首页 > AI前沿 > WM-Quant: Rollout-Aware Post-Training Quantization for JEPA-Style World Models

WM-Quant: Rollout-Aware Post-Training Quantization for JEPA-Style World Models

arXiv机器学习 2026-11-17 04:51 8 阅读 查看原文
用户现在需要把给定的新闻正文排版成语义化的HTML片段,规则很明确。首先,原文是一段介绍WM-Quant的内容,没有分小节的需求?不对,看原文内容,是不是可以用h2?不对,原文是连贯的介绍,然后有三个组件的内容。首先,第一个自然段落,整体是介绍的话,然后里面的三个组件是列表吗?不对,原文里说“three components that build on existing PTQ methods: early-encoder precision, control-path preservation, and rollout-aware scaling.”然后后面分别解释这三个。首先,先处理段落,第一个

可以放开头到rollout-aware scaling之前?不对,看原文结构:第一句到“near-random.”是介绍背景和问题,然后接下来介绍三个组件,然后分别解释每个组件,然后最后讲效果。那结构应该是: 首先,

?不对,原文没有明确的小节标题,哦,不对,用户说长文用h2/h3划分小节,这里的原文是不是可以分成几个部分?不对,再仔细看原文: 第一部分:介绍WM-Quant是首个针对JEPA式世界模型的PTQ框架,背景和问题:世界模型参数少但规划需要大量推理,PTQ能降成本,但自回归MPC rollout里量化误差会累积导致动作排序接近随机。这部分可以作为一个

? 然后第二部分:介绍WM-Quant的三个组件,这里可以用

吗?不对,或者三个组件的解释用列表?不对,原文里三个组件的具体内容是分别解释的,比如“Early-encoder precision protects...”“Control-path preservation allows...”“Rollout-aware scaling jointly calibrates...”这三个是并列的,所以可以用
  • 包裹每个组件的解释? 然后最后一部分是效果:“Across multiple world models...”这部分作为另一个

    。 等下,要保留所有信息,不能增删改。那梳理: 1. 第一个段落:We present WM-Quant, to our knowledge, the first post-training quantization (PTQ) framework designed specifically for JEPA-style world models. Although world models typically feature relatively few parameters, action planning algorithms often require thousands of inferences, which can quickly become computationally expensive. PTQ is a natural way to reduce this cost. However, we show that in world models using autoregressive model predictive control (MPC) rollouts, quantization error can compound with each predictor step, resulting in action rankings from the quantized model becoming near-random. 这部分用

    包裹。 2. 然后,接下来介绍三个组件,这里可以用

    ?不对,有没有合适的标题?不对,原文没有标题,哦,用户说长文用h2/h3划分小节,如果原文有逻辑分段的话可以用。或者,这里可以用

    ?不对,再看,原文中“WM-Quant addresses this with three components that build on existing PTQ methods: early-encoder precision, control-path preservation, and rollout-aware scaling.”这句话是引出三个组件,然后后面三个分别解释,所以可以: 用

    WM-Quant的核心组件

    ?不对,不行,用户说不能增删内容,哦对!规则4是保留原文所有信息,不得增删改内容,所以不能加原文没有的标题!哦对,我差点忘了,不能加原文没有的文字,所以h2/h3的标题必须是原文里的?但原文里没有小标题,那是不是就不用h2/h3?不对,再看原文,有没有适合的?哦,原文里没有现成的小标题,那可能不需要用h2/h3?不对,再看规则2:“长文用

    /

    划分小节”,如果没有现成的小节标题,是不是就不用?那再处理: 然后,引出三个组件的句子,然后三个组件的解释用
      ?比如:

      WM-Quant addresses this with three components that build on existing PTQ methods: early-encoder precision, control-path preservation, and rollout-aware scaling.

      然后
      • Early-encoder precision protects accurate latent state and goal information to mitigate long-term error propagation.
      • Control-path preservation allows the world model predictor to distinguish between fine-grained actions, significantly improving action rankings and execution.
      • Rollout-aware scaling jointly calibrates encoder and per-step predictor activation ranges with respect to an entire rollout rather than a single prediction, reducing the cumulative rollout error.
      然后最后一个段落:

      Across multiple world models, environments, and PTQ methods, WM-Quant substantially improves planning success at low bit-widths, approaching full-precision performance at W4A4.

      不对,等下,那前面的第一个段落和这个引出组件的段落,是不是可以合并?不,原文的标点是这样的: 原文的完整内容是: We present WM-Quant, to our knowledge, the first post-training quantization (PTQ) framework designed specifically for JEPA-style world models. Although world models typically feature relatively few parameters, action planning algorithms often require thousands of inferences, which can quickly become computationally expensive. PTQ is a natural way to reduce this cost. However, we show that in world models using autoregressive model predictive control (MPC) rollouts, quantization error can compound with each predictor step, resulting in action rankings from the quantized model becoming near-random. WM-Quant addresses this with three components that build on existing PTQ methods: early-encoder precision, control-path preservation, and rollout-aware scaling. Early-encoder precision protects accurate latent state and goal information to mitigate long-term error propagation. Control-path preservation allows the world model predictor to distinguish between fine-grained actions, significantly improving action rankings and execution. Rollout-aware scaling jointly calibrates encoder and per-step predictor activation ranges with respect to an entire rollout rather than a single prediction, reducing the cumulative rollout error. Across multiple world models, environments, and PTQ methods, WM-Quant substantially improves planning success at low bit-widths, approaching full-precision performance at W4A4. 哦,原来原文的标点是,前面到near-random.是一个句子结束,然后接下来是WM-Quant addresses...这个句子,然后是三个分别解释的句子,然后是最后一个句子。那现在排版的话: 首先,第一个

      可以包含从开头到“near-random.”的内容,对吧?因为是一个完整的语义段。 然后,接下来,“WM-Quant addresses this with three components that build on existing PTQ methods: early-encoder precision, control-path preservation, and rollout-aware scaling.”这句话,后面跟着三个组件的解释,这三个解释是并列的,所以可以用

        ,那这句话是不是可以放在
          的前面?或者,把这句话也放在一个

          里,然后后面是

            ? 不对,有没有更合适的?比如,把引出三个组件的句子和列表放在一起?或者,因为三个组件的解释是对前面三个组件名称的详细说明,所以:

            We present WM-Quant, to our knowledge, the first post-training quantization (PTQ) framework designed specifically for JEPA-style world models. Although world models typically feature relatively few parameters, action planning algorithms often require thousands of inferences, which can quickly become computationally expensive. PTQ is a natural way to reduce this cost. However, we show that in world models using autoregressive model predictive control (MPC) rollouts, quantization error can compound with each predictor step, resulting in action rankings from the quantized model becoming near-random.

            WM-Quant addresses this with three components that build on existing PTQ methods: early-encoder precision, control-path preservation, and rollout-aware scaling.

            • Early-encoder precision protects accurate latent state and goal information to mitigate long-term error propagation.
            • Control-path preservation allows the world model predictor to distinguish between fine-grained actions, significantly improving action rankings and execution.
            • Rollout-aware scaling jointly calibrates encoder and per-step predictor activation ranges with respect to an entire rollout rather than a single prediction, reducing the cumulative rollout error.

            Across multiple world models, environments, and PTQ methods, WM-Quant substantially improves planning success at low bit-widths, approaching full-precision performance at W4A4.

            这样是不是符合要求?检查规则: 1. 只输出HTML片段,没有围栏和解释,对。 2. 每个自然段落用

            ,这里三个

            ,都是自然段落,对。有没有需要h2/h3的?因为原文没有现成的小节标题,所以不需要,规则是长文用,这里如果没有现成的标题,就不用,避免添加内容。 3