Routing Queries to Multiple Embedding Models Without Curse

👁 116 reads

New research from MIT and collaborators addresses a practical but under-studied problem: how recommendation systems should route queries across multiple embedding models efficiently.

The work comes from Yan Dai, Negin Golrezaei, and Patrick Jaillet and appeared on arXiv on June 12, 2026 (revised September 24, 2026). Recommendation systems use multiple embedding models simultaneously and must decide which model should handle each query. The authors write explicitly that despite the practical importance of this problem, it remains poorly studied under realistic conditions.

The Core Problem

Embedding model routing works like this: a system receives a query (context) and must select an item (action) using one of the available embedding models (experts). Each expert operates in a low-dimensional hidden representation space - hence the term “low-rank experts”. The conditions are hostile: queries may be adversarially chosen, the system sees only limited bandit feedback, and there’s no full visibility into how the models work.

The authors first show that standard regret measures don’t work well here - they either suffer from structural misspecification or are statistically unsolvable. This is an important observation because it means you can’t simply reach for off-the-shelf tools from the contextual bandits literature.

The Proposed Solution

The authors identify a policy class they call log-quadratic. They argue it’s expressive enough to capture query-dependent routing while being structured enough to enable efficient online learning. On this foundation, they build a policy gradient algorithm called Hypentropy Policy Gradient, or HPG for short.

HPG has several claimed properties. First, it adapts to the unknown low-rank structure under partial information. Second, it achieves policy regret on the order of O-tilde of s times the square root of M times T, where s is the true rank of the experts, M is the number of models, and T is the number of rounds. The authors emphasize that this result avoids the curse of dimensionality - the cost doesn’t grow with the full dimension of the embedding space, only with the actual rank. Third, HPG’s implementation is computationally efficient and parameter-free.

What’s Missing from the Paper

The abstract and arXiv page contain no experimental results or comparisons with other methods - the paper presents only theoretical claims and their justification. There’s also no information about what real data or systems the authors tested the approach on, nor any numbers illustrating practical effectiveness. It’s also unclear whether code is publicly available - the code section on the arXiv page doesn’t point to any repository.

Formal Context

The work is classified in three arXiv categories: machine learning (cs.LG), artificial intelligence (cs.AI), and machine learning in statistics (stat.ML). The authors note that the problem of optimizing log-quadratic policies under bandit feedback is interesting in its own right, independent of the routing application.

What This Means

In our view, this is solid theoretical work that solves a well-defined mathematical problem. The practical significance depends on how large the actual ranks of embedding models are in production systems - if s is large, the theoretical guarantee loses its appeal. The lack of experiments means we don’t yet know whether HPG works better than the simpler heuristics engineers use today. It’s worth watching to see if a version with empirical validation appears.

Source: arXiv cs.AI: Policy Regret for Embedding Model Routing: Contextual Bandits with Low-Rank Experts (document from 2026-09-28). This overview is based on the full text of the paper, not other media reports.

// read moreEmbeddingGemma 2: Multimodal Embedding Model ReleasedNewsEmbeddings and Vector Databases Explained (Polish Test)GuidesNash-MOPD: Balancing Training Across Multiple TeachersNews

Kinetyra