简体中文 · English

RECAP: Closed-loop recommendation feedback for streaming semantic profiles¶

Fidelity: core-mechanism reproduction. Trains a causal Transformer updater, a two-tower evaluator, and clipped GRPO, and runs a fixed-capacity profile state machine.

Paper information¶

Field	Details
Paper	arXiv 2607.15730
Institution	Kuaishou Technology / USTC
First public release	2026-07-17 (arXiv v1)
Author code	No official or author-released code found in the paper (checked 2026-08-09).
Local adapter or method key	`recap`
Local reproduction code	`src/auto_research/reproductions/recap/`

Original paper summary¶

Background and main changes¶

Conventional LLM profiles aim to summarize users plausibly without directly optimizing future recommendations. RECAP makes the profile a structured, capacity-constrained streaming state. The LLM handles semantic updates, while deterministic logic handles lifecycle, merging, and eviction. An LLM judge screens for label-consistent behavior pairs, a two-tower evaluator learns from them, and its matching score becomes the GRPO reward for closing the profile-update loop.

flowchart LR
  B[streaming behaviors] --> U[LLM profile updater]
  U --> S[bounded structured state]
  S --> C[decay / merge / evict]
  B --> J[LLM label-consistency judge]
  J --> E[dual-tower evaluator]
  E --> R[semantic reward]
  R --> G[GRPO]
  G --> U
  C --> REC[recommendation]

Key figure from the paper¶

Figure 1: The core architecture, its main modules, and their connections. From the paper; copyright belongs to the authors. Click for the source.

Core equations¶

RECAP samples a group of profile updates \(o_i\) for the same context and normalizes advantages within the group:

\[ A_i=\frac{r_i-\operatorname{mean}(r)}{\operatorname{std}(r)+\epsilon},\qquad r_i=s_{\mathrm{eval}}(p_i,b^+). \]

The policy update uses clipped ratios with reference KL, format, and diversity constraints. The state machine always satisfies \(|P_t|\le C\).

Offline and online results¶

Compared with the base generator, offline uAUC rises +0.0084 absolute and Recall@2000 rises about +4.9%. In a 7-day online A/B test, time spent in the app per user rises significantly by +0.139%.

Local reproduction¶

Local comparison: The baseline is an open-loop SFT streaming profile. The experimental arm adds the feedback evaluator and GRPO. Relative NDCG@10 changes by -6.77%.

MovieLens-100K uses 220 users and 360 items. Genre tokens provide interpretable semantics, and profile capacity is 4. GRPO mean reward rises from 0.5245 to 0.7096, but test NDCG falls from 0.04838 to 0.04511 and head share from 0.2400 to 0.1914. Local evaluator reward gains do not transfer reliably to next-item metrics. See metrics/movielens-100k-seed42.json.

auto-research reproduce --paper recap --seed 42

Scope¶

Natural-language profiles and industrial LLMs are scaled down to a category-token causal Transformer. There are no private Kuaishou logs, production LLM judge, or online feedback loop. The core SFT → evaluator → GRPO → bounded-state pipeline executes, but it does not reproduce online gains.