Google paper frames LLM routing as Pandora's Box costly-search problem, with closed-form value-of-information policies
_onionesque · x · 2026-10-10
A new paper by Adam Fisch, Jacob Eisenstein et al. (arXiv:2608.20316) formalizes multi-model routing as the classic Pandora's Box problem: estimating each specialist model's expected value has a cost, with cheap estimators (embeddings) fast but noisy and accurate ones (retrieval-augmented or partially-reasoning fine-tuned models) expensive. Under a Gaussian signal model, the resulting policies have closed-form value-of-information expressions deciding when refining estimates pays off. They propose the centralized Pandora's Router and a decentralized Pandora's Bidder where specialists self-assess before claiming queries at an offered price, validated on multi-LLM benchmarks, retrieval-augmented specialists, and variable inference-time reasoning. The post also advertises a 16-week 2027 student researcher position in e-values, e-processes, and sequential decision theory.
More from Research
- 20-minute explainer breaks down how Tesla trains FSD: 8 cameras, 36fps, 2B signals to 2 outputs — PTrubey · 2026-10-10
- 21 researchers release white paper on Visual General Intelligence as a path to AGI — HirokatuKataoka · 2026-10-10
- Epoch AI: Over half of arXiv math papers in 3 subfields now acknowledge AI use — burny_tech · 2026-10-10
- Researcher faults standard Transformer diagrams for hiding self-attention as a black box — PlisSergey · 2026-10-10
- Xiaomi's MiMo-V2.6: agents run their own RL loop, DeepSWE score hits 72.6 for $2.6M — rohanpaul_ai · 2026-10-10
- New fastest deterministic 3SUM algorithm hits n^1.9961, matching randomized bound — basedjensen · 2026-10-10