Fri 13 Nov
LiveGet it free
ZOTPAPER

News without the noise

← All concepts

model routing

When you have many AI models to choose from, model routing decides which one to use for each incoming request. Instead of always picking the biggest or cheapest model, it examines the request and picks the one that balances speed, cost, and accuracy. Some systems route to different thinking modes, such as direct answering or step-by-step reasoning, or even assign different models to different steps of a task.

The papers all find that intelligent routing dramatically improves efficiency while preserving or even boosting performance. They tackle challenges like sparse data, costly evaluations, and long conversation histories. Common solutions include learning shared representations, using reinforcement learning to adapt routing decisions, and caching past solutions. A key insight is that routing must consider not just model choice but also context construction and the cost of evaluating options.

Papers
82
Released code
6
First seen
June 2026
Latest
Nov 2026

75 papers in the last two months, against 5 in the two before.

Start here

Three papers to read first, in the order they came out
  1. 1Language models’ own confidence becomes harder to outperform as they improve

    Performance prediction only beats a model's own confidence when the model is suboptimal or the predictor adds significant capacity.

    Oct 2026
  2. 2AI reading 3D scenes gives steadier answers when the camera moves

    A geometry-conditioned rotation of visual features makes 3D vision language models answer more consistently when the camera viewpoint changes.

    Nov 2026
  3. 3Small model predicts per-query cost of AI reasoning models, saving up to a third on compute

    A small frozen LLM's prefill activations predict both success probability and output token count for routing reasoning models, cutting cost by up to a third.

    Nov 2026

How fast it is moving

JunJulAugSeptOctNov so far

Often studied with

confidence calibrationConfidence calibration aligns a model's predicted confidence with its actual accuracy across predictions.11 togethermixture of expertsA neural architecture with many specialized sub-networks, only a few activated per input via a router, balancing capacity and efficiency.11 togetherlarge language modelA neural network trained on massive text corpora to generate coherent text and perform diverse language tasks.8 together

Who is working on it

Amazon 5Shanghai Jiao Tong University 5University of Illinois Urbana-Champaign 3Tsinghua University 3Apple 3Georgia Institute of Technology 2

The papers

Most central to this idea first, not most recent.

Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.