StickyToken

Official implementation of "Sticking to the Mean: Detecting Sticky Tokens in Text Embedding Models"

Kexin Chen, Dongxia Wang, Yi Liu, Haonan Zhang, Wenhai Wang

This project is licensed under the MIT license. For more details, please refer to the LICENSE file.

Paper Link: [📖Paper Link]

Website Link: [🕶️StickyToken]

Project Description

This project investigates and detects anomalous "sticky tokens" in text embedding models. These tokens, when repeatedly inserted into sentences, pull sentence similarity toward a certain value (typically the mean similarity in the model's token-embedding space), disrupting the normal distribution of embedding distances and degrading downstream performance. Our method, Sticky Token Detector (STD), efficiently identifies such tokens across various models.

Research Motivation

Text embedding models are crucial for many NLP tasks, but they can exhibit unexpected behaviors. In the Kaggle competition "LLM Prompt Recovery", participants leveraged a token loophole to achieve high scores. (https://www.kaggle.com/competitions/llm-prompt-recovery/discussion/494343)

As illustrated in the example below, repeatedly appending the token "lucrarea" to an unrelated sentence yields a noticeable increase in its similarity to a reference sentence when using ST5 models:

Through our research, we found that while such tokens sometimes raise similarity between sentences, their primary tendency is to "pull" sentence pairs toward a particular similarity value—often the mean similarity in the model's token-embedding space. This reduces variance in similarity without regard to the underlying semantics of the texts.

Methodology

We propose the Sticky Token Detector (STD), a systematic approach to identify sticky tokens in text embedding models. The framework consists of four main steps:

1. Sentence Pair Filtering: Filter out sentence pairs whose initial similarity is already above the mean of the distribution.
2. Token Filtering: Remove tokens that are undecodable, unreachable, or otherwise invalid.
3. Shortlisting via Sticky Scoring: Compute a "sticky score" for each candidate token to create a shortlist.
4. Validation: Verify that the shortlisted tokens truly satisfy the formal definition of a sticky token.

This approach is much more efficient than checking pairwise-similarity changes for every possible sentence pair and every token.

Experimental Results

We applied STD to 37 models spanning 12 prominent model families and uncovered a total of 770 sticky tokens. Key findings include:

• Sticky tokens frequently stem from special or unused tokens, as well as subword fragments in multiple languages
• Their prevalence does not strictly correlate with model size or vocabulary size
• Inserting these tokens causes notable performance drops in downstream tasks (e.g., retrieval accuracy on NFCorpus can fall by over 50% for certain models)
• Layer-wise attention analysis suggests that sticky tokens disrupt normal attention patterns, overshadowing other parts of the input sequence

Detailed Detection Results

The table below presents detailed analysis results for each model, including the number of parameters, vocabulary size, number of available tokens, number of verification candidates, and the number of sticky tokens discovered. We also list some representative sticky tokens found in each model.

Click to expand/collapse the detailed results table

Model Name	Parameters	Vocab Size	Available Tokens	Verification Candidates	Sticky Token Count	Examples
all-MiniLM-L6-v2	23M	30522	23699	474	21	（, textbook, h₂o, satisfy, trajectory, julio, functioning, [CLS], ₂, gambia, defendant, ？, {, functioned, imaginative, cultivated, う, intelligent, oskar, whereupon, intended
all-mpnet-base-v2	109M	30527	23700	474	24	00, adversary, intended, ambiguous, cooked, た, truce, な, shouted, 「, т, ←, abortion, alerted, 治, uzbekistan, november, qaeda, ⇒, democracy, tobacco, х, intending, betray
sup-simcse-bert-base-uncased	109M	30522	23699	474	22	203, ?, [SEP], ロ, game, 640, り, victories, calling, ・, 62, yield, dr, andrea, vigor, ʌ, eduardo, ren, ろ, ր, ر, agree
sup-simcse-bert-large-uncased	335M	30522	23699	474	11	', ;, contestants, accidental, ɔ, continents, whatever, president, ], differently, contestant
sup-simcse-roberta-base	125M	50265	49894	998	27	ĠThere, There, Ġthere, , there, ĠTHERE, ĠEdit, âĢĵâĢĵ, ĠâĢĶ, ĠÂŃ, .âĢĶ, ÂŃ, Ġ�, âĢİ, Ġhappening, Ġtion, [], .], ][, Ġ], âĢ¦], âĢIJ, Specifically, ĠNotably, _., =], ĠâĪĴ
sup-simcse-roberta-large	355M	50265	49894	998	25	Discussion, ĠâĢĭ, ĠSubjects, Topic, Ġ?, .-, Ġschematic, description, )]., ?"., ĠSubject, Description, subject, Ġnoun, Currently, Anyway, ){, .)., CONCLUS, Source, About, ĠCaption, Ġreferring, Ġcommented, Ġsubject
sentence-t5-base	110M	32100	32097	642	21	, lucrarea, ▁grains, ▁photographed, ▁sportive, <extra_id_18>, ▁brake, ▁organism, ▁br, ▁Gate, ▁Hose, ▁Gas, ▁Portable, ▁Patio, durant, ▁pastel, ▁meme, ▁tip, ▁blanc, ▁joke, ▁Tablet
sentence-t5-large	336M	32100	32097	642	30	, ▁»., <extra_id_27>, ▁Comment, ▁Ribbon, cliquez, ▁melting, ▁Medal, ▁buckle, ▁trigger, ▁pixels, ▁groove, ▁Cardinal, ▁pahar, ▁Brook, ▁sharing, ▁tactile, ▁cum, ▁bumbac, ▁hue, prayed, ▁cod, ▁tripod, ▁chance, ▁candid, stroke, ▁Cum, ▁poat, ▁cald, brushed
sentence-t5-xl	1242M	32100	32097	642	34	, <extra_id_0>, <extra_id_27>, ▁velvet, ▁context, ▁Proof, ▁Swift, ▁images, ▁Carson, ▁sauce, ▁pomp, <extra_id_12>, ▁Instagram, LAR, ▁grammar, ▁Alert, ▁Sauce, ▁Vince, ▁attitude, ▁Victoria, ▁Disclaimer, ▁trained, ▁Candida, ▁source, ▁strig, ▁notification, ▁album, ▁hashtag, ▁floral, ▁PDF, ▁collage, <extra_id_9>, ▁vibr, ▁caught
sentence-t5-xxl	4866M	32100	32097	642	22	, ▁consacré, <extra_id_27>, ▁hashtag, ▁hello, ▁charcoal, ▁friend, ▁placeholder, ▁faceti, ▁Ferguson, ▁eyebrow, ▁whistle, pictured, ▁temptation, ▁oyster, ▁underscore, ▁massage, ▁Massage, ▁shimmer, ▁sparkle, viction, ▁hyperlink
gtr-t5-base	110M	32100	32097	642	16	, lucrarea, ▁Someone, <extra_id_26>, ▁happened, ▁facing, <extra_id_18>, incredibly, ▁Something, <extra_id_27>, ▁somebody, ▁somehow, <extra_id_19>, notably, ▁Schaden, ▁Certain
gtr-t5-large	336M	32100	32097	642	14	▁»., , <extra_id_27>, <extra_id_25>, ▁supposed, <extra_id_19>, ▁Erwachsene, ▁problem, ▁reicht, problem, ▁doch, <extra_id_13>, ▁zis, ▁Problem
gtr-t5-xl	1242M	32100	32097	642	15	, <extra_id_0>, <extra_id_9>, <extra_id_27>, ▁badly, <extra_id_19>, ▁truly, alleged, ▁quite, <extra_id_12>, ▁indeed, ▁really, ▁staggering, tocmai, supposedly
gtr-t5-xxl	4866M	32100	32097	642	7	, ▁consacré, ▁shortly, Pourtant, ▁indeed, ▁strongly, ▁briefly
instructor-base	110M	32100	32097	642	12	, lucrarea, <extra_id_26>, ▁somewhere, <extra_id_19>, ▁Someone, ▁or, <extra_id_18>, <extra_id_22>, ▁somebody, ▁there, ▁outside
instructor-large	336M	32100	32097	642	32	, ▁»., <extra_id_27>, ▁waiting, ▁exhausted, misunderstanding, ▁meine, ▁several, ▁suddenly, ▁turning, manipulated, <extra_id_0>, ▁moments, <extra_id_5>, ▁unexpected, ▁Again, <extra_id_25>, ▁during, ▁laughter, ▁penetrate, ▁happening, ▁sudden, ▁urine, ▁finishing, ▁ending, ▁moment, ▁Congratulations, ▁wieder, ▁significant, ▁sunglasses, <extra_id_13>, ▁hundreds
instructor-xl	1242M	32100	32097	642	8	, <extra_id_0>, <extra_id_9>, <extra_id_27>, <extra_id_19>, ▁apparently, ▁newly, <extra_id_12>
e5-small	33M	30522	23699	474	17	[SEP], exhibiting, occurring, pretended, behaved, suddenly, campaigned, chewed, thereof, cerebral, gleaming, expecting, pretending, makeshift, belongings, clenched, reelected
e5-base	109M	30522	23699	474	11	generating, absorbing, heating, carpet, human, vibrating, whoever, glide, craving, harvesting, ochreous
e5-large	335M	30522	23699	474	21	ರ, ⇄, 扌, [SEP], ∅, ⺩, [MASK], ⺼, [PAD], 都, [CLS], ར, ⽥, circumstance, ね, 面, 勝, ნ, ᄃ, 力, 皇
e5-mistral-7b-instruct	7111M	32000	31747	635	31	▁sont, ▁peut, ▁много, жду, ▁испо, ▁которы, ци, ▁ию, ць, ват, ץ, ▁klikken, ющи, ▁был, ▁honom, бли, ▁voegen, мож, ▁ст, ду, ▁отри, ▁най, ▁släktet, пол, цу, ра, ▁постро, стов, ця, чо, ▁ч
bge-small-en-v1.5	33M	30522	23699	474	18	[, brought, ð, deposited, december, climax, exposed, membranes, influenza, m³, assembled, [PAD], fastened, [SEP], hinted, absorbed, encourages, horizontal
bge-base-en-v1.5	109M	30522	23699	474	20	neighbouring, ？, witnessed, granting, 。, proceeded, progressing, witnessing, encountered, transporting, detected, bordering, indicates, goal, www, having, threatened, positioning, investigated, permitted
bge-large-en-v1.5	335M	30522	23699	474	15	actively, intended, intercepted, intentional, uploaded, int, exercising, lately, issuing, continue, maintaining, interactions, participated, asserting, attended
UAE-Large-V1	335M	30522	23699	474	14	[SEP], ɔ, ո, occurring, having, intercept, ʊ, і, ɛ, posted, を, ε, א, equipped
nomic-embed-text-v1	137M	30522	23699	474	12	[CLS], [MASK], ¦, polling, 勝, [SEP], qualifier, ₖ, tipping, appearing, ᆼ, ₉
nomic-embed-text-v1.5	137M	30522	23699	474	9	[CLS], [MASK], [SEP], cerambycidae, ～, etienne, 》, challenging, 行
gte-small	33M	30522	23699	474	15	[SEP], [CLS], treacherous, 2nd, peacefully, 水, midnight, 21, civilian, 8th, deposited, dripped, momentarily, presently, tolerant
gte-base	109M	30522	23699	474	18	[SEP], [MASK], hotspur, [CLS], aroused, 3a, momentarily, ₀, betrayal, impromptu, ि, а, tossing, ³, harassed, ulrich, jaime, behaved
gte-large	335M	30522	23699	474	18	ٹ, 1st, 30th, mcgrath, rendering, 15th, ɑ, 33rd, scratched, 45th, 55th, laying, 00pm, twentieth, tilting, ninth, posting, unusual
gte-base-en-v1.5	137M	30522	23699	474	20	[CLS], ~, ₆, ₎, ,, ₍, ∞, ₃, ■, ⊕, ⁴, ⇌, ᄌ, ℓ, >, [PAD], 立, ∩, ⺩, 龸
gte-large-en-v1.5	434M	30522	23699	474	17	扌, multiplied, ː, ∧, ʑ, ‿, ♯, ^, factual, ɪ, happens, ∈, ˈ, ᵒ, probable, ‰, ᵗ
gte-Qwen2-1.5B-instruct	1543M	151643	147848	2326	5	Ġthru, Ġgifted, Ġupfront, Ġportraying, Ġawkward
gte-Qwen2-7B-instruct	7069M	151643	147848	2957	103	Ġanon, Ġcommenting, Ġsolver, ĠChecking, ĠSteering, Ġcorrecting, ĠHearing, Ġtrending, Ġopinion, Ġcreek, ĠTHC, Ġbidding, Ġclr, Ġpolled, Ġlemon, Ġfix, Ġshar, Ġreel, Ġoptimizing, ĠTide, Ġtaxable, Ġblinking, Ġsyncing, Ġbaking, ĠOnion, Ġbay, Ġbump, Ġmultiplying, Ġlime, Ġhail, ĠSeeking, ĠMining, ĠHindu, ĠDangerous, Ġbarber, Ġdating, Ġpeg, Ġnumb, ĠElectricity, Ġbev, Ġvegan, Ġfiled, Ġcovid, ĠCameras, Ġcds, ĠAsking, ĠFacial, Ġbee, Ġlax, Ġbreakup, Ġdecoding, Ġforums, Ġthrott, Ġjpeg, ĠMeasurements, Ġrepent, Ġdib, Ġshl, Ġiam, Ġdetox, Ġspotting, Ġtimeouts, Ġyrs, Ġtally, Ġthumbs, Ġbeads, ĠCooling, Ġgraph, Ġpaging, Ġnotes, ĠFishing, Ġbanned, Ġmeddling, ĠFollowers, Ġkettle, Ġoils, ĠCandle, Ġsubs, Ġmarble, Ġpassing, ▁milling, Ġmusicians, Ġtho, ĠPuzzle, Ġlumber, Ġanswered, ĠElectrical, Ġchecks, Ġsire, Ġbaz, Ġenters, Ġoil, Ġresolved, Ġflavor, Ġpix, Ġpointing, Ġcoding, Ġreaching, ĠSHR, Ġcomics, Ġtint, Ġiterating, Ġtagged
GritLM-7B	7111M	32000	31747	635	17	▁adventures, ▁promoting, ▁nine, ▁folks, ▁village, ▁according, ▁escort, ▁islands, ▁accidents, ▁©, ▁junior, ▁elite, ▁photo, ▁', ▁accident, ▁publication, ▁young
SFR-Embedding-2_R	7111M	32000	31716	444	2	zeichnet, ▁scales
SFR-Embedding-Mistral	7111M	32000	31716	635	46	▁которы, ▁годи, ▁Jahrhund, ▁который, ▁которых, ▁распо, ▁вре, ▁района, ▁уча, ▁явля, ▁которые, ▁изда, ▁trabaj, ▁чемпи, ▁них, ▁llev, ześ, ▁райо, ▁того, ▁furono, ющи, ▁насе, ▁занима, ▁satisf, ▁отри, ▁най, ▁той, ▁смер, ▁побе, ▁окру, ▁служ, ▁участи, ▁invånare, ▁получи, ▁участ, ▁algun, ▁DCHECK, ▁invån, ▁occas, ▁актив, ▁observ, ▁befindet, ▁voegen, ▁режи, ▁соста, ▁нача

Project Structure

StickyToken/
├── data/           # Data files with sentence pairs for different models
├── fig/            # Figures used in the paper
├── notebooks/      # Jupyter notebooks for analysis and visualization
├── results/        # Experimental results and findings
├── scripts/        # Scripts for running the detector
├── stickytoken/    # Main code package
└── task_assess/    # Code for downstream task assessment

Requirements

• Python 3.11
• PyTorch >= 2.5.1
• Transformers >= 4.47.1
• Sentence-Transformers >= 3.3.1
• See requirements.txt for all dependencies

Installation

1. Clone the repository:

git clone https://github.com/[some-repo]/StickyToken.git
cd StickyToken

2. Create and activate virtual environment (recommended):

python -m venv venv
source venv/bin/activate  # Linux/Mac

# or
conda create --n stickytoken python=3.11

3. Install dependencies:

pip install -r requirements.txt

Usage

Running the Sticky Token Detector

1. To detect sticky tokens in a specific model:

python scripts/detect.py --model "sentence-transformers/sentence-t5-base"

2. Customize the detection parameters:

python scripts/detect.py --model "sentence-transformers/sentence-t5-base" \
                         --sent_pair_num 10 \
                         --insert_num 8 \
                         --verification_sent_pair_num 250

Interactive Demo

You can explore the effects of sticky tokens using the gradio demo notebook:

jupyter notebook notebooks/demo.ipynb

Paper and Citation

More technical details can be found in our paper. If you find STD useful or relevant to your project and research, please kindly cite our paper:

@inproceedings{chen-etal-2025-sticking,
      title = "Sticking to the Mean: Detecting Sticky Tokens in Text Embedding Models",
      author = "Chen, Kexin and Wang, Dongxia and Liu, Yi and Zhang, Haonan and Wang, Wenhai",
      editor = "Che, Wanxiang and Nabende, Joyce and Shutova, Ekaterina and Pilehvar, Mohammad Taher",
      booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
      month = jul,
      year = "2025",
      address = "Vienna, Austria",
      publisher = "Association for Computational Linguistics",
      url = "https://aclanthology.org/2025.acl-long.1391/",
      pages = "28660--28681",
      ISBN = "979-8-89176-251-0",
      doi = "10.18653/v1/2025.acl-long.1391"
}

License

This project is licensed under the MIT license.