一句话结果:让模型只看最近 4 步,每步正确率从 35.0% 提到 66.8%,长时程衰减完全消失——而且比"把历史里所有错误都改成正确答案"(现实中做不到的上限)还要好。
论文《The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs》(arXiv:2509.09677)提出 self-conditioning:当上下文里出现模型自己犯过的错,它后续出错的概率会上升;每步错误率随任务推进而升高,而且把模型做大并不能消除这个效应。
我们注意到一件事:在那个设定里,"上下文里有自己的错"和"上下文变长"是绑在一起的,没有被拆开。所以本实验把它们做成一个 2×2,并且只测不需要正确答案、现在就能部署的干预。
给定一张 40 项的字典 k00..k39 → 两位数,模型逐步执行 100 步:每步取出指定 key 的值,加到当前累计和上,只输出新的累计和。每步是否正确可机器判定;唯一的难度是稳定地执行很多步。100 条独立序列,同一模型(Qwen2.5-7B-Instruct),贪心解码。
| 喂给模型的上下文 | 平均每步正确率 | 95% CI | 前20步 → 后20步 |
|---|---|---|---|
| 完整历史(基线) | 35.0% | [33.5, 36.5] | 71.1% → 11.0%(−60.1) |
| 完整历史 + 显式复述当前状态 | 26.9% | [25.9, 27.8] | 56.8% → 6.8%(−49.9) |
| 只保留最近 4 步 | 66.8% | [65.9, 67.8] | 64.6% → 68.4%(+3.8) |
| 只保留最近 4 步 + 显式复述状态 | 57.6% | [56.6, 58.6] | 60.6% → 57.1%(−3.5) |
| 以下两行需要"上帝视角"(把历史里的值替换成正确答案),现实中做不到,仅作参照上限 | |||
| 完整历史 + 全部改正 | 58.0% | [57.1, 58.9] | 74.7% → 46.1%(−28.6) |
| 最近 4 步 + 全部改正 | 59.5% | [58.5, 60.6] | 62.2% → 58.5%(−3.6) |
horizon_stats.json。Starting from self-conditioning (arXiv:2509.09677), we disentangled two factors that were confounded in the original setting — presence of the model's own errors in context vs context length — using a 2×2 over deployable interventions only (no oracle). On a 100-step running-sum execution task with Qwen2.5-7B-Instruct (100 sequences, greedy): self-conditioning replicates in long context (removing one's own errors: +31.5 pts of degradation, significant), but vanishes once the context is truncated (−0.2 pts, n.s.). Truncating to the last 4 steps raises mean per-step accuracy from 35.0% to 66.8% and removes degradation entirely (+3.8 pts across the run) — beating the oracle condition (58.0%) that replaces all history with ground truth. Counterintuitively, adding an explicit state restatement hurts when the context is already short (−7.2 pts, significant). Main caveat: the task is Markovian, so truncation discards nothing essential; results may not transfer to tasks requiring long-range recall. All code and per-step raw outputs are published above, including the statistics from our own initially-broken metric.
你也有想验证的方向?一句话提交,免费层直接跑:证据地图、查新 idea、实验方案、初步信号。
提交你的研究方向 →