Paper deep dive
Two Views, One Voice: Evidence-Grounded Conversational Music Recommendation
Sungwook Yoo, Sewook Yoo
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 8/1/2026, 10:31:53 AM
Summary
This paper presents a third-place solution for the ACM RecSys Challenge 2026 Blind-B industry track, focusing on conversational music recommendation. The proposed system decouples retrieval and response generation to preserve entity cues and improve explanation credibility. It employs a hybrid retrieval architecture combining a lexical-dense pool (BM25 + 0.6B Qwen embeddings) and a task-adapted conversational pool (fine-tuned Qwen 8B adapters). Candidates are fused and calibrated using LightGBM. The response generation utilizes an evidence-grounded Propose-Assign-Select (PAS) framework, which structures explanations based on candidate-level evidence and a few-shot demonstration bank, ensuring factual grounding and reliability.
Entities (10)
Relation Signals (8)
Team swyoo → achievedrankin → ACM RecSys Challenge 2026
confidence 98% · This paper presents the third-place solution by team “swyoo” for the Blind-B industry track.
LightGBM → calibrates → Candidate Rankings
confidence 96% · Candidates are calibrated via LightGBM
PAS → structures → Response Generation
confidence 95% · routed to an evidence-grounded propose-assign-select (PAS) framework to structure responses.
Qwen-8B → usedin → Conversational Pool
confidence 95% · task-adapted pool driven by fine-tuned Qwen 8B adapters
BM25 → partof → Hybrid Pool
confidence 94% · The hybrid pool (bm25_qmr) combines sparse and dense mechanisms. The sparse component runs two BM25 searches
RRF → fuses → Retrieval Pools
confidence 93% · an RRF layer ... merges their rankings.
TalkPlayData-Challenge → supplies → Track Catalog
confidence 92% · The official TalkPlayData-Challenge dataset suite supplies a 47,071-track catalog
QLoRA → usedforfinetuning → Qwen-8B
confidence 90% · we fine-tune an 8B parameter encoder model using 4-bit QLoRA
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Traditional conversational recommenders entangle retrieval and response generation within a single text interface, so exact entity cues fade as the dialogue's intent evolves, which compromises explanation credibility. We address this within the ACM RecSys Challenge 2026, which mandates both top-20 ranking and evidence-grounded response generation. This paper presents the third-place solution by team "swyoo" for the Blind-B industry track. We decouple retrieval and response into separate pipelines connected strictly via ranked tracks and metadata. Retrieval combines a hybrid lexical-dense pool for exact matching with a task-adapted pool driven by fine-tuned Qwen 8B adapters. Candidates are calibrated via LightGBM, then routed to an evidence-grounded propose-assign-select (PAS) framework to structure responses. This system also ranked second on the explanation-quality leaderboard in the final blind evaluation. Our findings demonstrate that: (i) isolating retrieval and response preserves both catalog cues and fluid intent; (ii) structuring generation via explicit evidence assignment is key to this near-best-in-class explanation reliability.
Tags
Links
- Source: https://arxiv.org/abs/2607.24846v1
- Canonical: https://arxiv.org/abs/2607.24846v1
Trouble viewing inline? Open PDF directly →
Full Text
26,565 characters extracted from source content.
Expand or collapse full text
Two Views, One Voice: Evidence-Grounded Conversational Music Recommendation Sungwook Yoo l22491360@gmail.com NaverSouth Korea and Sewook Yoo sewookyoo@gmail.com SamsungSouth Korea (2026) Abstract. Traditional conversational recommenders entangle retrieval and response generation within a single text interface, so exact entity cues fade as the dialogue’s intent evolves, which compromises explanation credibility. We address this within the ACM RecSys Challenge 2026, which mandates both top-20 ranking and evidence-grounded response generation. This paper presents the third-place solution by team “swyoo” for the Blind-B industry track. We decouple retrieval and response into separate pipelines connected strictly via ranked tracks and metadata. Retrieval combines a hybrid lexical-dense pool for exact matching with a task-adapted pool driven by fine-tuned Qwen 8B adapters. Candidates are calibrated via LightGBM, then routed to an evidence-grounded propose-assign-select (PAS) framework to structure responses. This system also ranked second on the explanation-quality leaderboard in the final blind evaluation. Our findings demonstrate that: (i) isolating retrieval and response preserves both catalog cues and fluid intent; (i) structuring generation via explicit evidence assignment is key to this near-best-in-class explanation reliability. conversational recommendation, music recommendation, hybrid retrieval, learning to rank, grounded generation †copyright: none†conference: Proceedings of the Recommender Systems Challenge 2026; September 28–October 2, 2026; Minneapolis, MN, USA†journalyear: 2026†ccs: Information systems Recommender systems 1. Introduction Conversational music recommendation (CRS) must balance precise catalog entity retrieval with fluid dialogue tracking (Melchiorre et al., 2025; Kemper et al., 2024). Traditional architectures route both tasks through a single text interface, so as later turns emphasize evolving preferences, earlier entity mentions are diluted within the accumulated query text and dropped from retrieval, undermining the exact catalog matches that credible explanations require (Qin et al., 2024). To resolve this, we propose an operationally decoupled architecture that separates retrieval and response. As shown in Figure 1, these pipelines connect strictly via ranked candidates and metadata evidence, directly aligning with the ACM RecSys Challenge 2026111https://w.recsyschallenge.com/2026/ objectives of top-20 ranking and evidence-grounded response generation. Figure 2b illustrates this single-interface vulnerability. At turn 4, a listener requests “melancholic darkwave” citing Switchblade Symphony; by turn 6, they add “gothic rock” and a “heavier beat” without repeating the artist. A keyword baseline (bm25_qmr) drops the ground-truth track—“Dollhouse (Razed In Black Mix)”—to 20th place due to this entity shift. Our system instead runs a hybrid lexical–dense pool for exact matching alongside a conversational pool driven by fine-tuned Qwen 8B adapters (Figure 2a): the 8B pool alone recovers the target to fourth place, RRF of the two views lifts it to second, and LightGBM ranks it first. After ranking, generating explanations requires strict factual grounding. Instead of relying on free-form generation, we adapt an evidence-grounded PAS framework (Figure 3). PAS converts response generation into a joint reasoning step conditioned on candidate-level evidence and a matched demonstration, followed by deterministic validation to anchor outputs to verified catalog attributes. We release our full pipeline222https://github.com/yoobros/music-crs-challenge. Our main technical contributions are: • An operationally decoupled CRS framework separating retrieval from response generation. • A multi-view calibration pipeline merging hybrid and Qwen 8B adapter pools via LightGBM. • The PAS response architecture enforcing structural evidence assignment to improve explanation reliability. Dialogue + catalogrequest + historyexternal metadataTwo retrieval viewshybrid lexical–densetask-adapted conversationalMerge + LightGBMseparate ranks + scoresTop 20tracks + evidencePAScommit, then writeRanked tracks+ evidence-grounded response Figure 1. System overview. Two retrieval views merge only at calibration; PAS grounds the response in the fixed top 20. Conversation and externally enriched catalog inputs feed the bm25 qmr and qemb two-tower 8B pools. Their candidates are merged and calibrated by LightGBM, after which the top 20 and evidence flow through PAS to the ranked tracks and response. 2. Method As shown in Figure 1, our pipeline sequentially enriches the track catalog with external metadata, routes dialogue through two complementary retrieval pools, calibrates candidate rankings via LightGBM, and structures the final output using an evidence-grounded PAS framework. 2.1. Official Data and External Metadata The official TalkPlayData-Challenge dataset suite333https://huggingface.co/datasets/talkpl-ai/TalkPlayData-Challenge-Dataset supplies a 47,071-track catalog along with dialogue sessions and conversation goals. To mitigate data sparsity, we integrate external metadata from TalkPlayTools-Env444https://huggingface.co/datasets/talkpl-ai/TalkPlayTools-Env for descriptive audio, LRCLIB555https://lrclib.net/ for lyrics fallback, and MusicBrainz666https://musicbrainz.org/ for structured fields. As shown in Table 1, these sources provide complementary coverage, enriching 96.6% of the catalog with at least one external field. Table 1. Coverage of external metadata over the 47,071-track catalog. Source rows may overlap. Source Information used Tracks Coverage TalkPlayTools-Env lyrics, caption, chord, tempo, key 13,745 29.2% LRCLIB lyrics fallback 28,162 59.8% MusicBrainz IDs, release, label, country, tags 38,585 82.0% Any source at least one sourced field 45,475 96.6% For example, the track “Dollhouse” in Figure 2b expands its baseline gothic and industrial tags with external fields such as 130.21 BPM and D# minor. This enriched profile flows downstream: keywords expand BM25 documents, structured properties populate conversational track briefs, and captions and lyrics ground PAS response generation; missing fields are strictly omitted. 2.2. Pool-Specific Retrieval Interfaces Rather than issuing a single query, our architecture routes retrieval through two parallel interfaces (Figure 2). The hybrid pool (bm25_qmr) combines sparse and dense mechanisms. The sparse component runs two BM25 searches (Robertson et al., 1994): a baseline catalog search and a variant enriched with catalog tags and external keywords. Its query fuses metadata from prior music turns with filtered current-request tokens, and an RRF layer (Cormack et al., 2009) merges their rankings. The dense component uses a 0.6B Qwen3 embedding model (Zhang and others, 2025), joining historical metadata with the lowercased current request to query official document vectors. A second RRF fuses the sparse and dense rankings without score calibration, producing the hybrid pool’s candidate list. (a) Final pool topology Dialogue history + track catalog BM25: catalog + external keywords → two searches → RRF 0.6B dense: catalog strings → embedding search cross-component RRF Hybrid pool bm25_qmr Dialogue history + track catalog 8B structured query + track brief five adapted 8B searches five-fold RRF Conversational pool qemb_twotower_8b (b) Shortened turn-6 query/document views with temporal alignment Component Query-side text (Inputs) Document-side text (Target) Preserved signal Hybrid: BM25 [Turns 1–5] Burn Witch Burn; Ego Likeness; 2006 [Turn 6] gothic rock; female vocals; industrial; driving beat Dollhouse (Razed In Black Mix); Switchblade Symphony; 2001; tags: gothic, darkwave, industrial. Exact entities and current search tags. Hybrid: 0.6B dense [Turns 1–5] title: burn witch burn, artist: ego likeness [Turn 6] Lowercase text-based search tokens. Dollhouse; Switchblade Symphony; 2001; popularity 11; tags: punk, ebm, remix. Catalog semantics; official fields only. Conversational: 8B Fixed four-field structured prompt template. (Instantiated in the table below.) Dollhouse by Switchblade Symphony; 130 BPM, D# minor, gothic/darkwave. Evolving conversational intent across turns. Actual 8B Structured Query Fields for Turn-6 Execution History [Turns 1–5] Recent listens: Burn Witch Burn [T5]; In The Throat Of The Unbounded; Sorrow Is Her Name; Quiet Moments; Frozen. Session so far: User anchored in Lycia’s ethereal darkwave, expanding through Violet Tears and Ego Likeness toward gothic rock. Current [Turn 6] Looking for: gothic rock + industrial elements + heavier driving beat. Context: female vocals; Switchblade Symphony; gothic/darkwave. Figure 2. Two final-pool pipelines and temporal separation of inputs. (a) Topology of the hybrid and conversational pipelines. (b) Input-target pairs for Turn 6. The bottom subtable details the 8B model’s query structure. Panel a shows the hybrid pipeline, where BM25 and a 0.6B dense component merge through cross-component RRF into the bm25 qmr pool, and the conversational pipeline, where five adapted 8B searches merge through five-fold RRF into the qemb twotower 8B pool. Panel b lists the turn-6 query-side and document-side text for each component and the structured 8B query fields. The conversational pool (qemb_twotower_8b) relies entirely on natural language on both sides to track long-range session dynamics. The query format strictly structures the dialogue history into four ordered fields: Recent listens: lists prior tracks chronologically; Session so far: supplies an optional cached preference trajectory summary under 25 words; Context: aggregates explicit goals, eras, and popularity cues; and Looking for: contains the current explicit request. When context window limits are reached, a strict truncation priority rule preserves the user’s immediate intent: because Looking for: sits adjacent to the final embedding token, it is never dropped, whereas lower-priority historical context fields are truncated first. To optimize this contextual mapping, we fine-tune an 8B parameter encoder model using 4-bit QLoRA (Dettmers et al., 2023) under an in-batch InfoNCE contrastive objective (van den Oord et al., 2018) with LogQ frequency correction (Yi and others, 2019). The dense retrieval is executed across five session-level cross-validation folds, and their rankings are combined via RRF into the unified qemb_twotower_8b pool (bagging). Turn 1 lacks dialogue history: BM25 and dense queries collapse into the current request tokens, and the 8B query retains only the Context: and Looking for: fields. Document representations remain unchanged, and the downstream LightGBM feature row records the turn number. 2.3. Candidate Fusion and Calibration To consolidate retrieval, the system unions the top-100 candidates from each pool. A LightGBM (Ke et al., 2017) ranker then scores and calibrates this final list. Table 2 outlines the input feature schema. The model ingests individual ranks and scores from both pools, alongside their cross-pool RRF metrics. Turn number, request category, and goal specificity define the session context. Non-retrieved candidates receive sentinel values. Training uses out-of-fold scoring to prevent data leakage from the 8B retrieval model. Table 2. LightGBM feature schema. Source Fields Decision evidence Hybrid pool rank, score exact/general support Conversational pool rank, score dialogue support Cross-pool RRF rank, score retrieval agreement Context turn, category, specificity request regime Evidence Request + dialogue Top-5 details + closed top-20 title list Request evidence “driving bassline”; “dark and intense”; prior Flying Lotus preference Propose Assign each candidate to its first qualifying group under fixed precedence Candidate-level groups “Camel” similarity: 0.60 ≥ threshold → query-aligned LLM generation ONE CALL Use proposed groups + one routed demo to perform Assign + Select and emit the response Generation fields core = “Camel”; reference = prior Flying Lotus; supplement = none; theme = “darker, rawer pocket” Validate / repair Check titles, claims, supplements, and repeated phrasing Validation result title allowed; claim supported; no supplement; soft-pattern checks pass Grounded response Return the verified explanation Final excerpt “Camel” sits in a darker, rawer pocket: vibrating sub-bass and stripped-down, menacing production. Few-shot demo bank ⋅· Fig. 4 retrieve a bucket-matched example within a single LLM call Figure 3. PAS runtime execution mapped to a single-shot demonstration workflow. Candidate-level evidence groups and a bucket-matched example condition a single LLM call, followed by deterministic validation. Five stages flow left to right: evidence collection from the request and dialogue, proposal of candidate-level evidence groups, a single LLM generation call conditioned on a routed demonstration from the few-shot demo bank, deterministic validation and repair, and the final grounded response. Eligible training turns Sample completed recommendation turns Compiler input 1,000 turns, not sessions Generate + judge GLM-5.2 candidate generation Personalization, Explanation, and lexical novelty Quality gates 15–130 words; total score ≥ 0.7 Coverage selection Balance category, specificity, and turn position Selected bank 18 demonstrations Bucketed demo bank Index by category × specificity × turn Matched bucket key by A — HL — mid compiled once; reused at runtime Retrieve one demo Select one compiled response example from the matched bucket Selected example core track: “Camel” by Flying Lotus Condition Fig. 3 Insert bucket-matched example as one-shot response guidance Destination single LLM generation call COMPILE ONCE OFFLINERETRIEVE AT RUNTIME Current request signature category = A — specificity = HL — turn = mid Figure 4. Offline demonstration compilation and runtime routing mechanisms. The current request signature queries the shared 18-example bank. The A|HL|mid route then selects a matching example. An offline compile-once stage samples eligible training turns, generates and judges candidate responses, applies coverage selection, and stores 18 demonstrations in a bucketed bank. At runtime the current request signature retrieves one bucket-matched demonstration, which conditions the single LLM generation call of the PAS figure. Table 3. Public-test retrieval results; 0.6B is not task-adapted. Configuration nDCG R@20 R@100 Enriched BM25 ensemble .1599 .3321 .4793 Untrained 0.6B dense .1210 .2385 .3531 Hybrid lexical–dense .1550 .3170 .4639 8B five-fold RRF .2165 .4363 .6619 Full cross-pool RRF .2198 .4406 .6630 + LightGBM .2208 .4400 .6630 Table 4. Blind-A response check on fixed track IDs. Generator Demos Judge Composite GLM-5.2 PAS 3 4.80 .6556 Gemini-3.1-Flash-Lite PAS 3 4.75 .6518 GLM-5.2 PAS 1 4.90 .6631 Table 5. Category nDCG@20 and relative gains (%) by retrieval pool and dialogue stage. Bold marks the best retrieval result in each category. Category (N) BM25 0.6B Adapt. 8B Fusion Δcrawl _crawl Δbest _best Early Mid. Late Sound traits (61) .161 .147 .219 .216 +5.5+5.5 −1.7-1.7 −7.2-7.2 −0.4-0.4 +0.8+0.8 Lyrics (142) .178 .126 .224 .229 +6.3+6.3 +2.5+2.5 +5.5+5.5 +4.7+4.7 −1.8-1.8 Artwork (58) .139 .109 .174 .176 +2.7+2.7 +0.8+0.8 −0.8-0.8 −3.2-3.2 +7.5+7.5 Activity fit (86) .152 .119 .200 .205 +2.9+2.9 +2.3+2.3 +3.8+3.8 −2.4-2.4 +7.2+7.2 Discovery (95) .173 .127 .214 .217 +4.9+4.9 +1.1+1.1 −0.6-0.6 −1.0-1.0 +5.0+5.0 Memory (95) .193 .153 .275 .281 +3.2+3.2 +2.1+2.1 +4.0+4.0 +2.3+2.3 +0.2+0.2 Mood (77) .156 .098 .197 .201 +7.7+7.7 +1.7+1.7 −2.3-2.3 +1.4+1.4 +5.2+5.2 Artist (135) .177 .150 .242 .245 +2.4+2.4 +1.2+1.2 +2.2+2.2 −3.1-3.1 +5.2+5.2 Exact hit (18) .118 .070 .164 .180 +24.4+24.4 +10.1+10.1 +9.3+9.3 +14.7+14.7 +6.5+6.5 Trends (77) .136 .095 .211 .217 +2.9+2.9 +3.2+3.2 −1.6-1.6 +1.1+1.1 +11.6+11.6 Era (156) .132 .094 .196 .195 +7.9+7.9 −0.1-0.1 −3.0-3.0 +1.5+1.5 +1.2+1.2 2.4. PAS Response Generation We adapt the PAS framework (Wang et al., 2023) to enforce structural constraints on response claims for the top-20 ranked tracks. Figure 3 illustrates the complete runtime execution alongside a sample trace. First, the propose stage partitions the 20 candidates into distinct evidence groups under fixed precedence rules: query alignment, lyric resonance, taste continuity, pool coherence, and discovery. In the sample trace, tokens matching “driving bassline” and “dark and intense” yield a 0.60 similarity score for “Camel” by Flying Lotus, satisfying the query-alignment threshold. A single structured LLM call processes the proposed candidate groups together with one routed demonstration. This demonstration provides response selection guidance rather than introducing additional recommendation candidates. Second, within this single call, the assign step identifies usable personalization anchors and supported themes. Third, the select step chooses a core track, optional supplementary tracks for requests seeking multiple tracks, and optional historical references. The same joint call generates the raw response text. Lastly, deterministic rules validate and repair title formatting errors or citation violations. If the draft matches any blacklisted session phrases, PAS triggers a single generation retry. Figure 4 outlines the offline compilation process for the reusable demonstration bank. The compilation pipeline samples 1,000 eligible recommendation turns from the TalkPlayData training split. GLM-5.2 generates the initial candidate responses. An independent LLM-as-judge scores personalization and explanation metrics on a 1-to-5 scale. We normalize these outputs into a unified quality metric. The final score weights this quality metric at 75% and lexical novelty at 25%. Viable candidates must contain between 15 and 130 words and achieve a minimum combined score of 0.7. A coverage-aware filtering step balances the final selection across request category, specificity, and turn position. This selection process yields a fixed 18-example runtime demonstration bank. At runtime, each incoming request derives a routing key based on its category, specificity, and turn. The router evaluates the exact matching bucket first. If that bucket is empty, the system sequentially falls back to category matches, specificity or turn matches, and the global bank. The router injects the highest-scoring demonstration from the first nonempty layer. Figure 4 displays the key A|HL|mid, where A represents a sound-focused category, HL encodes high goal specificity with low track specificity, and mid indicates a mid-dialogue turn. The trace shown in Figure 3 instantiates this routed example. 3. Experiments 3.1. Evaluation Setup We score 8,000 turns from the 1,000-session public test split against all 47,071 tracks using nDCG@20, R@20, and R@100. Blind-A and Blind-B contain 80 truncated sessions each and are used only for final external checks. Response ablations fix track IDs. In Table 4, Judge is the score returned by the organizers’ undisclosed LLM judge, and Composite is the official aggregate score. 3.2. Results and Controlled Ablations Table 4 summarizes public-test retrieval results. Individual 8B folds range from .2018 to .2138 in nDCG@20, from .4100 to .4298 in R@20, and from .6369 to .6435 in R@100. Five-fold RRF provides the largest ranking gain. It improves over the mean individual fold by .0076 nDCG@20 and raises R@100 from .6409 to .6619. The pools recover different targets. At least one pool places the target in the top 20 on 3,838 turns. Both succeed on 2,188; the conversational pool alone succeeds on 1,302, and the hybrid pool alone succeeds on 348. Their top-20 candidate sets have a mean Jaccard overlap of .231. This complementarity makes fusion useful: cross-pool RRF raises nDCG@20 from .2165 to .2198. LightGBM further raises it to .2208, while R@20 changes from .4406 to .4400 and R@100 remains .6630. These mixed changes support interpreting LightGBM as candidate-order calibration rather than as a broad retrieval gain. With tracks fixed, Table 4 isolates response generation. Among GLM-5.2 runs, one routed demonstration outperforms three. On Blind-B, PAS raises the official composite from .55 to .58. This supports the routed one-shot choice under the automatic judge but does not support a claim about human preference. 3.3. Pool Behavior by Request and Turn Table 5 tests whether fusion helps uniformly. For each subset, Δbest _best is 100(Fusion/best−1)100(Fusion/best-1), where best is the strongest of BM25, 0.6B, and 8B. Δcrawl _crawl is 100(BM25/BM25nocrawl−1)100(BM25/BM25_no\ crawl-1). Early, Mid., and Late report Δbest _best for turns 1–2, 3–5, and 6–8; N is the number of sessions. Bold values in Table 5 identify each category’s best result. The adapted 8B pool is the strongest individual pool in all eleven categories; fusion is best in nine. Its benefit appears more often later in the dialogue: fusion improves over the best individual pool in five categories during early turns, six during middle turns, and ten during late turns. Sound-focused and era-focused requests are the two aggregate exceptions. Crawl-derived fields raise enriched BM25 from .1524 to .1599 nDCG@20, an aggregate 4.9% gain, with category-level gains ranging from +2.4% to +24.4% (Table 5). The untrained 0.6B dense component does not improve BM25 on the public test split; adding it lowers the hybrid result from .1599 to .1550. Replacing BM25’s music-only query with the full dialogue further lowers hybrid nDCG@20 to .1527. The released BM25 interface therefore retains prior track metadata and the current request but excludes earlier user and assistant text. 4. Analysis and Limitations Three design lessons emerge. Tailoring input text to each scoring mechanism lets the hybrid interface exploit exact catalog cues while the conversational interface tracks evolving language, as Figure 2 shows. The low candidate overlap and complementary target hits between these channels create productive pool disagreement, explaining why cross-pool RRF outperforms single-pool baselines (Section 3). Finally, explicit evidence and role assignments let PAS prioritize evidence-backed selection over surface realization. These findings point toward extending the LightGBM ranker, which currently relies on only two upstream pools and compact features; stronger retrieval components and richer track–dialogue features would give the calibrator more informative signals and further extend its reranking gains. Several limitations remain. Uneven external metadata can constrain PAS despite deterministic validation, and the in-house judge may introduce model bias during offline compilation. The 18-example synthetic bank also cannot cover every dialogue state or sparse request bucket, and the five adapted 8B folds plus LLM-based synthesis require substantial compute. This study therefore leaves runtime latency, serving cost, and human preference unevaluated. 5. Conclusion Isolating retrieval from response generation optimizes recommendation accuracy and system transparency. Pool-specific text interfaces safeguard exact catalog cues, cross-pool RRF captures complementary ranking signals, and the PAS framework prevents ungrounded claims through structured constraints. Empirical evaluation confirms that external crawl enrichment drives performance gains across all request categories, while fusion delivers its largest benefits in late-dialogue turns. This clear separation of concerns supported our third-place finish in the Blind-B industry track, proving that decoupling retrieval from surface realization maintains competitive performance while keeping all decisions fully inspectable. References G. V. Cormack, C. L. A. Clarke, and S. Buettcher (2009) Reciprocal rank fusion outperforms condorcet and individual rank learning methods. In Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval, Cited by: §2.2. T. Dettmers, A. Pagnoni, A. Holtzman, and L. Zettlemoyer (2023) QLoRA: efficient finetuning of quantized llms. In Advances in Neural Information Processing Systems, Cited by: §2.2. G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T. Liu (2017) LightGBM: a highly efficient gradient boosting decision tree. In Advances in Neural Information Processing Systems, Cited by: §2.3. S. Kemper, J. Cui, K. Dicarlantonio, K. Lin, D. Tang, A. Korikov, and S. Sanner (2024) Retrieval-augmented conversational recommendation with prompt-based semi-structured natural language state tracking. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, p. 2786–2790. External Links: Document Cited by: §1. A. B. Melchiorre, E. V. Epure, S. Masoudian, G. Escobedo, A. Hausberger, M. Moussallam, and M. Schedl (2025) Just ask for music (JAM): multimodal and personalized natural language music recommendation. In Proceedings of the Nineteenth ACM Conference on Recommender Systems, p. 615–620. External Links: Document Cited by: §1. P. Qin, C. Huang, Y. Deng, W. Lei, and T. Chua (2024) Beyond persuasion: towards conversational recommender system with credible explanations. In Findings of the Association for Computational Linguistics: EMNLP 2024, p. 4264–4282. External Links: Document Cited by: §1. S. E. Robertson, S. Walker, S. Jones, M. M. Hancock-Beaulieu, and M. Gatford (1994) Okapi at trec-3. In Proceedings of the Third Text REtrieval Conference, Cited by: §2.2. A. van den Oord, Y. Li, and O. Vinyals (2018) Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748. Cited by: §2.2. Z. Wang, J. Shang, and R. Zhong (2023) Goal-driven explainable clustering via language descriptions. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, p. 10626–10649. External Links: Document Cited by: §2.4. X. Yi et al. (2019) Sampling-bias-corrected neural modeling for large corpus item recommendations. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Cited by: §2.2. Y. Zhang et al. (2025) Qwen3 embedding: advancing text embedding and reranking through foundation models. arXiv preprint arXiv:2506.05176. Cited by: §2.2.