - Federal Reserve officials advocate for the use of the SRF tool under economically reasonable circumstances to support effective monetary policy implementation and promote market stability [1] - The New York Fed plans to expand SRF operations from only afternoon sessions to include morning sessions, allowing for same-day settlement, which is a significant step to enhance the tool's effectiveness [1] Group 2: Market Liquidity and Balance Sheet - There are indications that market liquidity is tightening, and the Fed's balance sheet reduction may take some time to complete [1] - As the Fed continues to reduce its balance sheet and lower reserve levels, upward pressure on money market rates may increase [1]

Core Viewpoint - The New York Federal Reserve plans to incorporate early settlement operations of a key liquidity support tool into its regular schedule to enhance and strengthen this liquidity tool, supporting stable financial market operations [1] Group 1: Federal Reserve Actions - The Federal Open Market Committee (FOMC) agreed in March to "significantly slow" the pace of balance sheet reduction to prevent excessive liquidity withdrawal from the market [1] - The New York Fed's recent normalization of early repurchase operations is seen as a measure to ensure market stability and liquidity, especially in light of volatility in the U.S. Treasury market due to tariff policies [1][6] - The Fed is expanding the liquidity "insurance layer" in financial markets to prevent short-term funding mismatches and liquidity shortages during high yield fluctuations [1][7] Group 2: Market Reactions and Concerns - The global financial market's renewed focus on the Standing Repo Facility (SRF) follows significant volatility in the U.S. Treasury market triggered by new trade policies [5] - Concerns about market movements due to trade policy uncertainty have led to a "real and significant" deterioration in financial market liquidity, although the repo market has shown resilience [5][6] - The New York Fed had already begun providing additional daily repo operations before the recent market volatility, aiming to prevent repo market rates from exceeding the Fed's target range [6] Group 3: Future Implications - The adjustments in monetary policy reflect the Fed's intention to add safety measures against potential funding mismatches in a high-rate environment, avoiding a repeat of the 2019 liquidity crisis [7] - The FOMC noted that while reserves remain high at approximately $3 trillion, certain indicators are nearing buffer limits, necessitating caution to avoid approaching critical points similar to those in 2019 [7] - If the 10-year Treasury yield exceeds 5% again, the Fed may consider pausing balance sheet reduction or even targeted expansion to stabilize the financial market's benchmark interest rate corridor [7]