首页 > AI前沿 > Late Entry in Looped Language Models

Late Entry in Looped Language Models

arXiv自然语言 2026-11-18 23:09 19 阅读 查看原文
用户要求将一段纯文本的英文新闻正文排版为语义化的HTML片段。需要遵循以下规则: 1. 只输出HTML片段本身,不要Markdown围栏,不要解释。 2. 每个自然段落用`

`包裹,长文用`

`/`

`划分小节。 3. 列表用`
  • `,引用用`
    `,代码用`
    `,重要句子可用``。
    4. 保留原文所有信息,不得增删改内容。
    5. 不要输出``、``、``等文档级标签。
    
    分析原文:
    原文是一段连续的英文科技新闻文本,主要讲述了“循环语言模型”的“延迟进入(late entry)”解码策略及其性能表现。
    原文共有5个句子,可以划分为几个逻辑段落或小节,以符合“长文用`

    `/`

    `划分小节”的要求。 原文拆解: 句子1:Looped language models deepen reasoning at a compact parameter count by iterating a shared layer stack T times per token, turning recurrence depth directly into sequential decoding latency. 句子2:Conventional early-exit decoding halts after initial passes without cache completion, truncating cached key-value states of past tokens to shallow depth. 句子3:We study the opposite strategy, late entry, where new tokens skip initial passes and enter directly into the final K < T passes over a deep context cache. 句子4:Once lightweight fine-tuning resolves hidden-state norm mismatch at the entry boundary, deep cached context substitutes for skipped token-local passes, allowing late entry to outperform matched early exit across model families and pass budgets. 句子5:To turn this asymmetry into wall-clock speedups without stalling to build deep caches, candidate blocks verify through only the final K passes on the critical path while previously committed tokens complete passes 1 -> T alongside subsequent steps at zero extra serial calls. 句子6:Projecting these finished T-pass wake states inside the existing output GEMM (Wake verification, governed by logit-margin tolerance tau in [0, inf]) unifies greedy-exact T-pass decoding (tau = 0, matching K = T block decoding within bf16 reduction order), margin-gated verification (tau > 0), and non-rollback lossy late entry (tau = inf) inside a single static CUDA graph. 句子7:On Ouro-2.6B (T = 4), K = 2 Wake reaches 5.2x greedy-exact speedup at tau = 0 (+26% over K = 4 block decoding on GSM8K) and 6.8x at tau = inf (+16.9 points over trained early exit at matched speed). 句子8:Scaling test-time recurrence across T in {8, 16, 32, 64} on Huginn-3.5B cuts K = T/4 rollbacks 21.5x with depth, driving greedy-exact (tau = 0) speedups to 6.1x (T = 32) and 8.4x (T = 64, +153% over K = 64 block) and tau = inf speedups to 10.6 to 12.5x. 为了结构清晰,可以划分如下:

    背景与策略

    句子1。

    句子2。

    句子3。

    句子4。

    加速机制

    句子5。

    句子6。

    实验结果

    句子7。

    句子8。

    检查是否有需要加粗的重要句子。 “We study the opposite strategy, late entry...” 可以加粗。 “Once lightweight fine-tuning resolves hidden-state norm mismatch at the entry boundary, deep cached context substitutes for skipped token-local passes, allowing late entry to outperform matched early exit across model families and pass budgets.” 可以加粗。 或者将关键的技术名词加粗,比如 late entry。规则说“重要句子可用”,所以对关键句子整体加粗或部分加粗都可以。为了自然,可以对核心结论句子加粗。 调整排版:

    研究背景与策略

    Looped language models deepen reasoning at a compact parameter count by iterating a shared layer stack T times per token, turning recurrence depth directly into sequential decoding latency. Conventional early-exit decoding halts after initial passes without cache completion, truncating cached key-value states of past tokens to shallow depth.

    We study the opposite strategy, late entry, where new tokens skip initial passes and enter directly into the final K < T passes over a deep context cache.

    Once lightweight fine-tuning resolves hidden-state norm mismatch at the entry boundary, deep cached context substitutes for skipped token-local passes, allowing late entry to outperform matched early exit across model families and pass budgets.

    加速机制与统一框架

    To turn this asymmetry into wall-clock speedups without stalling to build deep caches, candidate blocks verify through only the final K passes on the critical path while previously committed tokens complete passes 1 -> T alongside subsequent steps at zero extra serial calls.

    Projecting these finished T-pass wake states inside the existing output GEMM (Wake verification, governed by logit-margin tolerance tau in [0, inf]) unifies greedy-exact T-pass decoding (tau = 0, matching K = T block decoding within bf16 reduction order), margin-gated verification (tau > 0), and non-rollback lossy late entry (tau = inf) inside a single static CUDA graph.

    实验表现

    On Ouro-2.6B (T = 4), K = 2 Wake reaches 5.2x greedy-exact speedup at tau = 0 (+26% over K = 4 block decoding on GSM8K) and 6.8x at tau = inf (+16.9 points over trained early exit at matched speed).

    Scaling test-time recurrence across T in {8, 16, 32, 64} on Huginn-3.5B cuts K = T/4 rollbacks 21.5x with depth, driving greedy-exact (tau = 0) speedups to 6.1x (T = 32) and 8.4x (T = 64, +153% over K = 64 block) and tau = inf speedups to 10.6 to 12.5x.

    检查规则: 1. 只输出HTML片段本身,不要Markdown围栏、不要解释。 -> 符合。 2. 每个自然段落用

    包裹,长文用

    /

    划分小节。 -> 符合,用了h2和p。 3. 列表内容用
    • ,引用用
      ,代码用
      ,重要句子可用。 -> 原文没有列表和引用,使用了code和strong。这里1 -> T,不需要
      ,因为只是行内。如果规则要求代码用
      ,行内代码用也是合理的,或者把数学公式/参数用
      ?不,
      通常用于多行代码块。这里用即可,或者严格点用
      ?"代码用
      ",如果只有行内,足够,但为了严格遵守,如果不使用代码也可以。不过把 "1 -> T" 和 "[0, inf]" 和 "{8, 16, 32, 64}" 用  包裹是合理的,如果不用代码标签也没问题。为了安全,不用代码标签,避免被判定违规,或者用。规则说“代码用
      ”,意味着如果有代码块就用这个,行内代码用。这里没有代码块,所以不用
      。
      4. 保留