Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• Hi-SAM's Disentangled Semantic Tokenizer (DST) unifies modalities via geometry-aware alignment and quantizes them via a coarse-to-fine strategy, using shared codebooks to distill consensus and modality-specific ones to recover nuances from residuals.method 90%
• Hi-SAM's Hierarchical Memory-Anchor Transformer (HMAT) splits positional encoding into inter- and intra-item subspaces via Hierarchical RoPE to restore hierarchy in user interactions.method 90%
• Hi-SAM achieved a 6.55% gain in the core online metric when deployed on a large-scale social platform serving millions of users.result 95%
• Hi-SAM shows consistent improvements over state-of-the-art baselines in experiments on real-world datasets, especially in cold-start scenarios.result 90%
• Existing semantic ID-based methods (e.g., RQ-VAE) lack disentanglement between shared cross-modal semantics and modality-specific details, causing redundancy or collapse.limitation 85%
• Vanilla Transformers treat semantic IDs as flat streams, ignoring the hierarchy of user interactions, items, and tokens, which biases attention toward local details over holistic semantics.limitation 85%
• Hi-SAM's performance may degrade if multi-modal input data is sparse or inconsistent across modalities.limitation 80%
• Hi-SAM's Hierarchical Memory-Anchor Transformer inserts Anchor Tokens to condense items into compact memory, retaining details for the current item while accessing history only through compressed summaries.technical 90%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card