RARS training objective teaches generative retrievers multiresolution relevance at zero inference cost
_reachsumit · x · 2026-10-01
- An arXiv paper introduces RARS (Resolution-Aligned Relevance Supervision) for generative retrieval with semantic identifiers (SIDs).
- It formulates multiresolution relevance as conditional distributions induced by a single document-level relevance measure across the SID hierarchy, supervising a shared query representation at each refinement level.
- A prefix-conditioned predictor allocates relevance among sibling branches during training, weighted by relevance mass reaching each parent; the predictor is discarded at inference, preserving standard autoregressive retrieval with no inference cost.
- Consistent improvements on three multilingual ESCI locales.
More from Research
- Protein binding folding models face data starvation after PDB has been fully juiced — anshulkundaje · 2026-10-01
- Stanford's UniEvo-VL Self-Distillation Lifts Qwen-image GenEval From 0.747 to 0.808 — stanfordnlp · 2026-10-01
- NVIDIA's Mid-Harness Scales Actions at the Model-Harness Boundary, Lifting TerminalBench Pass@1 to 68.03% — nvidia · 2026-10-01
- Amazon's SMART Self-Evolving Multi-Agent System Tops All 15 Subtitle Arena Directions, Cuts Penalty 6.9% — amazon · 2026-10-01
- Meta's Loop Scaling Laws: Sparsity Gives ~3x Active-Param Efficiency, Recurrence ~2x on Reasoning — facebook · 2026-10-01
- AI Protein Design Still Can't Solve Binder Prediction and the Age-Old Docking Problem, Researcher Explains — anshulkundaje · 2026-10-01