BERT-style encoders produce one vector per input token. To get a sentence-level vector, you have to combine them, this is **pooling**.
Two common choices are CLS pooling (take the vector at position 0, the special CLS token) and mean pooling (average all token vectors). The question is: which one gives better sentence embeddings for tasks like semantic search?
The empirical answer is **mean pooling**, established by the SBERT paper in 2019. They measured both on STS (Semantic Textual Similarity) benchmarks and mean pool consistently beat CLS by 3 to 7 points.
The reason is about what BERT was trained to do. BERT's pretraining task is masked language modeling, predicting masked words from context. That task does not specifically train the CLS token to be a sentence summary. CLS just happens to be the position whose output BERT's classifier head reads after fine-tuning.
Using raw BERT's CLS as a sentence embedding asks the model for behavior it was never trained for. Mean pool, in contrast, averages over tokens that WERE trained to encode their bidirectional context, which is a sounder basis for sentence representation.
Most modern open-weight embedders (BGE, E5, mxbai-embed, Snowflake Arctic) use mean pool by default.
The two strategies differ in what they extract from the encoder output and, more importantly, in whether the training objective justifies that extraction.
**Mean pool** averages all token outputs (with attention mask) into a single vector:
$$v_{\text{mean}} = \frac{1}{|\text{mask}|} \sum_i \text{mask}_i \cdot h_i$$
Every position contributes equally; the sentence vector reflects collective token semantics.
**CLS pool** takes only the vector at position 0:
$$v_{\text{CLS}} = h_0$$
This works in BERT as classifier setups because fine-tuning with a classification head drives summarization signal into CLS for that specific task. It does not work well for raw pretrained BERT used in cosine-similarity retrieval, because MLM pretraining never trained CLS to encode sentence meaning.
SBERT (Reimers and Gurevych, 2019) measured the gap directly on STS benchmarks and found mean pool beat CLS by 3 to 7 points across BERT variants. The result has been replicated and the pattern stuck: BGE, E5, multilingual-e5-large, mxbai-embed-large-v1, Snowflake Arctic Embed L 2.0, and Nomic Embed v1.5 all use mean pool.
The distractor options miss the architecture. Mean pool is NOT only for fine-tuning, it is used at inference too. Vectors are NOT identical for short inputs; the two pooling strategies project to different points in the same encoder's output space. CLS is NOT universally preferred, the opposite is true for sentence similarity.
Quick reference
TL;DR
Mean pooling wins for sentence similarity because BERT's MLM pretraining does not train CLS to summarize sentence meaning; SBERT measured the gap at 3 to 7 STS points.
Imagine you ask 32 jurors to vote on a verdict (mean pool) versus asking only one juror who happened to sit in seat 1 and never had the case explained to them (CLS pool). The first gives you the collective opinion; the second gives you one person's underprepared guess. Mean pool wins because every token in BERT was trained to predict its context, so averaging them collects sentence-wide signal. CLS in raw BERT was never specifically trained to summarize, so leaning on it alone is asking for a job the model was not assigned.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Using vanilla pretrained BERT CLS as a sentence embedding for cosine-similarity retrieval MLM does not train CLS to summarize sentence meaning, you get measurably worse retrieval than mean pool on the same encoder.
Switching pooling strategies mid-pipeline (index built with mean pool, query encoded with CLS) The two strategies project to different geometric subspaces of the encoder output; cosine similarity across them is meaningless.
Applying mean pool to a model whose card explicitly specifies a different pooling strategy The model was contrastive-trained with a specific pooling head; using a different one at inference breaks the training contract and degrades quality.
Quick scan · 2–4 min
Everything important, quickly.
6 to 8 min: mean vs CLS definitions + MLM-mismatch argument + SBERT empirical gap + why every modern encoder embedder uses mean + when CLS is acceptable.
Real-world examples
Real products, models, and research that use this idea.
1SBERT (2019) measured CLS vs mean on STS-B and showed mean beats CLS by 3 to 7 points across BERT variants.
2BGE-large-en-v1.5 and BGE-M3 from BAAI use mean pool as the explicit default.
3E5-large-v2 and multilingual-e5-large from Microsoft use mean pool, the dominant 2026 multilingual choice.
4mxbai-embed-large-v1 from Mixedbread and Snowflake Arctic Embed L 2.0 both default to mean pool.
5Nomic Embed v1.5 uses mean pool with Matryoshka representation learning on top, MRL is orthogonal to the pooling decision.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QCould you fine-tune BERT so CLS becomes a strong sentence embedding?
Sign in or go Pro to see the approach for this follow-up.
QWhy does the gap between mean and CLS narrow after large-scale contrastive fine-tuning?
Sign in or go Pro to see the approach for this follow-up.
QHow does max pool fit into the mean vs CLS comparison?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Picking CLS because BERT papers use CLS for classification. Classification is not similarity; SBERT specifically showed mean beats CLS for STS by 3 to 7 points.