Third-Place Solution for ACM RecSys Challenge 2026 Music Recommendation
The team known as 'swyoo' secured third place in the Blind-B industry track of the ACM RecSys Challenge 2026, which demanded a top-20 ranking alongside evidence-based response generation. Their innovative system separates retrieval and response processes into distinct pipelines linked through ranked tracks and metadata. The retrieval mechanism utilizes a hybrid lexical-dense pool for precise matching, supplemented by a task-specific pool enhanced with fine-tuned Qwen 8B adapters. Candidates undergo calibration via LightGBM before being directed to an evidence-grounded propose-assign-select (PAS) framework for response structuring. Additionally, the system achieved second place on the explanation-quality leaderboard during the final blind evaluation. This research tackles the limitations of traditional conversational recommenders, where precise entity cues diminish as dialogue intent shifts, undermining explanation reliability. The associated paper is available on arXiv with ID 2607.24846.
Key facts
- Team 'swyoo' won third place in the Blind-B industry track of the ACM RecSys Challenge 2026.
- The challenge mandates both top-20 ranking and evidence-grounded response generation.
- The system decouples retrieval and response into separate pipelines.
- Retrieval uses a hybrid lexical-dense pool and a task-adapted pool with fine-tuned Qwen 8B adapters.
- Candidates are calibrated via LightGBM.
- An evidence-grounded propose-assign-select (PAS) framework structures responses.
- The system ranked second on the explanation-quality leaderboard.
- The paper is available on arXiv with ID 2607.24846.
Entities
Institutions
- ACM RecSys Challenge 2026
- arXiv