In briefVerifiable agent insurance: making AI agents financially accountable after misbehavior
SummaryThe paper proposes verifiable agent insurance (VAI), in which agent providers financially commit to programmatic policies through a smart contract with escrowed funds on an immutable ledger. Users who detect violations can have a verifier review execution traces, and confirmed violations trigger payment from escrow to the reporting user. Implemented on an Ethereum layer-2 network and tested on three benchmarks, VAI detected violations in 42.2% of violating runs, rising to 83.0% when agents emit explicit structured authorization and security records.
In briefInterpretable quantum generators for financial returns via problem-informed circuit design
SummaryThe paper proposes a framework that uses relational structure and symmetry-inspired parameter sharing to build quantum circuits with few trainable parameters. Three quantum generator architectures are tested within a hybrid Wasserstein quantum GAN to generate 16-step Bitcoin return windows. Single-layer circuits remain competitive with deeper ones, and ZZ-CRX shows a downward shift in median leverage relative to ZZ at matched depths.
In briefFLAI Protocol: Blockchain-Rewarded Decentralized Federated Learning with Contribution Scoring
SummaryThis paper presents the FLAI Protocol, a decentralized federated learning framework that lets parties train models collaboratively without centralizing their data. It combines aggregation verified inside a trusted execution environment, a sparse Peer Evaluation Graph (sPEG) scoring method for participant contributions, and on-chain settlement of token rewards. Evaluations on image, text, and healthcare datasets show utility comparable to FedAvg, with improved fairness and accountability.
In briefTransferable solvation corrections for machine-learning potentials in solution chemistry
SummaryMost machine-learning interatomic potentials are trained on gas-phase data, limiting their use for solution chemistry. The paper introduces AnySolv, a transferable implicit solvation potential that adds solvent-dependent energy and force corrections to existing gas-phase potentials without retraining. It is trained on 1.4 million DFT-SMD solvation energies and forces across 36 solvents and supports 179 solvents, including unseen ones.
WhyProposes a transferable solvation correction added to gas-phase potentials without retraining, validated on large DFT-SMD data and experimental solvation free energies.
In briefEvoQuantLab: Training-free dual-loop evolution for auditable quant research
SummaryThis paper presents EvoQuantLab, a training-free framework that uses a dual-loop design to automatically discover quantitative alpha factors. An inner genetic-programming loop evolves typed factor expressions, while an outer loop audits evidence across cycles to refine the protocol, and multi-metric validation filters out spurious factors. Out-of-sample tests on daily CSI 500 and S&P 500 data and hourly cryptocurrency data report annualized returns of 18.34%, 21.30%, and 37.91%, respectively.
