Four papers, read for what they mean for PyrusLLM
Four independent groups, four disciplines — systems, cryptoeconomics, mechanism design, control theory — published between January 2025 and January 2026. None cite each other. Read together, they sketch the state of the art for decentralized LLM inference from every angle except one, and the missing angle is the one PyrusLLM is built around.
The four papers
PolyU Hong Kong + China Mobile HK, arXiv 2510.02395, Oct 2025. A decentralized LLM inference marketplace with a trustless quality-verification protocol (TIQE) and a three-role incentive model — the closest thing to PyrusLLM published as an academic system, with a real 20-device deployment.
UC Berkeley (Dawn Song's group), arXiv 2501.14784, Jan 2025. Gives up on winning on latency and competes on offline batch throughput instead — 445 tok/s for a 70B model, pipelined across a high-latency WAN, with an optimistic verification scheme that costs one signature.
University of Exeter + HUST, arXiv 2503.04521, in review at IEEE TMC. An auction mechanism for edge inference pricing with proven truthfulness and envy-freeness, clearing in ~3 milliseconds — small enough to run inside a confidential enclave without discussion.
FLock.io + University of Oxford, arXiv 2601.09961, Jan 2026. A PID controller for token buyback-and-burn, positioned to stabilize the token that pays for all of the above — undercut by a simpler alternative the paper never considers.
The synthesis, in one paragraph
The three papers that have to match a request to a machine all leave that role centralized — two admit it in writing. AERIA is the most revealing: it doesn't leave the central coordinator in by accident — the auctioneer is the design, and the paper details exactly what it needs to know to clear the market: budget, deadline, accuracy requirement, and which model each user wants. That isn't an accidental leak. It's the algorithm's required input. It can't be encrypted, because the algorithm has to read it in the clear. It can only be moved somewhere nobody watching can see it — which is what a confidential-compute enclave is for, and is the thesis PyrusLLM is built on.
Where the four line up
| PolyLink | DeServe | AERIA | DCBM | |
|---|---|---|---|---|
| Layer | quality & incentives | systems & cost | price discovery | monetary policy |
| Throughput / systems | weak (7 tok/s, sharded) | strong (445 tok/s) | n/a (CPU sim) | n/a |
| Provider economics | inverted incentive | real cost table | reserve price + margin γ | n/a |
| Quality / reputation | TIQE, measured | absent | declared σ, unverified | n/a |
| Verification | committee, 30% permanent tax | optimistic, ~0% tax | absent | absent |
| Price discovery | undefined δ | fixed reference price | proven auction | stabilizes the unit |
| Privacy | none | none (linkability is a stated feature) | none — defines what the auctioneer must see | n/a |
| Decentralized matching | no (cloud API server) | no (admitted open problem) | no (the auctioneer is the design) | n/a |
| Real deployment | 20 devices, 4 cities | GCP east–west | simulation only | simulation only |
| Public code | yes | yes | yes | no |
| Applies to autoregressive LLMs | yes | yes | no — single-forward-pass cost model | n/a |
Sources
| Paper | arXiv | Venue / date | |
|---|---|---|---|
| PolyLink | 2510.02395v1 | cs.CR, 1 Oct 2025 | 2510.02395v1.pdf |
| DeServe | 2501.14784v1 | cs.DC, 4 Jan 2025 | 2501.14784v1.pdf |
| AERIA | 2503.04521v2 | in review, IEEE Trans. Mobile Computing | 2503.04521v2.pdf |
| DCBM | 2601.09961v1 | cs.GT, 15 Jan 2026 | 2601.09961v1.pdf |