Research map / Architectures and efficiency / Attention and transformers

MARS: Multi-resolution Adaptive Routing for Sequential Recommendation

NeurIPS 2026 · Architectures and efficiency › Attention and transformers · approach: State space and recurrent sequence models

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.

Proposes MARS, a multi-resolution cached user memory that writes history into half-life-anchored recurrent tracks and reads them via sparse seed-to-track routing, improving CTR prediction over VISTA and others while keeping a candidate-independent serving interface.

Problem

Summarization-based sequential recommenders compress long user histories into a single cached memory, which entangles short-, medium- and long-term behavioral signals ("temporal aliasing"), limiting accuracy for long-history CTR prediction.

Approach

Write the full history into multiple gated rank-one recurrent state tracks, each with a different half-life prior, and let each memory seed sparsely route (Top-k with Gumbel/straight-through) to a subset of tracks, yielding a fixed-size cached memory scored by an unchanged target-attention scorer.

Key results

• On KuaiRec (longest histories), MARS raises AUC from 0.793 (VISTA) to 0.804 with NE 0.791 vs 0.819.
• Gain over VISTA grows with history length: +0.0005 short, +0.0073 medium, +0.0114 long bins on KuaiRec.
• Warm-cache end-to-end latency for 1,000 candidates per user is 15.48 ms for MARS vs 15.15 ms for VISTA (about 1.02x).

Surprising findings

• Simply increasing VISTA's memory-side capacity (wider, deeper, more heads) raises AUC only to at most 0.7949, far below MARS, indicating structure rather than capacity drives the gain.
• A single-resolution bank with routing adds little (+0.002 AUC), showing temporal diversity is needed for routing to help.

Limitations

• Evidence is offline and scoped to CTR-style prediction with cached memory.
• Gains on Amazon-Electronics with short histories are small (comparable to VISTA).
• Industrial results are single-run and do not quantify seed variability.

Experimental setup

Datasets: Amazon-Electronics KuaiRand-1K KuaiRec industrial CTR dataset (four tasks, anonymized)

Benchmarks: Amazon-Electronics KuaiRand-1K KuaiRec

Metrics: AUC normalized entropy (NE) linear probe R^2 latency

Baselines: DIN DIEN TTSN MHA SASRec RetNet Mamba4Rec HSTU VISTA

Builds on: VISTA gated linear attention Mamba HSTU

Closest papers by meaning

• UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence (NeurIPS 2026 · similarity 0.87)
• LLM4RSR: Large Language Models as Data Correctors for Robust Sequential Recommendation (AAAI 2025 · similarity 0.87)
• SLMRec: Distilling Large Language Models into Small for Sequential Recommendation (ICLR 2025 · similarity 0.86)
• LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.86)
• LIME: Link-based User-item Interaction Modeling with Decoupled XOR Attention for Efficient Test Time Scaling (NeurIPS 2026 · similarity 0.86)
• Iterative Sparse Attention for Long-sequence Recommendation (AAAI 2025 · similarity 0.85)