Kaley Brauer 💫 (@kaleybrauer): "If you ask a frontier LLM a multi-hop reasoning question, e.g., "Who won the Nobel Prize for Chemistry in (1900 + Mozart's age when he died)?", it usually can't answer correctly immediately (no thinking)

BUT if you ask the same question & append 300 dots, suddenly it can answer?" | XCancel

Kaley Brauer 💫@kaleybrauer

8h

If you ask a frontier LLM a multi-hop reasoning question, e.g., "Who won the Nobel Prize for Chemistry in (1900 + Mozart's age when he died)?", it usually can't answer correctly immediately (no thinking)

BUT if you ask the same question & append 300 dots, suddenly it can answer?

Jul 17, 2026 · 6:31 PM UTC

54

95

1,572

78,497

Kaley Brauer 💫@kaleybrauer

8h

Clearly the models are using the dot tokens to do extra hidden computation (looking up facts, doing math)

We investigated how + what the models are doing and showed the computation is readable from the hidden states over the "meaningless" filler tokens: arxiv.org/abs/2607.03502

Reading Between the Dots: Decoding Hidden Computation across Filler Tokens

Frontier LLMs can perform multi-step reasoning over content-free filler tokens like dots or counting sequences, producing correct answers with no visible chain-of-thought (CoT). This is a limit...

arxiv.org

264

9,163

Kaley Brauer 💫@kaleybrauer

8h

Just presented this at #ICML 2026 Mechanistic Interpretability Workshop in Seoul! So many wonderful people doing fascinating work

112

7,987

Sort replies:<br>Relevant<br>Recent<br>Liked

Alman Gonzaleshvili

@gonzaleshvili

16m

Replying to @kaleybrauer

Excellent paper, this is one of those arguments that starts with a hunch and ends with substantial findings. Thank you, this supports my proposal that semantically/semioticslly each token can carry many more data than just "words" .

54

samlaf

@samlafer

1h

Replying to @kaleybrauer

How is this different from train of thought? Couldn’t the LLM learn to generate its own “dots whiteboard” with optimal length by itself..?

109

Ashraf@Ashraf_Medhat93

1h

Replying to @kaleybrauer

Will it work if you just added empty spaces or underscore instead of dots?

218

1) What@anachronistical

2h

Replying to @kaleybrauer

Emergent pause-token computation?

1,144

Cam Turner@CameronTurner55

5m

Replying to @kaleybrauer @burny_tech

Adding more space to a prompt leaves room for reasoning🤯

Austin Ray

@austospumanto

2m

Replying to @kaleybrauer

Makes sense. Fun!

14

Tonic@decaf100

1h

Replying to @kaleybrauer

I think it’s why reasoning is so powerful, it’s not just literal reasoning but giving itself more tokens to reason through.

145

CK 🏴‍☠️

@cyprianpl

2h

Replying to @kaleybrauer

How many dots until AGI?

21

1,085

Bearuo Bear@BearuoB60982

2m

Replying to @kaleybrauer

Haiku 4.5 (no deep reasoning)

Cam Turner@CameronTurner55

7m

Replying to @kaleybrauer @burny_tech

I love weird quirks like this.

29

Bearuo Bear@BearuoB60982

8m

Replying to @kaleybrauer

When did you last try that claim on the benefit of appending 300 dots? Sonnet 4.6 Medium is not exactly a frontier LLM.

73

Chris Nota

@chris_nota_rl

4h

Replying to @kaleybrauer

The cool thing is you can use this to do “prefill reasoning,” which . The downside is that you lose the recurrence that you get from real tokens.

830

David

@DavidSHolz

4h

Replying to @kaleybrauer

this is fun and i wonder how much of 'reasoning models' get their boost from this effect versus the actual words they say

134

4,853

yungmulababy

@yungmulababylol

2h

Replying to @kaleybrauer @cremieuxrecueil

llm’s have invisible scratch paper

1,481

Kai Sheng Tai@kaishengtai

1h

Replying to @kaleybrauer

Hah, @lateinteraction and I pitched investigating this “thinking with dots” phenomenon as a project idea for the Stanford NLP class in 2019(?). We didn’t get any takers unfortunately. Glad that someone looked into it.

614

RB@530RB

18m

Replying to @kaleybrauer

It’s just another paper on attention sinks. Giving token space allows attention to work as thoughts (adding reasoning steps is just self-created tokens for the same effect). This is well known.

pekora@pekora45402

1h

Replying to @kaleybrauer

woah, that's cool! did you ever test parallelizable otherwise 1-step fact addition questions?

825

Tim Jackowski

@TakaseStudios

4h

Replying to @kaleybrauer

This is going to be fun to dig into - thank you for sharing. My favorite stupid AI trick is whatever the answer I get just say "Really?" ... LOL.

732

Delip Rao e/σ

@deliprao

3h

Replying to @kaleybrauer

Is it only dots? I’ve seen dots in reasoning traces, and also dots are correlated with thinking-related vocab. Wondering if you can reproduce this behavior by considering other commonly occurring ngrams from reasoning traces?

2,265

Patty

@Patty_H93

3h

Replying to @kaleybrauer @cremieuxrecueil

You see stuff like this and realize maybe AGI isn’t as close as we think

624

Clif

@clifcode

4h

Replying to @kaleybrauer...