Selected work

Representative Papers

Why Formal Monitors Fail: Attack Distribution Entropy as a Coverage Bound for LTL-Based LLM Agent Safety

Aug 02, 2026

This study addresses the significant variation in attack interception efficacy of existing LTL/FSA-based safety monitors across different large language model (LLM) architectures, a phenomenon previously lacking theoretical explanation. The work reveals, for the first time, that the distributional entropy of attack trigger-completion patterns is the fundamental determinant of formal monitor coverage. To enable pre-deployment evaluation independent of model architecture, the authors propose an entropy-based testing methodology. Leveraging Shannon entropy and statistical correlation analysis, they establish a theoretical upper bound on the recall of invariant-based monitors. Experiments across eight state-of-the-art LLMs demonstrate a strong negative correlation (r = −0.87) between monitor coverage and attack distribution entropy: GPT-style models achieve 96% coverage with a single pattern, whereas Gemini-style models require multiple clustered strategies yet attain only 6–13% coverage.

0 citationsRead paper

Incentives and Market Structure in Intent-Based Exchanges: Evidence from a Solver-Reward Reform

Jul 24, 2026

This study investigates how solver incentive mechanisms in intent-based decentralized exchanges shape market structure and the distribution of value capture. Leveraging the CoW Protocol’s CIP-74 reform—which replaced a fixed reward cap with a protocol-revenue-linked scheme and introduced ad valorem trading fees—as a natural experiment, the authors employ difference-in-differences and triple-difference designs, alongside Herfindahl–Hirschman Index (HHI) and Spearman rank correlation analyses. Their findings reveal, for the first time empirically, that reward rule changes systematically redistribute value across order sizes: post-reform, small-order markets became less concentrated, while large-order concentration rose significantly, with volume-weighted HHI increasing from 0.176 to 0.241. Notably, average execution quality remained stable at approximately seven basis points, underscoring the pivotal role of incentive design in shaping market structure.

0 citationsRead paper

Cosine-Gated Adam-Decay: Drop-In Staleness-Aware Outer Optimization for Decoupled DiLoCo

May 09, 2026

This work addresses momentum instability in asynchronous DiLoCo systems, where the outer-loop optimizer receives pseudo-gradients delayed by multiple rounds. Standard Nesterov momentum fails under large delays because it ignores gradient age. To resolve this, the authors propose CGAD, a plug-and-play, age-aware optimizer that introduces delay awareness into outer-loop optimization for the first time. CGAD weights delayed gradients via exponential decay and cosine gating, then incorporates these into Adam’s first- and second-moment estimates. Theoretically, its bias depends only on the decay coefficient, not the maximum delay. The method provides default hyperparameters transferable across model scales and a variant, PA-CGAD, suited for partially synchronous schedules. Experiments on Llama-style models (25M–7B parameters) demonstrate that CGAD enables stable training across diverse delays, significantly outperforming Adam Decay and Nesterov baselines, with failure probability far below random chance.

0 citationsRead paper

Bittensor Protocol: The Bitcoin in Decentralized Artificial Intelligence? A Critical and Empirical Analysis

Jun 29, 2025

This paper investigates whether Bittensor can serve as the “Bitcoin of decentralized AI,” focusing on the decentralization degree of its tokenomics, consensus mechanism, and incentive architecture. Analyzing on-chain data from 64 active subnets via empirical analysis, statistical modeling, and cybersecurity simulation, we identify critical issues: high concentration of stake and rewards, and misalignment between incentives and contribution quality. To address these, we propose a novel two-track protocol optimization: (1) performance-weighted token issuance and a composite scoring mechanism incorporating a trust-based reward multiplier to align incentives with service quality; and (2) a stake cap at the 88th percentile to significantly enhance resilience against 51% attacks. Experimental evaluation across daily, weekly, and monthly time horizons demonstrates robust efficacy—improving both network security and the correlation between rewards and actual contribution.

0 citationsRead paper

Recent publications

Latest Papers

Why Formal Monitors Fail: Attack Distribution Entropy as a Coverage Bound for LTL-Based LLM Agent Safety

Aug 02, 2026

This study addresses the significant variation in attack interception efficacy of existing LTL/FSA-based safety monitors across different large language model (LLM) architectures, a phenomenon previously lacking theoretical explanation. The work reveals, for the first time, that the distributional entropy of attack trigger-completion patterns is the fundamental determinant of formal monitor coverage. To enable pre-deployment evaluation independent of model architecture, the authors propose an entropy-based testing methodology. Leveraging Shannon entropy and statistical correlation analysis, they establish a theoretical upper bound on the recall of invariant-based monitors. Experiments across eight state-of-the-art LLMs demonstrate a strong negative correlation (r = −0.87) between monitor coverage and attack distribution entropy: GPT-style models achieve 96% coverage with a single pattern, whereas Gemini-style models require multiple clustered strategies yet attain only 6–13% coverage.

0 citationsRead paper

Incentives and Market Structure in Intent-Based Exchanges: Evidence from a Solver-Reward Reform

Jul 24, 2026

This study investigates how solver incentive mechanisms in intent-based decentralized exchanges shape market structure and the distribution of value capture. Leveraging the CoW Protocol’s CIP-74 reform—which replaced a fixed reward cap with a protocol-revenue-linked scheme and introduced ad valorem trading fees—as a natural experiment, the authors employ difference-in-differences and triple-difference designs, alongside Herfindahl–Hirschman Index (HHI) and Spearman rank correlation analyses. Their findings reveal, for the first time empirically, that reward rule changes systematically redistribute value across order sizes: post-reform, small-order markets became less concentrated, while large-order concentration rose significantly, with volume-weighted HHI increasing from 0.176 to 0.241. Notably, average execution quality remained stable at approximately seven basis points, underscoring the pivotal role of incentive design in shaping market structure.

0 citationsRead paper

Cosine-Gated Adam-Decay: Drop-In Staleness-Aware Outer Optimization for Decoupled DiLoCo

May 09, 2026

This work addresses momentum instability in asynchronous DiLoCo systems, where the outer-loop optimizer receives pseudo-gradients delayed by multiple rounds. Standard Nesterov momentum fails under large delays because it ignores gradient age. To resolve this, the authors propose CGAD, a plug-and-play, age-aware optimizer that introduces delay awareness into outer-loop optimization for the first time. CGAD weights delayed gradients via exponential decay and cosine gating, then incorporates these into Adam’s first- and second-moment estimates. Theoretically, its bias depends only on the decay coefficient, not the maximum delay. The method provides default hyperparameters transferable across model scales and a variant, PA-CGAD, suited for partially synchronous schedules. Experiments on Llama-style models (25M–7B parameters) demonstrate that CGAD enables stable training across diverse delays, significantly outperforming Adam Decay and Nesterov baselines, with failure probability far below random chance.

0 citationsRead paper

Bittensor Protocol: The Bitcoin in Decentralized Artificial Intelligence? A Critical and Empirical Analysis

Jun 29, 2025

This paper investigates whether Bittensor can serve as the “Bitcoin of decentralized AI,” focusing on the decentralization degree of its tokenomics, consensus mechanism, and incentive architecture. Analyzing on-chain data from 64 active subnets via empirical analysis, statistical modeling, and cybersecurity simulation, we identify critical issues: high concentration of stake and rewards, and misalignment between incentives and contribution quality. To address these, we propose a novel two-track protocol optimization: (1) performance-weighted token issuance and a composite scoring mechanism incorporating a trust-based reward multiplier to align incentives with service quality; and (2) a stake cap at the 88th percentile to significantly enhance resilience against 51% attacks. Experimental evaluation across daily, weekly, and monthly time horizons demonstrates robust efficacy—improving both network security and the correlation between rewards and actual contribution.

0 citationsRead paper