Term IDs (TIDs): Robust Identifiers in LLM Recommenders

• Term IDs (TIDs) are structured, standardized sequences of concise keywords that serve as robust item identifiers in LLM-based generative recommendation systems.
• They leverage Context-aware Term Generation and Integrative Instruction Fine-tuning to transform free-text metadata into semantically dense and discriminative identifier sequences.
• Empirical results show improved Recall@5 metrics and nearly perfect grounding accuracy, demonstrating significant performance gains and reduced hallucination.

Term IDs (TIDs) are structured, standardized sequences of short human-readable keywords designed to serve as robust item identifiers within LLM-based generative recommendation systems. TIDs address challenges inherent to previous identifier schemes, specifically the vastness and ambiguity of free-text output and the vocabulary alignment gap of Semantic IDs (SIDs), by leveraging semantically dense and discriminative native tokens for both item representation and recommendation generation (Zhang et al., 11 Jan 2026).

1. Formal Definition and Properties

Let $T = \{ t_1, t_2, \ldots, t_K \}$ denote the fixed vocabulary of candidate terms, where each $t_i$ is a concise, standardized textual keyword (e.g., "iPhone", "128GB"). For each item $i \in \mathcal I$, an ordered sequence of Term IDs is assigned:

$T_i = (t_{i}^1, t_{i}^2, \ldots, t_{i}^N) \in T^N$

The mapping from item metadata to Term IDs is written as:

$M: \mathcal I \rightarrow T^N, \quad M(i) = T_i$

Key constraints are imposed:

• The sequence has fixed length $N$ (typically $N=5$).
• All terms are sourced from the controlled set $T$ (raw size $\approx 50K$).
• Terms are individually human-readable, unambiguous, and semantically discriminative, enhancing item-level distinction and interpretability.

2. Context-aware Term Generation (CTG)

Context-aware Term Generation (CTG) systematically converts an item's free-text metadata $m_i$ (title, description, attributes) into its structured Term ID sequence $T_i$. The CTG process comprises several algorithmic steps:

1. Metadata Embedding & Neighbor Retrieval: All item metadata $m_j$ is embedded into $\mathbf v_j \in \mathbb R^d$ using a frozen embedding model. For a target item $i$, cosine similarities $\cos(\mathbf v_i, \mathbf v_j)$ are computed to select the top-$k$ nearest neighbors $\mathcal N_i = \{j_1, \ldots, j_k\}$.
2. Structured Prompting: A prompt $\mathcal P(m_i, \{m_j\}_{j\in\mathcal N_i})$ is constructed to instruct the LLM to extract terms both globally consistent with neighbors and locally discriminative for $i$.
3. Term Generation: The LLM generates the Term ID sequence:

$$T_i = \mathrm{LLM}(\mathcal P(m_i, \{m_j\}_{j\in\mathcal N_i}))$$

1. (Optional) CTG Fine-tuning: CTG is optionally fine-tuned with next-token prediction:

$$\mathcal L_{\rm CTG} = -\sum_{n=1}^N \log P(t_i^n | m_i, \{m_j\}, t_i^{<n})$$

A concrete example involves mapping the title "Sony WH-1000XM4 Wireless Noise-Canceling Headphones" and its similar neighbors to the sequence $$(\text{“Sony”},\text{“WH-1000XM4”},\text{“Wireless”},\text{“Noise-Canceling”},\text{“Headphones”})$$.

3. Integrative Instruction Fine-tuning (IIFT)

Integrative Instruction Fine-tuning (IIFT) jointly optimizes two objectives to internalize the semantics of TIDs and to improve user sequence recommendation:

• Generative Term Internalization (GTI): Given $m_i$, predict $T_i$ using the loss

$$\mathcal L_{\rm GTI} = -\sum_{n=1}^N \log P(t_i^n | m_i, t_i^{<n})$$

• User Behavior Sequence Prediction: Each user history item $i_j$ is represented as $x_j = [T_{i_j}; m_{i_j}^{\rm title}]$. The loss over a sequence $S = \{i_1, \ldots, i_n\}$ with prefix $\mathcal I$ is

$$\mathcal L_{\rm rec} = -\sum_{k=2}^n \log P(x_k | \mathcal I, x_1, \ldots, x_{k-1})$$

The total loss is a weighted sum:

$$\mathcal L = \mathcal L_{\rm GTI} + \lambda \mathcal L_{\rm rec}$$

During fine-tuning, distinct instruction structures are provided for both GTI ("Extract 5 concise terms from the following metadata: $\langle m_i\rangle$ →") and recommendation ("Given your interaction history as $[T_1;Title_1],\ldots,[T_{k-1};Title_{k-1}]$, predict the next item's 5 terms →").

4. Elastic Identifier Grounding (EIG)

Elastic Identifier Grounding (EIG) reliably translates generated TID sequences back to real item identities. EIG implements a hybrid retrieval mechanism:

1. Direct Mapping: An exact string match against the prebuilt library $\{T_j\}_j$ yields immediate identification.
2. Structural Mapping: If direct mapping fails, structural similarity is used. For generated sequence $(t_{\rm gen}^1, \ldots, t_{\rm gen}^N)$, EIG selects

$$i^* = \arg\max_{j\in\mathcal C} \sum_{n=1}^N w_n \mathbf 1\{t_{\rm gen}^n = t_j^n\}, \qquad w_n = \frac{1}{n+1}$$

This approach prioritizes early-term matches with decaying weights, enhancing both robustness and precision in identifier grounding.

5. Comparative Assessment and Empirical Results

Relative to alternative identifier schemes:

• Vs. Textual IDs: Raw text-based identifiers are lengthy and non-discriminative, which expands the output space and fosters hallucination. TIDs, by contrast, represent high-information content within concise tokens, significantly reducing hallucination.
• Vs. Semantic IDs (SIDs): SIDs consist of numerical codes necessitating costly vocabulary expansion and alignment. TIDs utilize the LLM’s native token inventory, circumventing such overhead and directly leveraging LLM world knowledge.

Empirical performance improvements (in-domain) include:

• +7.8% Recall@5 on Beauty
• +30.2% Recall@5 on Sports
• +14.9% Recall@5 on Toys

Hallucination is quantifiably suppressed, with VR@10 (Valid Rate) and DHR@10 (Direct Hit Rate) both exceeding 99%, confirming near-perfect identifier grounding without the need for constrained decoding.

6. End-to-End Operation and Illustrative Example

A comprehensive workflow involves representing user interaction history in TIDs. For a user purchasing:

1. "Apple iPhone 14 Pro 128GB" $\rightarrow$ $$(\text{“Apple”},\text{“iPhone”},\text{“14”},\text{“Pro”},\text{“128GB”})$$
2. "Samsung Galaxy S23 Ultra 256GB" $\rightarrow$ $$(\text{“Samsung”},\text{“Galaxy”},\text{“S23”},\text{“Ultra”},\text{“256GB”})$$

A recommendation prompt:

History: [Apple,iPhone,14,Pro,128GB; Apple iPhone 14 Pro] [Samsung,Galaxy,S23,Ultra,256GB; Samsung Galaxy S23 Ultra] → Predict next 5 terms.

Model outputs:

$$(\text{“Apple”},\text{“AirPods”},\text{“Pro”},\text{“Wireless”},\text{“Earbuds”})$$

Direct grounding links this to “Apple AirPods Pro Wireless Earbuds,” reliably mapping generated TIDs to concrete catalog items via native tokens, eliminating requirements for external indices or post-hoc alignment.

7. Significance and Implications for Generative Recommendation

The deployment of Term IDs as the backbone of generative recommendation advances the field by enabling precise, low-hallucination item identification within LLM-native output spaces. TIDs facilitate robust, context-sensitive recommendations without specialized vocabulary expansion, opening new directions for system generalizability and performance. This framework demonstrates that semantically rich, human-readable identifiers can unlock both interpretability and operational efficiency in next-generation recommender architectures, as evidenced in GRLM's empirical superiority and alignment with practical deployment needs (Zhang et al., 11 Jan 2026).