Atomic Item IDs

Definition

Atomic Item ID

An atomic item ID assigns one unique, indivisible token per catalogue item: each item $i\in \mathcal{I}$ gets its own integer index and its own learned embedding row. It is the simplest item-tokenization scheme (level L1 of the tokenization ladder) and the implicit choice in classical sequential recommenders. In generative recommendation it is the special case of a semantic ID with length $L=1$ and codebook = the entire catalogue. The tokens are arbitrary: item_3487 carries no information about what the item is, and similar items share no structure.

Intuition

One token per item — simple, but it does not scale or generalize

Atomic IDs are the “natural” identifier: like giving every product a unique barcode. They are trivial to look up and a recommendation is exactly one autoregressive step. But two problems follow directly from “one token per item”:

1. The vocabulary equals the catalogue. A platform with $10^6$ items needs $10^6$ output tokens (and a softmax over all of them); $10^9$ items is intractable. Capacity is tied 1-to-1 to vocabulary size.
2. No shared structure. Inception → item_3487 and Interstellar → item_124 are unrelated symbols even though the films are similar. There is no reusable “subword” the model can generalize over, so a brand-new item needs a brand-new token and a freshly trained embedding before it can ever be recommended — strict cold start.

This is precisely the pain that semantic IDs (a few shared codebook tokens per item) were designed to remove.

Mathematical Formulation

In classical sequential recommendation each item maps to a single token with a learned embedding, and the next item is produced by scoring, not generating. Given a SASRec-style history encoding $F_t^{(b)}$ and the shared item embedding table $M$ (one row $M_i$ per atomic ID), the score and the next-item objective are:

$r_{i,t}=F_t^{(b)}{M}_i^{\top },p(i_{t+1}\mid i_1,\dots ,i_t)={\mathrm{softmax}}_{i\in \mathcal{I}}\left(F_t^{(b)}{M}_i^{\top }\right)$

where:

• $i_j$ — the $j$-th interacted item, encoded as a single atomic token $⟨i_j⟩$
• $M\in {\mathbb{R}}^{|\mathcal{I}|\times d}$ — embedding table; grows linearly with the catalogue size $|\mathcal{I}|$
• $F_t^{(b)}{M}_i^{\top }$ — inner-product score of candidate $i$ against the encoded history
• softmax denominator runs over all $|\mathcal{I}|$ items — the output space is the catalogue

In the generative view, an item identifier is a length-$L$ token sequence decoded autoregressively. Atomic IDs are the degenerate case:

$p_{\theta }({\mathbf{z}}_i\mid \mathbf{x})={\prod }_{\ell =1}^L{p}_{\theta }(z_{i,\ell }\mid \mathbf{x},z_{i,<\ell })\stackrel{L=1}{\to }{p}_{\theta }(z_{i,1}\mid \mathbf{x})$

where:

• $\mathbf{x}=(x_1,\dots ,x_t)$ — user history; ${\mathbf{z}}_i$ — item identifier
• $L=1$ — a single decoding step per item (vs. $L$ steps for a semantic ID)
• codebook at position 1 has $K=|\mathcal{I}|$ entries — so $K^L=|\mathcal{I}|$, capacity is not decoupled from vocabulary (contrast: semantic IDs reach $K^L\gg K\cdot L$ tokens, e.g. $256^4\approx 4.3\times 10^9$ items from only $4\times 256$ tokens)

Key Properties / Variants

• Strengths: simplest possible scheme; no tokenizer stage; direct, collision-free item↔token lookup; one autoregressive step per recommendation; works fine for finite, stable catalogues.
• Weaknesses: vocabulary explodes with catalogue size; tokens are semantically arbitrary; similar items share nothing; no cold-start generalization; popularity bias enters through the large output softmax; embedding table memory grows linearly.
• Cold start (atomic vs semantic): with atomic IDs a new item $i^{\star }$ has no token in the vocabulary — an embedding row must be added and learned from interactions, and until then $i^{\star }$ is unrecommendable. With semantic IDs the new item is run through a frozen tokenizer, its sub-tokens already exist, and similar items share prefixes, giving warm-start generalization “for free”.
• Where atomic IDs still appear in GenRec: GPTRec and P5 use atomic IDs; there is even a counter-current paper “Atomic IDs are enough” (arXiv:2508.10478) arguing they remain competitive in some regimes.
• Other ladder rungs (plain text, not the focus here): L2 text/description-based IDs (very long, no collaborative info); L3 codebook-based semantic IDs from RQ-VAE (compact + semantic); L4 codebook + collaborative signal (LETTER, TokenRec); L5 adaptive IDs.

Building training examples — Atomic IDs (next-item prediction)
──────────────────────────────────────────────────────────────
Input : chronological user sequence (i_1, i_2, ..., i_t, i_{t+1})
Map   : each item -> ONE atomic token,  i_j -> <i_j>
        history = [<i_1>, <i_2>, ..., <i_t>]   (t tokens)
        target  = <i_{t+1}>                    (1 token)
Learn : p(i_{t+1} | i_1, ..., i_t)             (single AR step)
Vocab : exactly |I|  (one token per catalogue item)
 
Cold start (new item i*):
  token <i*> does NOT exist in vocab
  -> add embedding row, train from interactions  (strict cold start)

Connections

• Level L1 / special case ($L=1$) of: Item Tokenization, Semantic IDs
• The scheme Semantic IDs and RQ-VAE were designed to replace
• Used by score-and-rank classical models: SASRec, Next-Item Prediction over a Sequential Recommendation history
• Contrasts with: Generative Retrieval, Trie-Constrained Decoding (semantic-ID decoding must stay valid; atomic IDs are trivially valid)
• Drives the Cold Start Problem in generative recommenders
• Feeds the Top-K Recommendation softmax in discriminative Collaborative Filtering / Matrix Factorization

Appears In

• RS-L03b - From LLMs to LRMs
• RS-L04 - Generative Recommendation