“Related? "Let's Think Dot by Dot: Hidden Computation in Transformer Language Models" Let's Think Dot by Dot”
Let's Think Dot by Dot: Hidden Computation in Transformer Language Models
无意义的填充 token 也能替代思维链,带来额外计算
论文问思维链带来的提升,究竟来自类似人类的任务分解,还是来自额外 token 提供的更多计算。作者表明 Transformer 可以用无意义的填充 token(如“......”)代替思维链,解决两个没有中间 token 时解不了的困难算法任务,但学会使用填充 token 很困难,需要特定的密集监督才能收敛。论文还用一阶公式的量词深度刻画了填充 token 有用的问题类别,并指出这引发对 LLM 进行不可审计、与可见思维链脱节的隐藏计算的担忧。
评论者把它当作“思维链不必有语义内容,填充 token 甚至只是点也能带来额外隐藏计算”的出处,用来讨论推理轨迹是否反映模型内部真实的计算。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“let it even be dots”
“See for example, Let's Think Dot by Dot.”
“Let's Think Dot by Dot: Hidden Computation in Transformer Language Models Let's Think Dot by Dot”
“which found that you can get many of the benefits of chain-of-thought simply by adding filler tokens”
“[1] Let's Think Dot by Dot: Hidden Computation in Transformer Language Models: Let's Think Dot by Dot”
看全部 16 条 →
“There was a paper[1] from last year where the authors discovered getting the model to output anything during times of uncertainty, improved the generations overall.”
“"Let's Think Dot by Dot: Hidden Computation in Transformer Language Models" - Let's Think Dot by Dot”
“There are several such papers, off the top of my head one is Let's Think Dot by Dot”
“a 2024 paper noted that simply adding dots to the output allowed some models to perform more latent computation resulting in higher-skill answers.”
“Chain of thought does not need to be semantically useful to improve LLM performance. Let's Think Dot by Dot”
“I think these pretty much map out a little chronology of the research:”
“But at least to some extent it's just model taking time thinking through the problem, and the trace produced doesn't have to be representative of what happens internally: Let's Think Dot by Dot”
“even if you make thinking tokens literally just '....' (absolutely meaningless; zero information), you still see significant performance improvements: Let's Think Dot by Dot and [OTHER_PAPER] for some starters.”
“I think this is a known phenomenon: even in non-reasoning models, adding useless/filler tokens before an answer improves task performance. The model is doing some computation during the filler. See: Let's Think Dot by Dot”
“this seems supported by the abstract of the paper cited in the above article:”
- Let's Think Dot by Dot: Hidden Computation in Transformer Language Models ↗Jimmc414 · 2024-04-27 · 159 分