Research map / Vision-language and multimodal / Multimodal LLMs
IntentLens: Grounding Underspecified Multimodal Queries for Recommendation via Tool-Augmented Reasoning
xiao chen, Dong Fang, HAITAO LI, Qing Li
Read the paperCheck your idea against 137,217 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.
Proposes IntentLens, a tool-augmented reasoning approach for grounding underspecified multimodal queries (reference image plus fuzzy text) into recommendations.
Problem
Recommenders struggle when users give only a reference image and a fuzzy natural-language query, leaving visual attributes, personal preferences, and situational constraints implicit.
Approach
Uses tool-augmented reasoning to ground implicit intent from underspecified multimodal evidence before recommending; details not stated in the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add tool-augmented grounding (ground-then-rank) to MLLM recommender (ablation) | IntentLens (7B) · Google Review-MMTool | MRR: 33.5 -> 40.2 (w/o grounding) | Improved |
| add visual tools (ablation) | IntentLens (7B) · Google Review-MMTool | MRR: 35.7 -> 40.2 | Improved |
| add user memory tools (ablation) | IntentLens (7B) · Google Review-MMTool | MRR: 38.1 -> 40.2 | Improved |
| add item attribute tools (ablation) | IntentLens (7B) · Google Review-MMTool | MRR: 38.0 -> 40.2 | Improved |
| add ranker-aligned reinforcement fine-tuning over SFT (ablation) | IntentLens (7B) · Google Review-MMTool | MRR: 33.1 -> 40.2 | Improved |
| add counterfactual no-tool baseline to reward (ablation) | IntentLens (7B) · Google Review-MMTool | MRR: 35.0 -> 40.2 | Improved |
| add item-side tool-selection reward (ablation) | IntentLens (7B) · Google Review-MMTool | MRR: 35.5 -> 40.2 | Improved |
| mask observation tokens in GRPO losses (ablation) | IntentLens (7B) · Google Review-MMTool | MRR: 35.1 -> 40.2 | Improved |
| scale backbone from 3B to 7B | IntentLens · Google Review-MMTool and Yelp-MMTool | MRR: 34.5 -> 40.2; 33.5 -> 46.7 | Improved |
| use predicted grounded query instead of raw query (ablation) | IntentLens ranker · Google Review-MMTool | MRR: 33.6 -> 40.2 (oracle-grounded 44.4) | Improved |
| evaluate on unseen categories, cross-dataset, unseen tools | IntentLens (7B) · Google and Yelp | MRR: Google 40.2 -> 36.2 / 34.7 / 35.8 | Hurt |
| use IntentLens instead of GPT-5.4 with evidence on Google | IntentLens (7B) · Google Review-MMTool | Hit@1 / MRR: Hit@1 31.0 vs 31.0; MRR 40.2 vs 41.6 | No clear effect |
Closest papers by meaning
- VisualLens: Personalization through Task-Agnostic Visual History (NeurIPS 2025 · similarity 0.89)
- DOGE: LLMs-Enhanced Hyper-Knowledge Graph Recommender for Multimodal Recommendation (AAAI 2025 · similarity 0.88)
- MTSTRec: Multimodal Time-Aligned Shared Token Recommender (ICML 2025 · similarity 0.86)
- Semantic Convergence: Harmonizing Recommender Systems via Two-Stage Alignment and Behavioral Semantic Tokenization (AAAI 2025 · similarity 0.86)
- Harnessing Multimodal Large Language Models for Multimodal Sequential Recommendation (AAAI 2025 · similarity 0.86)
- UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
