[HN Gopher] How attention sinks keep language models stable
       ___________________________________________________________________
        
       How attention sinks keep language models stable
        
       Author : pr337h4m
       Score  : 143 points
       Date   : 2025-08-08 08:53 UTC (14 hours ago)
        
 (HTM) web link (hanlab.mit.edu)
 (TXT) w3m dump (hanlab.mit.edu)
        
       | Havoc wrote:
       | > The first few tokens often carried minimal semantic information
       | --sometimes just a start-of-sequence marker or common words like
       | "the" or "a."
       | 
       | I wonder if it makes sense to use the first word as a title of
       | sorts rather than going straight in grammatically correct
       | sentence when prompting
        
         | optimalsolver wrote:
         | "Magnets. How do they work?"
        
           | gjm11 wrote:
           | The heuristic doesn't work quite so well when applied to the
           | actual original version of that line.
        
         | xg15 wrote:
         | Some people start their prompts with "Hello" or "Please" or
         | something similar, out of some habitual sense of politeness, I
         | think. It would be hilarious if those prompts really work
         | better because the model can use those words as attention
         | sinks.
        
           | CamperBob2 wrote:
           | One point that Karpathy has made in some of his videos is
           | that using additional tokens in the prompt can facilitate
           | computation. If you ask a transformer to do some basic math,
           | it will be more likely to get the right answer (or at least a
           | better approximation) with a more verbose prompt. To me, this
           | backs up the use of more conversational language ("Please,"
           | etc.) when prompting.
           | 
           | However, that seems to be contradicted by what was shown
           | recently with the successful International Math Olympiad
           | effort. Their prompts, such as
           | https://github.com/aw31/openai-
           | imo-2025-proofs/blob/main/pro... , were very terse. It's hard
           | to tell where the prompt stops and the CoT response starts,
           | in fact.
           | 
           | So there is probably some interplay between the need for
           | attention sinks and the use of step-by-step reasoning. It
           | might not be too surprising if the latter works because it's
           | an indirect way to optimize the former.
        
             | xg15 wrote:
             | I wonder if the model could also just make its own sink
             | tokens if the prompt doesn't have any. E.g. if the model
             | first emits some "fluff" like "The answer to this question
             | is:" before starting with the actual answer, it could use
             | those tokens as attention sinks. Same with "thinking
             | tokens" that don't directly contribute to the answer or
             | invisible formatting tokens, etc.
        
               | CamperBob2 wrote:
               | True, along with "You're absolutely right! What an
               | insightful observation. You're going places, bro," yadda
               | yadda yadda.
               | 
               | It would be amusing if all that gratuitous sycophancy
               | actually helped with inference accuracy. It would also be
               | worth treating that as a bug to be fixed, of course.
        
             | yorwba wrote:
             | > It's hard to tell where the prompt stops and the CoT
             | response starts, in fact.
             | 
             | That's because you're looking at the final output that
             | includes neither the prompt nor the intermediate chain of
             | thought.
        
               | CamperBob2 wrote:
               | Good point -- I can see that, but it all ends up in the
               | same context, anyway. Point being, the model seems to
               | prefer to conserve tokens.
               | 
               | That said, now I'm wondering if all those dashes it spews
               | out are more than just window dressing.
        
       | Calavar wrote:
       | > Researchers had observed similar patterns in BERT, where "a
       | surprisingly large amount of attention focuses on the delimiter
       | token [SEP] and periods," which they argued was used by the model
       | as a sort of no-op. The same summer at Meta, researchers studying
       | vision transformers found similar behavior, observing that models
       | would repurpose uninformative background patches as computational
       | scratchpads.
       | 
       | This seems to go beyond just transformers. For example, I recall
       | reading a paper a while ago that showed a similar effect in an
       | image to image model with a GAN/U-Net architecture [1].
       | 
       | [1] https://arxiv.org/abs/1712.02950
        
         | SpaceManNabs wrote:
         | I miss GANs. I understand that they are much harder to train
         | than transformers for the same performance even with high data
         | regime and high parameter regime, but there was such good
         | optimization research and tricks that came out of them.
         | 
         | The work on the capacity of discriminators was super cool.
        
       | am17an wrote:
       | This is nice and useful because the new GPT-OSS model uses this
       | technique. Kudos to the original authors!
        
         | diggan wrote:
         | And, as always, the FOSS ecosystem moves quickly, llama.cpp
         | already fully support them! https://github.com/ggml-
         | org/llama.cpp/pull/15157
        
       | esafak wrote:
       | > Barbero et al. have shown that attention sinks serve as
       | "pressure valves" preventing what researchers call "over-mixing"
       | --a pathological state where deep models processing long
       | sequences blur important distinctions between tokens. The
       | presence of a sink draws attention away from other tokens,
       | limiting the spread of information (and noise) and resulting in
       | more stable embeddings.
       | 
       | This sounds like it is working for the wrong reasons. Surely the
       | right behavior is for the _right_ neurons to receive attention
       | rather than the first handful. Jamming everything there is the
       | complementary sin of blurring. I would investigate attention
       | _equalization_ paired with a sparsity prior or something similar
       | to prevent blurring.
        
         | yorwba wrote:
         | The point is that there's not always a right token to attend
         | to. If the information you're looking for is not there, no
         | clever attention scheme will find it. The best you can hope for
         | when that happens is that the value returned in the "not found"
         | case is distinguishable from the "found" case. Having an
         | attention sink serve as a fixed "not found" value is one way to
         | do this.
        
           | esafak wrote:
           | Good point. Does that make them mitigate hallucinations?
        
             | yorwba wrote:
             | In a sense? As the article notes, models trained using
             | standard attention develop attention sinks naturally and
             | removing them makes the model deteriorate completely, so
             | the hallucinations you're thinking of were most likely
             | output by a model that had already mitigated them in this
             | way.
        
       | canjobear wrote:
       | Seems like this was a better solution to the same problem
       | https://www.evanmiller.org/attention-is-off-by-one.html
        
         | markisus wrote:
         | Did this end up working? It sounds plausible but it needs some
         | empirical validation.
        
           | serialx wrote:
           | Yeah, attention sinks were applied to gpt-oss
        
         | danieldk wrote:
         | The attention sink as used in gpt-oss is similar to your link.
         | But rather than adding one to the denominator, they add a
         | trainable 'logit' (a different logit for each head).
        
       | Scene_Cast2 wrote:
       | I found a fairly large improvement in my toy transformer model
       | where I added a "global" token akin to the CLS token in ViT.
       | 
       | Another approach I've seen is the "Diff transformer" from MS
       | Research (https://github.com/microsoft/unilm/tree/master/Diff-
       | Transfor...).
        
       | innerlee wrote:
       | The singular defects (or high-norm tokens) [1] may be related to
       | attention sinks. It is interesting that the direction of all
       | high-norm tokens share the same direction. Maybe the theory
       | behind is not very complex and the issue can be fixed cleverly
       | during training.
       | 
       | [1] https://openreview.net/pdf?id=4yBnUokU2v
        
       | sanj wrote:
       | Is there a way to hint in the prompting what information should
       | be retained in the attention sinks?
        
       ___________________________________________________________________
       (page generated 2025-08-08 23:01 UTC)