SCASRec: A Self-Correcting and Auto-Stopping Model for Generative Route List Recommendation

Route recommendation systems commonly adopt a multi-stage pipeline involving fine-ranking and re-ranking to produce high-quality ordered recommendations. However, this paradigm faces three critical limitations. First, there is a misalignment between offline training objectives and online metrics. Offline gains do not necessarily translate to online improvements. Actual performance must be validated through A/B testing, which may potentially compromise the user experience. Second, redundancy elimination relies on rigid, handcrafted rules that lack adaptability to the high variance in user intent and the unstructured complexity of real-world scenarios. Third, the strict separation between fine-ranking and re-ranking stages leads to sub-optimal performance. Since each module is optimized in isolation, the fine-ranking stage remains oblivious to the list-level objectives (e.g., diversity) targeted by the re-ranker, thereby preventing the system from achieving a jointly optimized global optimum. To overcome these intertwined challenges, we propose SCASRec (Self-Correcting and Auto-Stopping Recommendation), a unified generative framework that integrates ranking and redundancy elimination into a single end-to-end process. SCASRec introduces a stepwise corrective reward (SCR) to guide list-wise refinement by focusing on hard samples, and employs a learnable End-of-Recommendation (EOR) token to terminate generation adaptively when no further improvement is expected. Experiments on two large-scale, open-sourced route recommendation datasets demonstrate that SCASRec establishes an SOTA in offline and online settings. SCASRec has been fully deployed in a real-world navigation app, demonstrating its effectiveness.

💡 Research Summary

SCASRec tackles three fundamental shortcomings of conventional route‑list recommendation pipelines: (1) the mismatch between offline training objectives (typically item‑level click signals) and online performance metrics such as list coverage and diversity; (2) reliance on rigid, handcrafted redundancy‑removal rules that cannot adapt to varying user intents and contextual factors; and (3) the fragmented optimization caused by separating fine‑ranking and re‑ranking stages, which prevents joint learning of item relevance and list‑level objectives.

To address these issues, the authors propose a unified encoder‑decoder generative architecture that produces the ordered route list token by token. Two novel mechanisms are introduced. First, the Stepwise Corrective Reward (SCR) computes, at each generation step t, the gap between the current list’s List Coverage Rate (LCR) and the maximal possible coverage derived from offline logs. Formally, (r^{SCR}t = \hat{p}{CR} - LCR(\bar{P}_t)). This reward highlights “hard” samples where substantial improvement remains, weighting the loss so that the model focuses on steps that can most increase coverage and, consequently, Mean Reciprocal Rank (MRR). SCR thus aligns offline training directly with online‑aligned list‑level metrics and implicitly encourages diversity because adding a redundant route yields little LCR gain.

Second, a learnable End‑of‑Recommendation (EOR) token replaces static stop‑criteria. When the ground‑truth route first appears at step (\hat{t}), the model receives a positive reward (\alpha) for emitting EOR at step (\hat{t}+1) and zero otherwise. This directly optimizes the redundancy penalty term (-\alpha|Z|) in the global objective
\