TencentGR-1M: Generative Ad Recommendation Benchmark

• TencentGR-1M is a large-scale, privacy-preserving benchmark for next-click prediction, utilizing de-identified Tencent Ads logs for generative recommendation.
• It combines collaborative IDs, categorical features, and multi-modal text and vision embeddings, covering 1M users, 4.78M ads, and a 660K candidate pool.
• A causal Transformer baseline with InfoNCE loss and ANN retrieval demonstrates effective autoregressive modeling over tokenized user-ad sequences.

Searching arXiv for the specified paper and topic to ground the article. TencentGR-1M is the preliminary-round dataset for the Tencent Advertising Algorithm Challenge 2025: All-Modality Generative Recommendation. It is a large-scale, privacy-preserving benchmark for generative recommendation in industrial advertising, built from real de-identified Tencent Ads logs. In this setting, the task is to predict the next ad a user will interact with from a large candidate set using all modalities: collaborative IDs, categorical metadata, and precomputed text and vision embeddings. The dataset contains 1,001,845 users, 4,783,154 ads, a maximum sequence length of 100, an average sequence length of 91.06, and a candidate set size of 660,000 ads (Pan et al., 4 Apr 2026).

1. Dataset definition and benchmark scope

TencentGR-1M is designed for next-click prediction in a multi-modal ad interaction sequence setting. It serves as the preliminary-round benchmark in the challenge, where the positive behavior is click and the prediction target is the next clicked ad. The released sequences incorporate collaborative signals such as user IDs, ad IDs, and advertiser, product, and category IDs; multi-modal content embeddings from text and image encoders; and sequential behavior signals including exposure and click (Pan et al., 4 Apr 2026).

The benchmark is positioned specifically for all-modality generative recommendation rather than conventional pointwise or pairwise CTR estimation. The underlying formulation reflects the broader GR paradigm described in the paper: collaborative identifiers and multi-modal content representations are mapped into discrete token spaces, and user behavior is modeled with autoregressive sequence models. This suggests that TencentGR-1M is intended not merely as a ranking dataset with auxiliary features, but as an evaluation substrate for sequence generation over tokenized user-item interactions.

2. Construction pipeline and privacy constraints

The construction process begins by setting an answer window $[t_{\text{begin}}, t_{\text{end}}]$ and sampling users who have at least one positive behavior in that window; for TencentGR-1M, the positive behavior is click. For each sampled user, the first clicked ad after $t_{\text{begin}}$ is identified, that click is attributed to its corresponding impression or exposure event, and the impression that triggered the first click is used as the prediction target. The user history consists of all behaviors strictly before the exposure time of the target impression. A user token is prepended, and the final sequence is truncated to at most 100 item tokens (Pan et al., 4 Apr 2026).

The released data are explicitly privacy-preserving. All personally identifiable information is removed, raw creative content is removed, and only hashed IDs together with embedding vectors extracted by production models are released. This is central to the benchmark’s public usability: it preserves industrial realism while avoiding disclosure of raw creatives or user-identifying fields. A common misconception is that “all-modality” here implies access to raw text, images, or videos; in TencentGR-1M, it instead denotes access to state-of-the-art modality-specific embeddings derived from those inputs.

3. Representation schema, action labels, and modality coverage

The sequence for user $u$ is defined as

$$S_u = \{ x_u, x_{u,1}, x_{u,2}, \ldots, x_{u,T_u} \},$$

where $x_u$ is a user-profile token and each $x_{u,t}$ is an item token corresponding to an ad impression at time $t$. Item tokens include categorical attributes, action or feedback signals, and precomputed multi-modal embeddings; user tokens include profile features. TencentGR-1M includes two action types, exposure and click, with labels

$r_{u,t} \in \{0,1\},$

corresponding to exposure and click. The reported action distribution is 90.19% exposure and 9.81% click (Pan et al., 4 Apr 2026).

The feature schema contains user features with IDs 103, 104, 105, 106, 107, 108, 109, and 110, including both single-value and multi-value categorical features. Ad features have IDs 100, 101, 102, 112, 114, 115, 116, 117, 118, 119, 120, 121, and 122. These are categorical fields with very high coverage in most cases, and some are multi-value or high-cardinality features.

For multi-modal content, raw creatives are not released. Instead, the paper states that up to six embedding vectors per creative are extracted using production models:

Embedding source	Modality	Dimension
Bert-finetune	text	32
Conan-embedding-v1	text	1024
gte-Qwen2-7B-instruct	text	3584
hunyuan_mm_7B_finetune	image	4096/32
QQMM-embed-v1	image	3584
UniME-LLaVA-OneVision-7B	image	3584

The paper further states that Bert-finetune and hunyuan_mm_7B_finetune are finetuned using Tencent’s real-world collaborative data with contrastive learning, while the remaining encoders are used without finetuning. This suggests that TencentGR-1M operationalizes “all-modality” through precomputed semantic representations rather than raw-modality ingestion.

4. Formal task and evaluation protocol

The task is defined as next-item recommendation on multi-modal ad interaction sequences. Given a prefix of $S_u$, the model must generate the next ad $x_{u,T_u+1}$ that the user is most likely to interact with from a large candidate pool. In the preliminary round, the candidate set is a global pool of 660,000 de-duplicated ads. The true target item is always included, negatives are sampled from the ads corpus, and about 40% of ads in the candidate pool appear as targets for at least one user (Pan et al., 4 Apr 2026).

The leaderboard metrics are HitRate@10 and NDCG@10, combined as

$t_{\text{begin}}$0

The paper states that official metrics are reported at $t_{\text{begin}}$1 because $t_{\text{begin}}$2 gave a larger coefficient of variation than $t_{\text{begin}}$3, yielding clearer separation between teams, and because the score weights were calibrated so HitRate@10 and NDCG@10 contribute roughly equally. Since there is exactly one relevant item per user, the evaluation protocol is a single-target retrieval problem over a large shared candidate space rather than a multi-relevance ranked list setting.

5. Baseline model and practical instantiation of generative recommendation

The released baseline formulates the problem as next-token prediction with retrieval. A causal Transformer encodes the user history, the last hidden state represents the user, candidate items are scored against this user representation, and retrieval is performed with ANN search over item embeddings. In the feature encoder, each categorical or ID field has its own embedding table, multi-modal fields are used directly as continuous vectors, and field embeddings are concatenated and passed through an MLP to create the final token embedding (Pan et al., 4 Apr 2026).

The released baseline pipeline uses a causal Transformer, InfoNCE loss, ANN retrieval with Faiss, and field-wise sparse embeddings plus dense multi-modal vectors. The implementation details given in the paper are: 1 Transformer block, hidden size $t_{\text{begin}}$4, 1 attention head, dropout 0.2, max length 101, Adam optimizer with learning rate 0.001, and one negative sample per positive in the released baseline implementation.

This baseline is significant because it clarifies how the challenge’s “generative recommendation” terminology is instantiated in practice. The system is generative in the sense of sequence modeling over tokenized user and item representations, but deployment-facing retrieval remains explicit. A plausible implication is that TencentGR-1M is especially suitable for research at the interface of autoregressive modeling, semantic tokenization, and scalable retrieval.

6. Relationship to TencentGR-10M and research significance

TencentGR-1M is the smaller preliminary dataset, whereas TencentGR-10M is the larger second-round or final-round dataset. TencentGR-1M has 1,001,845 users and 4,783,154 ads, uses click-only positive targets, contains exposure and click labels, and evaluates next-click prediction with 660,000 candidate ads. TencentGR-10M has 10,139,575 users and 17,487,676 ads, introduces exposure, click, and conversion labels, supports next click-or-conversion prediction, and uses 3,637,720 candidate ads. Both datasets have maximum sequence length 100 (Pan et al., 4 Apr 2026).

Within that pairing, TencentGR-1M functions as the simpler but still industrially realistic entry point. Its utility derives from several properties stated in the paper: it is built from real Tencent Ads logs, combines collaborative IDs with categorical attributes and multi-modal embeddings, supports long sequence modeling, uses a large candidate space, is designed specifically for generative recommendation in advertising, and is publicly released in a privacy-preserving form. The broader challenge summary reports that top teams benefited from autoregressive or decoder-only generative architectures, action-conditioning, temporal feature engineering, semantic ID construction, and large-scale contrastive training with ANN retrieval. Although those findings summarize the full competition rather than only the preliminary round, they reinforce TencentGR-1M’s role as a benchmark for studying how collaborative identifiers, multi-modal representations, and autoregressive recommendation interact under industrial-scale retrieval constraints.