arXiv subjects

Xinkai Zhang

Publications and source records attributed to Xinkai Zhang.

4 recordsLinked to original sources

arXiv2026

EFX Allocations for Three Agents and Seven or Eight Chores

We prove that every nonnegative additive chore instance with three agents and either seven or eight indivisible chores admits a chores-EFX allocation, in the zero-tolerant sense that every owned chore, including one of zero cost, is quantified in the trim. Both proofs are computer-assisted, but their machine formulas differ. For seven chores, hand-checkable canonicalization reduces nonexistence to a quantifier-free linear real arithmetic (QF_LRA) formula over 21 variables with one failure clause per complete allocation. For eight chores, instances in which two agents share a weakly cheapest chore are lifted from the seven-chore theorem through the matching insertion lemma of Kobayashi, Mahara, and Sakamoto, and the remaining pairwise-disjoint-argmin class reduces to a residual QF_LRA formula over 24 variables. Z3 5.1.0 and cvc5 1.3.4 report both formulas unsatisfiable. With the known $m\leq 2n$ theorem, this settles every three-agent instance with at most eight chores; $m=9$ is the next open cardinality, and additive chores can fail to admit EFX for every $n\geq 4$.

cs.GT↗

arXiv2026

MT-JailBench: A Modular Benchmark for Understanding Multi-Turn Jailbreak Attacks

Multi-turn jailbreaks exploit the ability of large language models to accumulate and act on conversational context. Instead of stating a harmful request directly, an attacker can gradually steer the conversation toward an unsafe answer. Recent methods demonstrate this risk, but they are usually evaluated as black-box pipelines with different budgets, judges, retry rules, and strategy generation procedures. As a result, it is often unclear whether reported gains reflect stronger attack mechanisms or different experimental conditions. We introduce MT-JailBench, a modular evaluation framework for benchmarking multi-turn jailbreaks under fixed conditions. MT-JailBench implements each attack as five interacting modules: evaluation function, attack strategy, prompt generation, prompt refinement, and flow control. This design enables fair comparison across attack methods and component-wise analysis of what drives attack success. Using MT-JailBench, we find that resource budgets and evaluation functions are major confounders: controlling turns, retries, interactions, sampled strategies, and judges substantially change the ranking of attacks. At the component level, prompt generation accounts for most performance variation, while refinement and flow control provide moderate gains. We also find that explicit dynamic strategy generation is not always necessary; stochastic sampling from a fixed strategy can rival more elaborate diversification mechanisms. Finally, recomposing the best components yields a strong attack configuration that outperforms its source attacks and generalizes across diverse target LLMs. MT-JailBench therefore provides a modular framework for comparing multi-turn jailbreaks, understanding the impact of components, and guiding stronger red-teaming evaluations.

cs.CR↗

arXiv2026

TEC: A Collection of Human Trial-and-error Trajectories for Problem Solving

Trial-and-error is a fundamental strategy for humans to solve complex problems and a necessary capability for Artificial Intelligence (AI) systems operating in real-world environments. Although several trial-and-error AI techniques have recently been proposed, most of them rely on simple heuristics designed by researchers and achieve limited performance gains. The core issue is the absence of appropriate data: current models cannot learn from detailed records of how humans actually conduct trial-and-error in practice. To address this gap, we introduce a data annotation platform and a corresponding dataset, termed Trial-and-Error Collection (TEC). The platform records users' complete trajectories across multiple trials and collects their reflections after receiving error feedback. Using this platform, we record the problem-solving processes of 46 participants on 58 tasks, resulting in 5,370 trial trajectories along with error reflections across 41,229 webpages. With this dataset, we observe that humans achieve substantially higher accuracy compared to LLMs, which demonstrates that humans are more effective in trial-and-error than LLMs. We believe that the TEC platform and dataset provide a valuable foundation for understanding human trial-and-error behavior and for developing more capable AI systems. Platform and dataset are publicly available.

cs.CL↗

arXiv2020

Trua: Efficient Task Replication for Flexible User-defined Availability in Scientific Grids

Failure is inevitable in scientific computing. As scientific applications and facilities increase their scales over the last decades, finding the root cause of a failure can be very complex or at times nearly impossible. Different scientific computing customers have varying availability demands as well as a diverse willingness to pay for availability. In contrast to existing solutions that try to provide higher and higher availability in scientific grids, we propose a model called Task Replication for User-defined Availability (Trua). Trua provides flexible, user-defined, availability in scientific grids, allowing customers to express their desire for availability to computational providers. Trua differs from existing task replication approaches in two folds. First, it relies on the historic failure information collected from the virtual layer of the scientific grids. The reliability model for the failures can be represented with a bimodal Johnson distribution which is different from any existing distributions. Second, it adopts an anomaly detector to filter out anomalous failures; it additionally adopts novel selection algorithms to mitigate the effects of temporary and spatial correlations of the failures without knowing the root cause of the failures. We apply the Trua on real-world traces collected from the Open Science Grid (OSG). Our results show that the Trua can successfully meet user-defined availability demands.

cs.DC↗