OnePiece: Applying LLM‑Style Reasoning to Item‑ID Sequences for Generative Recommendation

Problem Context

Meta’s HSTU paper demonstrated a scaling law for generative recommendation (GR) models, prompting many industrial attempts to adopt HSTU‑style flat feature injection. Directly transplanting those solutions failed because baseline optimizations, user behavior, data scale, and system architectures differ across companies.

Failure Analysis

Baseline optimizations, user habits, and data characteristics vary, making flat feature injection ineffective.

Feature, sample, and inference services are tightly coupled to existing pipelines, preventing low‑cost MVP experiments.

Key technical details in prior works (e.g., HLLM) were missing, leading to reproduced models that lag behind strong dual‑tower baselines.

Attention mechanisms are already deeply integrated in DLRM; replacing them incurs a loss that must be compensated.

Research Goal

Identify a verifiable MVP that works within a traditional recall‑ranking pipeline and migrate proven LLM techniques to industrial recommendation. The resulting OnePiece framework injects LLM‑style context engineering and implicit (latent) reasoning into a model that treats item IDs as language tokens.

OnePiece Architecture

OnePiece supports two modes (recall and ranking) and three components: context engineering, latent reasoning, and progressive multi‑task training.

1. Context Engineering

Four prompt parts are built for each query:

Interaction History (IH) : mixed, time‑ordered, de‑duplicated user actions.

Preference Anchor (PA) : curated hot items per query, providing an inductive bias.

Situational Descriptor (SD) : heterogeneous tokens describing scene or user context, embedded via a separate adapter.

Candidate Item Set (CIS) : visible only in ranking mode; candidates are split into blocks for parallel inference.

Experiments show that expanding the PA sequence and adding SD tokens yields consistent lift in both recall and ranking stages.

2. Latent Reasoning

After processing the full context, the model emits a latent token that is fed back into the transformer without a softmax step, allowing further reasoning in latent space. This reduces decoding steps from hundreds to 5‑6 tokens and cuts inference cost by tens to hundreds of times, consistent with findings from CoCoNut.

3. Progressive Multi‑Task Training

Latent reasoning preserves all information, but training can become unstable. Inspired by ReaRec and GNOLR (arXiv:2505.20900), progressive supervision signals are added step‑by‑step. The loss combines an InfoNCE contrastive term with a BCE calibration term; the latter prevents logits from exploding to NaN during transformer training.

Training Configuration

Both recall and ranking models use a two‑layer transformer (hidden size = 256, pre‑norm). Side‑info embeddings are trimmed to less than half the feature count of a typical DLRM, and embedding values are clipped to [-0.02, 0.02]. Separate learning rates for dense and embedding parameters avoid gradient explosion.

Deployment Considerations

Recall models deploy easily because only user embeddings are needed at inference time. Ranking models require block‑wise scoring, which conflicted with an existing point‑wise inference platform. Reducing the block size to 1 enabled a “residual” OnePiece version that launched with modest performance loss yet still outperformed the DLRM baseline.

Attention Visualization

Heat‑maps of attention heads reveal:

In recall mode, heads capture user session clusters in the IH sequence and highlight PA items that guide reasoning.

In ranking mode, SD tokens dominate attention, injecting scene‑specific signals into both IH and CIS sequences.

These patterns confirm that engineered prompts steer the model’s reasoning as intended.

Online Performance

On a production TensorFlow training and inference stack, OnePiece achieved:

+1.12 % GMV per user and +2.9 % ad‑revenue uplift in the ranking stage.

+1.08 % GMV per user and +0.98 % paid‑order uplift in the recall stage.

Exposure and click‑coverage analysis shows that OnePiece independently recalls many items later confirmed by downstream ranking, indicating a strong explore‑exploit balance.

Future Work

Planned directions for 2025 include model compression, efficient retrieval, and broader generalisation. Extending PA design (e.g., injecting graph‑based item‑to‑item information) is suggested to move toward a truly general‑purpose recommendation model.

GitHub: https://github.com/CongFu92