In briefFeature-Context Consistency: label-free, drift-stable adversarial detection on attributed graphs
SummaryThis paper addresses detecting adversarial nodes on evolving attributed graphs, where labels are scarce and data distributions drift over time. It proposes measuring feature-context consistency, the alignment between a node's attributes and its neighborhood's aggregated attributes, via a Context Boost score within a Restricted Boltzmann Machine framework (CB-RBM). The authors prove separation guarantees between legitimate and adversarial nodes and a Wasserstein-Lipschitz bound under drift. On Ethereum phishing and Reddit banned-user datasets, CB-RBM reports AUROC of at least 0.96 with zero labels and keeps the false positive rate near its 5% target under drift.
In briefAsymmetric q-Gaussian distributions derived from Tsallis maximum entropy
SummaryThis paper proposes an asymmetric q-Gaussian distribution within non-extensive statistical mechanics. Standard q-Gaussians are symmetric and cannot capture the asymmetric fluctuations seen in empirical data. The authors keep Tsallis entropy and impose a global escort mean together with separate downside and upside escort semi-variances, reformulating the problem as maximizing a strictly concave functional. Monte Carlo experiments and in-sample fits to equity index, commodity, and cryptocurrency returns indicate that the asymmetric model fits better than the symmetric q-Gaussian, most clearly in the tails.
Francesco Cesarone, Massimiliano Corradini, Nicol\`o Giunta et al.
In briefDemand under a mechanical price: how inelastic Ethereum data-availability demand is
SummaryEthereum sells data-availability capacity to rollups as "blobs" priced by a controller that adjusts fees with cumulative usage rather than through an auction. Using the Pectra upgrade of May 7, 2025, which doubled blob capacity, the paper estimates demand elasticity and finds it is inelastic, so the controller mainly acts as a transfer. It also estimates the deadweight loss from the reserve price introduced in December 2025 at most 2.6 ETH when capacity was slack.
WhyUses a scheduled upgrade as a quasi-experiment to estimate blob price elasticity and welfare, with placebo-based inference, relevant to protocol design.
In briefCryptoFlowBench: A Benchmark for Multi-Hop Crypto Fund Tracing on Blockchains
SummaryThis paper introduces CryptoFlowBench, a benchmark built from authoritative reports on cryptocurrency-related illicit incidents, with four tasks ranging from local to multi-hop fund tracing. It also proposes ChainSleuth, an agent that uses graph tools to iteratively reason and update its tracing path. Experiments show no baseline consistently leads on all four tasks, and ChainSleuth achieves notable gains on multi-hop tasks.
In briefGuidanceNFT: Reward fine-tuning for guided diffusion models
SummaryThe paper notes that DiffusionNFT, which fine-tunes diffusion models toward task rewards, may not keep full prompt fidelity, and that classifier-free guidance (CFG) can reduce the optimized reward. The authors attribute standard NFT's failure on guided samples to a mismatch between its reference velocity and the score of the guided distribution. They propose GuidanceNFT, which estimates a matching reference velocity via online flow matching and uses it to correct the loss. Experiments on SD3.5-M and models with built-in guidance report higher optimized rewards and improved prompt alignment than NFT.
