HLLM-Creator: Personalized Ad Titles

• HLLM-Creator is a hierarchical framework that generates personalized ad titles using structured user interest modeling, factual grounding, and scalable inference techniques.
• It leverages a three-tier architecture—item, user, and creative LLMs—to encode behaviors, aggregate user profiles, and generate ad-specific content with measurable performance gains.
• The system incorporates a chain-of-thought data construction pipeline and clustering-based inference pruning to ensure efficiency and reduce hallucination in industrial-scale ad personalization.

HLLM-Creator is a hierarchical LLM framework for personalized creative generation, introduced for personalized title generation in Douyin Search Ads. It addresses three coupled constraints that are central to industrial AIGC: user interest modeling, factual grounding to ad attributes, and inference efficiency at the scale of hundreds of millions of users and millions of ads. The framework combines hierarchical user representation learning, constrained creative generation, a chain-of-thought data construction pipeline, and cluster-based inference pruning; in reported deployment, the online A/B test shows a $0.476\%$ increase on Adss, alongside gains on Advv and RankAdvv (Chen et al., 25 Aug 2025).

1. Problem Formulation and Target Setting

The task studied by HLLM-Creator is personalized title generation for online advertising. Given a user $u$, the user’s historical behavior sequence

$H_u = \{I_1, I_2, \dots, I_n\},$

and a target ad $I$ with an original ad title $y_{\text{orig}}$ and a set of product selling points $\mathbf{s}$, the objective is to learn a generator

$p_\Theta(y \mid u, I)$

that produces a personalized creative title $y$ aligned with user interests while remaining factually grounded in the ad’s true attributes (Chen et al., 25 Aug 2025).

The framework is motivated by a concrete industrial asymmetry: advertisers typically provide one broad title, whereas real deployment requires user-centric variation. The paper identifies two main challenges: accurately modeling user interests from behavior history and integrating them into generation under factual constraints, and ensuring efficiency and scalability for industrial deployment. A third challenge is the scarcity of personalized creative data, since human-written user-specific ad titles are rarely available in practice (Chen et al., 25 Aug 2025).

A recurrent misconception in this problem class is to treat personalization as a purely prompt-level rewriting problem. HLLM-Creator is explicitly not organized that way. Its design assumes that personalization quality depends on a learned user representation, and that factuality requires ad-side constraints rather than unconstrained free-form generation. In the search advertising setting, the query is also treated as an immediate intent signal and is incorporated both during data construction and at generation time (Chen et al., 25 Aug 2025).

2. Hierarchical Architecture

HLLM-Creator is organized around three LLM roles: an Item LLM, a User LLM, and a Creative LLM. This hierarchy separates item-level semantic encoding, user-level aggregation, and ad-conditioned creative generation, rather than forcing a single model to consume raw click histories and ad metadata in one pass (Chen et al., 25 Aug 2025).

At the item level, each historical clicked ad is converted into text attributes flattened as a textual sentence $\mathcal{T}$. After tokenization, a special token $[item]$ is appended: $u$0 The hidden state of the final $u$1 token is taken as the item embedding $u$2. Thus the user history

$u$3

is mapped to

$u$4

At the user level, the sequence of item embeddings is fed into the User LLM. A special $u$5 token is appended, and the hidden state of that token is used as the user embedding $u$6. The paper describes this as a late fusion user modeling strategy. At the generation level, ad-side information is represented as token embeddings

$u$7

and the Creative LLM conditions on

$u$8

to generate a personalized title autoregressively (Chen et al., 25 Aug 2025).

This hierarchy has both modeling and systems consequences. Modeling-wise, it turns long, noisy behavior sequences into a compressed semantic user state. Systems-wise, it creates a reusable user representation that can later be clustered and reused in offline generation. In the reported configuration, the Item LLM is TinyLlama-1.1B, the User LLM is TinyLlama-1.1B, and the Creative LLM is Qwen3-8B. User history length is capped at $u$9, item title input length at $H_u = \{I_1, I_2, \dots, I_n\},$0 tokens, the user embedding is projected to the Creative LLM dimension via a single-layer FC, and the User Feature Extractor, User Interest Decoder, and Creative LLM share parameters to save memory (Chen et al., 25 Aug 2025).

3. User-Interest Representation and Training Objectives

The framework trains the hierarchical representation not only through title generation, but also through auxiliary supervision designed to make the user embedding semantically descriptive and recommendation-relevant. The overall objective combines one generation loss and three auxiliary losses; all auxiliary loss weights are set to $H_u = \{I_1, I_2, \dots, I_n\},$1 in the experiments (Chen et al., 25 Aug 2025).

The generative loss is standard supervised fine-tuning on synthetic personalized titles: $H_u = \{I_1, I_2, \dots, I_n\},$2 Here $H_u = \{I_1, I_2, \dots, I_n\},$3 is the synthesized target title and $H_u = \{I_1, I_2, \dots, I_n\},$4 is the response length. This term trains the Creative LLM to map the user embedding and ad-side inputs into the final title (Chen et al., 25 Aug 2025).

The recommendation objective $H_u = \{I_1, I_2, \dots, I_n\},$5 makes the user embedding more discriminative for click preference. A shallow predictor consumes user embedding $H_u = \{I_1, I_2, \dots, I_n\},$6 and item embedding $H_u = \{I_1, I_2, \dots, I_n\},$7 and outputs a click probability through a dense network and sigmoid: $H_u = \{I_1, I_2, \dots, I_n\},$8 Positive samples are clicked items and negative samples are randomly sampled (Chen et al., 25 Aug 2025).

The semantic alignment loss $H_u = \{I_1, I_2, \dots, I_n\},$9 aligns the learned user embedding $I$0 with an explicit user-interest description encoded as a feature $I$1: $I$2 The reconstruction loss $I$3 forces the user embedding to reconstruct the textual user-interest description: $I$4 In the paper’s interpretation, these terms play complementary roles: $I$5 makes $I$6 predictive of preference, $I$7 grounds $I$8 in explicit textual interest semantics, and $I$9 makes $y_{\text{orig}}$0 preserve descriptive user information (Chen et al., 25 Aug 2025).

The reported ablations support that interpretation. On the industrial dataset, the GSB advantage improves from $y_{\text{orig}}$1 with no auxiliary loss to $y_{\text{orig}}$2 with align only, $y_{\text{orig}}$3 with cls only, $y_{\text{orig}}$4 with recon only, $y_{\text{orig}}$5 with align + cls, and $y_{\text{orig}}$6 with all three. Within the framework’s own evidence, this is one of the strongest indications that the method is fundamentally a representation-learning system rather than a prompt-engineered generator (Chen et al., 25 Aug 2025).

4. Chain-of-Thought Data Construction and Hallucination Control

Because real personalized ad titles are scarce, HLLM-Creator constructs synthetic supervision through a chain-of-thought pipeline. The paper emphasizes that naively prompting an LLM with user history and ad information yields weak personalization and can hallucinate facts. Its solution is a three-stage decomposition: User Interest Profiling, Interest-Driven Title Generation, and Hallucination-Free Title Filtering (Chen et al., 25 Aug 2025).

In User Interest Profiling, an LLM such as DeepSeek-R1 reads the user’s historical sequence and summarizes user interests along multiple dimensions, including long-term interests, short-term preferences, and specific needs. In Interest-Driven Title Generation, the paper hypothesizes that personalization in advertising titles generally occurs in two ways: explicitly mentioning a user group, or selecting ad selling points that match the user. The model is therefore guided to identify which user interests are relevant to the ad, select selling points likely to attract that user, and then synthesize a final title combining those elements (Chen et al., 25 Aug 2025).

The final stage, Hallucination-Free Title Filtering, removes titles that invent unsupported facts. The paper is explicit that this stage is not optional in the advertising setting, because ad copy cannot contain misleading clickbait. After filtering, only about one quarter of the initially generated data remains. From an initial $y_{\text{orig}}$7 million samples drawn from Douyin Search Ads click logs, the final split is $y_{\text{orig}}$8K samples for training and $y_{\text{orig}}$9 samples for offline evaluation (Chen et al., 25 Aug 2025).

The paper’s ablations show that the CoT decomposition materially affects data quality. Reported GSB advantage rises from $\mathbf{s}$0 with no CoT and a single chat round, to $\mathbf{s}$1 with one CoT step, to $\mathbf{s}$2 with two chat rounds using CoT1 only, and to $\mathbf{s}$3 with two chat rounds using both CoT1 and CoT2. Hallucination control also improves measurably: the final hallucination pass rate is $\mathbf{s}$4 for HLLM-Creator, compared with $\mathbf{s}$5 for the Qwen3-8B baseline, and on the hallucination-free subset the model’s advantage rises to $\mathbf{s}$6 (Chen et al., 25 Aug 2025).

5. Industrial Inference Strategy and Scalability

HLLM-Creator is designed for offline large-scale deployment rather than per-request user-specific generation. The central engineering problem is straightforward: generating a distinct title for every $\mathbf{s}$7 pair is computationally infeasible. The framework therefore introduces two coupled deployment strategies, user clustering and user-ad-matching-prediction based pruning (Chen et al., 25 Aug 2025).

First, users are encoded into embeddings through the hierarchical model and clustered offline using k-means into $\mathbf{s}$8 groups. Each cluster center $\mathbf{s}$9 becomes a user segment representation. Creative generation is then performed for cluster centers rather than for individual users. The paper states the resulting cost reduction directly: instead of scaling roughly as $p_\Theta(y \mid u, I)$0, inference scales roughly as $p_\Theta(y \mid u, I)$1 (Chen et al., 25 Aug 2025).

Second, not every ad is relevant to every cluster. The system therefore uses an Item-User Predictor to score the relevance of a target ad embedding $p_\Theta(y \mid u, I)$2 against each cluster center $p_\Theta(y \mid u, I)$3, and retains only the top-$p_\Theta(y \mid u, I)$4 clusters for generation. This pruning stage further concentrates generation on the user segments most likely to benefit. In the reported model configuration, the Item-User Predictor is an 8-layer MLP-Mixer with hidden size $p_\Theta(y \mid u, I)$5, and the default clustering setup uses $p_\Theta(y \mid u, I)$6 clusters (Chen et al., 25 Aug 2025).

This deployment logic introduces an explicit quality-efficiency tradeoff. The paper notes that clustering helps scalability but slightly hurts quality, and that performance improves as the number of clusters increases, though gains diminish. Even so, with only $p_\Theta(y \mid u, I)$7 clusters the model still achieves a $p_\Theta(y \mid u, I)$8 advantage over original titles. This makes clear that the framework’s industrial claim is not merely that personalization is possible, but that personalization can be amortized through representation compression and cluster-level pre-generation (Chen et al., 25 Aug 2025).

6. Evaluation, Results, and Position Within the Broader Creator Paradigm

The online evaluation uses Adss, Advv, and RankAdvv, with

$p_\Theta(y \mid u, I)$9

and

$y$0

Offline evaluation uses GPT-based GSB preference and hallucination pass rate. In GSB preference, GPT-4.1 compares two titles and labels the result as Good, Same, or Bad, and the reported scalar is

$y$1

To reduce order bias, both $y$2 and $y$3 are tested and inconsistent outcomes are mapped to Same (Chen et al., 25 Aug 2025).

The headline deployment results are modest in absolute percentage terms but significant in context. On Douyin Search Ads, HLLM-Creator improves Adss by $y$4, Advv by $y$5, and RankAdvv by $y$6. On ads where personalized titles won the selection stage, the paper also reports a CTR increase of $y$7. Against baselines, the comparison is nuanced rather than uniformly dominant: with equal sequence length, Pigeon has a $y$8 advantage in offline GSB evaluation, but when HLLM-Creator is allowed longer user sequences of length $y$9, it achieves a $\mathcal{T}$0 advantage; compared with a general-purpose Qwen3-8B prompt baseline, HLLM-Creator has about a $\mathcal{T}$1 advantage (Chen et al., 25 Aug 2025).

The paper’s broader significance lies in how it instantiates a “creator” pattern. In adjacent literature, CREATOR reframes LLM reasoning by letting the model create tools on demand and then use them, separating abstract tool creation from concrete decision execution and adding a rectification loop driven by execution feedback (Qian et al., 2023). GenAI-DrawIO-Creator applies a related logic to diagram authoring: Claude 3.7 produces draw.io-compatible XML, which is then validated, semantically checked, and corrected if necessary (Yu et al., 8 Jan 2026). CollabToolBuilder extends the pattern into a human-in-the-loop multi-agent setting with Coach, Coder, Critic, and Capitalizer, emphasizing reusable tool storage and iterative expert-guided refinement (Xavier et al., 1 Dec 2025). This suggests that HLLM-Creator belongs to a broader family of creator-oriented LLM systems in which the model is not used merely to emit final text, but to construct constrained artifacts under explicit structural, semantic, or human-governed control.

Within its own domain, however, HLLM-Creator remains bounded by several limitations. Clustering loses some personalization granularity; the synthetic data pipeline depends on prompt quality and LLM capability; hallucination cannot be fully eliminated and still requires rule-based filtering and safety checks; and the method is optimized for offline generation rather than real-time per-request deployment (Chen et al., 25 Aug 2025). The resulting picture is not of unconstrained generative personalization, but of a layered industrial system in which representation learning, synthetic supervision, constrained generation, and deployment-time approximation are all necessary to make personalized creative generation operational at scale.