跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2610.24133 · v2

A Conditional No-Hair Theorem for Transformers: How Language Models Forget Their Prompt

首次提交 2026-10-17 03:03
版本更新 2026-11-17 18:37
雷达观测 2026-11-18 11:33
arXiv ↗PDF ↗

摘要

A black hole retains only a few conserved charges of the matter it formed from. We are inspired to make the analogy and we ask how much an autoregressive transformer retains of its prompt after generating tt tokens. For two prompts followed by the same continuation, the difference of their hidden states decays approximately as a power law t−αt^{-α}, where α=1α=1 is the rate at which a fixed prompt's share of uniform attention shrinks. Beyond the first layers, trained models forget faster (α≈0.55α\approx0.55--0.70.7) than untrained ones of the same architecture (α≈0.2α\approx0.2--0.50.5). For finite-depth pre-LN softmax transformers with bounded attention logits and non-degenerate pre-norm activations, we prove that this offset is O(P(log⁡t)L−1/t)O\big(P(\log t)^{L-1}/t\big) for prompt length PP and depth LL: the prompt's direct influence is diluted as the context grows, and \emph{relay} through continuation tokens adds only logarithmic factors. Conditional on the generated text, the prompt's influence on the next token thus vanishes. The bound fixes only the asymptotic shape and is uninformative at the lengths we measure in trained networks; an exactly solvable linear relay model shows that it cannot be improved and accounts for the untrained curves. Cancelling the dilution with a logit bias on the prompt stops the decay early in training, and training suppresses relay by about one billion tokens, before attention sinks form. A toolkit that measures αα for any Hugging Face causal language model is publicly available.

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

本站首次收录为 v2,暂无更早版本可对照。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v2 · 2026-11-17 18:37 · 最新观测版本

A Conditional No-Hair Theorem for Transformers: How Language Models Forget Their Prompt

A black hole retains only a few conserved charges of the matter it formed from. We are inspired to make the analogy and we ask how much an autoregressive transformer retains of its prompt after generating tt tokens. For two prompts followed by the same continuation, the difference of their hidden states decays approximately as a power law t−αt^{-α}, where α=1α=1 is the rate at which a fixed prompt's share of uniform attention shrinks. Beyond the first layers, trained models forget faster (α≈0.55α\approx0.55--0.70.7) than untrained ones of the same architecture (α≈0.2α\approx0.2--0.50.5). For finite-depth pre-LN softmax transformers with bounded attention logits and non-degenerate pre-norm activations, we prove that this offset is O(P(log⁡t)L−1/t)O\big(P(\log t)^{L-1}/t\big) for prompt length PP and depth LL: the prompt's direct influence is diluted as the context grows, and \emph{relay} through continuation tokens adds only logarithmic factors. Conditional on the generated text, the prompt's influence on the next token thus vanishes. The bound fixes only the asymptotic shape and is uninformative at the lengths we measure in trained networks; an exactly solvable linear relay model shows that it cannot be improved and accounts for the untrained curves. Cancelling the dilution with a logit bias on the prompt stops the decay early in training, and training suppresses relay by about one billion tokens, before attention sinks form. A toolkit that measures αα for any Hugging Face causal language model is publicly available.

查看此版本解读 →查看此版本原文 ↗