Three pooling strategies are standard in 2026 production embedding models:
**Mean pooling** is the SBERT-family default. You average the token-level output vectors (with attention mask) into one sentence vector. Used by BGE, E5, mxbai, Snowflake Arctic Embed, and most encoder-based embedders.
**Last-token pooling** is the default for decoder-only LLM-based embedders like e5-mistral and NV-Embed. Because decoder-only models use causal masking, only the last token has seen the full input context, so taking its vector gives you a sensible sentence representation.
**Attention pooling with a learned query token** is the third option, used by models like Jina v3. A small learned component computes attention weights over token outputs and produces a weighted sum. Slight quality lift over mean pool at the cost of one extra trained component.
The other options are wrong. CLS pooling underperforms for sentence similarity (BERT's MLM does not train CLS for that). Sum without normalization makes vector magnitude depend on input length, which breaks cosine similarity. Random-token pooling is not a real technique.
Three pooling strategies survive as standards: **mean pool**, last-token pool, and learned attention pool. Each fits a different model family.
**Mean pool** is the encoder-only default. SBERT measured CLS vs mean vs max in 2019 and mean won by 3 to 7 STS points. BGE, E5, mxbai-embed, Snowflake Arctic, Nomic Embed all use mean pool with the attention mask:
$$v = \frac{\sum_i m_i \cdot h_i}{\sum_i m_i}$$
**Last-token pool** is the decoder-only default. Under causal masking, token 0 has seen only itself and token N has seen the entire input. Averaging them is incoherent, so decoder-only embedders (e5-mistral, NV-Embed, Linq-Embed-Mistral) take the final position's hidden state instead, often with an instruction prefix to shape what the last token summarizes.
**Learned attention pool** introduces a learned query vector that attends over the token outputs. Jina v3 and some Cohere variants use this for a 1 to 2 percent retrieval-benchmark lift over mean pool. Cost is one extra trained component.
The three rejected options are wrong for different reasons. CLS pool is the BERT-classifier default and underperforms for sentence similarity. Sum without normalization makes the vector magnitude scale with sequence length, breaking cosine similarity assumptions. Random-token pooling is not a real strategy and was likely included as a distractor.
Quick reference
TL;DR
Standard pooling choices in 2026 are mean (encoder default), last-token (decoder-only embedders), and learned attention pool; CLS is fine for classification but not for sentence similarity.
Imagine a school class of 32 students and you need one report card that summarizes the class. Mean pool: average everyone's grades. Last-token pool: ask the last student, who has heard everyone else (only works in classes where students speak in order). Attention pool: ask a teacher to weight each student by importance. CLS pool: ask the kid who sits in seat 1 (they were not specifically asked to summarize, so the answer is patchy). Sum without normalization: add the grades but do not divide, so a class of 40 looks twice as smart as a class of 20. Random pick: pick a random student. Only the first three give sensible report cards.
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.
Trying to apply mean pool to a decoder-only LLM-based embedder Causal masking makes early-position outputs incomparable to late-position outputs; averaging produces a distorted representation. Use last-token instead.
Trying to apply last-token pool to an encoder-only model like BERT or RoBERTa Encoders use bidirectional attention, every position sees the full context, last-token has no architectural privilege and you lose the averaging signal.
Switching pooling between indexing and query time Vectors live in pooling-specific geometric spaces; a mean-pooled index and a CLS-pooled query produce nonsense cosine similarities.
Quick scan · 2–4 min
Everything important, quickly.
6 to 8 min: three standard pools by architecture + three rejects by failure mode + decoder-only causal-mask argument + attention-pool lift + production gotchas.
Real-world examples
Real products, models, and research that use this idea.
1BGE-M3 from BAAI uses mean pool and is the dominant open-weight multilingual embedder in 2026.
2e5-mistral-7b-instruct (Microsoft, 2024) introduced the decoder-only last token with instruction pattern at scale.
3NV-Embed-v2 from NVIDIA uses last-token pool and tops MTEB leaderboards as of late 2025.
4Jina v3 uses learned attention pooling and ships task-specific LoRA heads on top.
5Snowflake Arctic Embed L 2.0 and mxbai-embed-large-v1 both default to mean pool, consistent with the SBERT lineage.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QWhen would you train a custom attention pooling head on top of an existing encoder?
Sign in or go Pro to see the approach for this follow-up.
QWhy do decoder-only embedders use instruction prefixes like 'Instruct: retrieve relevant passages'?
Sign in or go Pro to see the approach for this follow-up.
QWhat goes wrong if you concatenate mean-pooled and CLS-pooled vectors as a richer embedding?
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 as a standard sentence-embedding pool. It is the BERT-classifier default, not the SBERT-lineage default. SBERT explicitly measured and rejected it for sentence similarity.