Searcharxiv⌕ Search

arXiv subjects

Eric Ren

Publications and source records attributed to Eric Ren.

4 recordsLinked to original sources

FraudBench: A Legal Evaluation of AI Deception

We introduce FraudBench, an evaluation of large language models' (LLMs) propensities to recommend, assist, or directly take actions that would constitute fraud under U.S. law. Our evaluation contains a novel dataset of 159 scenarios based on real-world U.S. court cases and validated by legal experts. We present each scenario to LLMs via three tasks at varying levels of autonomy: advising users, drafting messages for users to send, and acting directly on behalf of users (477 tasks total). In each scenario, FraudBench re-enacts the facts of a court case to give LLMs an opportunity, but no explicit instruction, to either misrepresent or conceal a material fact in a high-stakes commercial setting--e.g., whether to describe company-funded product testing as "independent" to customers (misrepresentation) or whether to hide unfavorable regulatory investigations from clients (concealment). We evaluate 20 LLMs across 25 experimental conditions (totaling >300,000 generations) and include a preregistered human baseline ($n_{\texttt{annotations}} = 2{,}957$, $n_{\texttt{humans}} = 996$). We score a response as fraudulent when it engages in, assists with, or recommends misrepresentation or concealment. Key findings include: (1) across LLMs, fraud rates ranged from near 0% to over 40%; (2) LLM fraud rates increased with autonomy, with overall rates of 3.0% when advising, 14.7% when drafting, and 21.4% when acting directly, without applying pressure; (3) LLMs sometimes engaged in fraud despite recognizing legal risks; (4) LLMs more often engaged in fraud by concealment than by misrepresentation; and (5) prompt-based interventions substantially affected fraud rates, with a prompt to obey the law reducing fraud rates, while prompts to maximize profits and act agentically increased fraud rates.

cs.AI↗

On numerical semigroup elements and the $\ell_0$- and $\ell_\infty$-norms of their factorizations

A numerical semigroup $S$ is a cofinite, additively-closed subset of $\mathbb Z_{\ge 0}$ that contains 0, and a factorization of $x \in S$ is a $k$-tuple $z = (z_1, \ldots, z_k)$ where $x = z_1a_1 + \cdots + z_ka_k$ expresses $x$ as a sum of generators of $S = \langle a_1, \ldots, a_k \rangle$. Much~of the study of non-unique factorization centers on factorization length $z_1 + \cdots + z_k$, which coincies with the $\ell_1$-norm of $z$ as the $k$-tuple. In this paper, we study the $\ell_\infty$-norm and $\ell_0$-norm of factorizations, viewed as alternative notions of length, with particular focus on the generalizations $Δ_\infty(x)$ and $Δ_0(x)$ of the delta set $Δ(x)$ from classical factorization length. We prove that the $\infty$-delta set $Δ_\infty(x)$ is eventually periodic as a function of $x \in S$, classify $Δ_\infty(S)$ and the 0-delta set $Δ_0(S)$ for several well-studied families of numerical semigroups, and identify families of numerical semigroups demonstrating $Δ_\infty(S)$ and $Δ_0(S)$ can be arbitrarily long intervals and can avoid arbitrarily long subintervals.

math.AC↗

Perspicacious $l_p$ norm parameters

Fix $t\in [1,\infty]$. Let $S$ be an atomic commutative semigroup and, for all $x\in S$, let $\mathscr{L}_t(S):=\{\|f\|_t:f\in Z(x)\}$ be the "$t$-length set" of $x$ (using the standard $l_p$-space definition of $\|\cdot\|_t$). The $t$-Delta set of $x$ (denoted $Δ_t(S)$) is the set of gaps between consecutive elements of $\mathscr{L}_t(S)$; the Delta set of $S$ is then defined by $\bigcup\limits_{x\in S} Δ_t(S)$. Though all existing literature on this topic considers the $1$-Delta set, recent results on the $t$-elasticity of Numerical Semigroups (Behera et. al.) for $t\neq 1$ have brought attention to other invariants, such as the $t$-Delta set for $t\neq 1$, as well. Here we characterize $Δ_t(S)$ for all numerical semigroups $\langle a_1,a_2\rangle$ and all $t\in(1,\infty)$ outside a small family of extremal examples. We also determine the cardinality and describe the distribution of that aberrant family.

math.AC↗

Improved upper bounds on longest-path and maximal subdivision transversals

Let $G$ be a connected graph on $n$ vertices. The Gallai number $Gal(G)$ of $G$ is the size of the smallest set of vertices that meets every maximum path in $G$. Grünbaum constructed a graph $G$ with $Gal(G)=3$. Very recently, Long, Milans, and Munaro, proved that $Gal(G)\leq 8n^{{3}/{4}}$. This was the first sublinear upper bound on $Gal(G)$ in terms of $n$. We improve their bound to $Gal(G)\leq 5 n^{{2}/{3}}$. We also tighten a more general result of Long et al. For a multigraph $M$ on m edges, we prove that if the set $L(M,G)$ of maximum $M$-subdivisions in $G$ is pairwise intersecting and $n\geq m^{6}$, then $G$ has a set of vertices with size at most $5 n^{{2}/{3}}$ that meets every $Q\in \mathcal{L}(M,G)$

math.CO↗