The Ledger
Researchers update Angelfish consensus preprint
A new version of the Angelfish research paper describes a consensus protocol that shifts between leader-based and DAG-based operation. In tests reported by the authors, it reduced latency against Sailfish under moderate throughput while matching its peak throughput.

A consensus design between leaders and DAGs
An updated arXiv paper proposes Angelfish, a Byzantine fault-tolerant consensus protocol designed to shift between leader-based and directed acyclic graph (DAG)-based operation as transaction demand changes. The paper was first posted in September 2025; the current arXiv version is version 4, dated November 17, 2026, and lists Qianyu Yu, Giuliano Losa, Nibesh Shrestha and Xuechao Wang as authors.
The paper describes a trade-off between leader-based protocols, which tightly couple transaction dissemination and consensus, and DAG-based designs that let participants disseminate data in parallel. Angelfish combines the approaches: a dynamically adjusted subset of participants can send lightweight votes using best-effort broadcast instead of reliably broadcasting more costly DAG vertices. The round leader still creates a vertex, while other participants can vote rather than contribute vertices in a given round.
The protocol can operate at either end of this design range, or between them, by changing how many participants propose vertices. The researchers say a participant with few transactions to propose can send a vote that uses less communication and takes fewer message delays than broadcasting a vertex. Votes do not carry transaction data, so the protocol uses them to help advance consensus rather than to add transaction throughput.
Tests report lower latency at moderate load
The evaluation ran Angelfish on a geo-distributed testbed with 50 nodes, comparing it with Sailfish and Autobahn. The authors deployed nodes across five Google Cloud regions; the paper says their experiments used transactions of 512 bytes and measured latency from transaction creation until all non-faulty nodes committed it.
At a 40% vertex-proposal rate, the paper reports that Angelfish reduced latency against Sailfish by at least 40% in a moderate-throughput range of 30,000 to 100,000 transactions per second. The comparison describes latency falling from roughly 1.0 seconds to 0.6 seconds, while peak throughput matched Sailfish. In another reported result, Angelfish processed roughly 27,000 transactions per second at 0.5-second latency with a 40% proposal rate.
The researchers also tested a multi-leader version with 10 and 20 leaders per round. The paper’s introduction reports that with 20 leaders and a 40% vertex-proposal rate, this variant reached a latency of roughly 300 milliseconds while serving 25,000 transactions per second, which it says is only 36% slower than the leader-based protocol Hydrangea ordering empty blocks. That baseline, the paper notes, was configured without a mempool.
Scope and remaining work
The evaluation compared Angelfish mainly with Sailfish and Autobahn. The authors say they did not compare it with Mysticeti because its implementation uses a substantially different codebase, and they report that no implementation of a state-of-the-art leader-based protocol using asynchronous information dispersal was available for their comparison.
The paper also reports tests involving slow nodes and crash failures, alongside its fault-free performance evaluation. Its conclusion says that the lightweight-vote design might be applied to other DAG-based protocols, but each would need a separate safety argument based on its own commit rule, timeout mechanism and DAG-validity conditions. A systematic exploration of those design choices is left for future work.
Sources
- Original source · Preprint Angelfish: Leader, DAG, or Anywhere in Between arXiv:2509.15847
Related stories
- Category Labs Researchers Propose Quantum Cadence Consensus
- Researchers map trade-offs across Tendermint consensus designs
- Preprint proposes fee-memory rule for Layer 2 posting costs
- Lightning force-closes cost 2–6 times more in fees, preprint finds
- Preprint proposes one-transaction proofs for Ethereum quantum migration