Research map / Large language models / Retrieval and knowledge

TRACE: Tourism Recommendation with Accountable Citation Evidence

NeurIPS 2026 · Large language models › Retrieval and knowledge · approach: Domain-specific QA datasets · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces TRACE, a 10,000-dialogue Yelp-based tourism CRS benchmark with review-span citations and rejection-repair turns, evaluated with 25 metrics across accuracy, grounding, and recovery over 14 baselines.

Problem

Existing conversational recommender benchmarks score only Recall@k over entity mentions and do not check whether recommendations are backed by verifiable review evidence or whether systems recover after rejection, which matters in high-stakes tourism recommendation.

Approach

Each item is a multi-turn tourism dialogue with verbatim review-span citations and explicit rejection turns, scored jointly on accuracy, grounding (CGS, citation density, NLI entailment) and recovery (rejection-repair rate, turns-to-first-correct), revealing a Three-Competency Gap across baselines.

Key results

• LLM Zero-Shot (gpt-5.4-mini) achieves closed-set Recall@1 0.533 and rejection recovery 0.988, but CGS only 0.636 versus 0.864 for Dense retrieval.
• Grounding Score agrees with human citation precision at Spearman rho=+0.80 across pooled annotations.
• Multi-Review Synthesis collapses on grounding and recovery (CGS 0.008, rejection recovery 0.263).

Surprising findings

• Injecting an itinerary proximity graph into LLM Zero-Shot does not raise accuracy, suggesting a representational-vs-algorithmic gap in spatial reasoning.
• Fuzzy grounding score near 1.0 for non-LLM retrievers is largely vacuous, with entailment grounding falling to 0.110-0.193.

Limitations

• Synthetic dialogues generated by a single model (gpt-5.4-mini); controlled multi-LLM leaderboard not run.
• Human agreement is modest (Krippendorff alpha 0.31-0.58) and based on only 90 items.
• English-only Yelp data across 8 U.S. cities; recovery finding in a Gemini rerun uses a different KB, so it is suggestive only.

Experimental setup

Datasets: Yelp Open Dataset (Academic)

Benchmarks: TRACE

Metrics: Recall@1 Recall@3 MRR BLEU-4 ROUGE-L CGS Citation Density Grounding Score Entailment Grounding Provenance Coverage

Baselines: Popularity TF-IDF Dense Spatial Itinerary LLM Zero-Shot DST RAG-Citation Multi-Review Synthesis Memory-Augmented

Builds on: DSPy ALCE all-MiniLM-L6-v2

Closest papers by meaning

• ESPRESSO: An Effective Approach to Passage Retrieval for High-Quality Conversational Recommender Systems (AAAI 2025 · similarity 0.84)
• SPOT-Trip: Dual-Preference Driven Out-of-Town Trip Recommendation (NeurIPS 2025 · similarity 0.84)
• ORBIT - Open Recommendation Benchmark for Reproducible Research with Hidden Tests (NeurIPS 2025 · similarity 0.82)
• Ready for You When You Are Back: Content-Driven Session-Based Recommendation for Continuity of Experience (AAAI 2025 · similarity 0.82)
• Re2LLM: Reflective Reinforcement Large Language Model for Session-based Recommendation (AAAI 2025 · similarity 0.82)
• RecFlow: An Industrial Full Flow Recommendation Dataset (ICLR 2025 · similarity 0.81)