arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2309.13638cs.CLcs.AI
收藏

自回归的余烬:通过大语言模型被训练去解决的问题来理解它们

Embers of Autoregression: Understanding Large Language Models Through the Problem They are Trained to Solve

Princeton University(普林斯顿大学)

R. Thomas McCoy, Shunyu Yao, Dan Friedman, Matthew Hardy, Thomas L. Griffiths

AI总结:

本文从互联网文本下一词预测这一训练目标出发,提出影响 LLM 准确率的任务、输出和输入概率三因素,并通过 GPT-3.5/GPT-4 的十一项任务验证低概率会导致显著失败。

摘要:

大语言模型(LLM)的广泛采用,使得认识其优势与局限变得重要。我们主张,为了形成对这些系统的全面理解,需要考察它们被训练去解决的问题:在互联网文本上进行下一个词预测。通过认识该任务所施加的压力,我们可以预测 LLM 将采取的策略,从而推断它们何时会成功或失败。这种方法——我们称之为目的论方法——使我们识别出三个假设会影响 LLM 准确率的因素:待执行任务的概率、目标输出的概率,以及所提供输入的概率。我们预测,当这些概率较高时,LLM 将比较低时获得更高准确率,即使是在概率本不应产生影响的确定性环境中也是如此。为检验这些预测,我们在十一项任务上评估了两个 LLM(GPT-3.5 和 GPT-4),并发现有力证据表明 LLM 会以我们所假设的方式受到概率影响。在许多情况下,实验揭示了令人惊讶的失败模式。例如,当输出是高概率词序列时,GPT-4 解码一种简单密码的准确率为 51%,而当输出为低概率时仅为 13%。这些结果表明,AI 从业者在低概率情形中使用 LLM 时应保持谨慎。更广泛地说,我们的结论是,不应像评估人类那样评估 LLM,而应将其视为一种独特类型的系统——一个由其自身特定一组压力塑造而成的系统。

英文摘要:

The widespread adoption of large language models (LLMs) makes it important to recognize their strengths and limitations. We argue that in order to develop a holistic understanding of these systems we need to consider the problem that they were trained to solve: next-word prediction over Internet text. By recognizing the pressures that this task exerts we can make predictions about the strategies that LLMs will adopt, allowing us to reason about when they will succeed or fail. This approach - which we call the teleological approach - leads us to identify three factors that we hypothesize will influence LLM accuracy: the probability of the task to be performed, the probability of the target output, and the probability of the provided input. We predict that LLMs will achieve higher accuracy when these probabilities are high than when they are low - even in deterministic settings where probability should not matter. To test our predictions, we evaluate two LLMs (GPT-3.5 and GPT-4) on eleven tasks, and we find robust evidence that LLMs are influenced by probability in the ways that we have hypothesized. In many cases, the experiments reveal surprising failure modes. For instance, GPT-4's accuracy at decoding a simple cipher is 51% when the output is a high-probability word sequence but only 13% when it is low-probability. These results show that AI practitioners should be careful about using LLMs in low-probability situations. More broadly, we conclude that we should not evaluate LLMs as if they are humans but should instead treat them as a distinct type of system - one that has been shaped by its own particular set of pressures.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑