Searcharxiv⌕ Search

arXiv subjects

Aideen Fay

Publications and source records attributed to Aideen Fay.

8 recordsLinked to original sources

Quantum Cohomology and irrationality of Küchle fourfolds

We prove that very general Küchle fourfolds of types (c1), (c6), (b8), and (b11) are irrational. The key ingredients are a new rationality criterion, based on the theory of Hodge atoms due to Katzarkov--Kontsevich--Pantev--Yu and its integral extension by Guéré, and a computation of the action of quantum multiplication on the vanishing cohomology of these fourfolds, which uses a version of the Quantum Lefschetz Hyperplane Principle. Quantum multiplication acts as a scalar on the vanishing cohomology, and we give an enumerative interpretation of this scalar.

math.AG↗

LLMail-Inject: A Dataset from a Realistic Adaptive Prompt Injection Challenge

Indirect prompt injection attacks exploit a fundamental weakness of LLMs: executing unauthorized tasks from untrusted sources. Despite the practical significance of this attack vector, existing benchmarks rely on small sets of template-based attacks (100-1,000 examples) that fail to capture the diversity and adaptiveness of real-world adversaries. Consequently, benchmarks are saturated and overfit to static attacks. We introduce LLMail-Inject, a large-scale benchmark of 208,095 distinct human-written attack prompts collected through a public red-teaming challenge with 839 participants. LLMail-Inject evaluates the attack's complete chain of execution: prompts must be retrieved from an email database, evade production-level defenses, and trigger unauthorized tool calls with correct arguments. Evaluation reveals that models that achieve 5% attack success rate (ASR) on existing benchmarks fail more on LLMail-Inject: Meta SecAlign--specifically trained against prompt injection--still yields 32% ASR, GPT-4o exhibits 58% ASR, and GPT-5 mini 47% ASR. We further demonstrate that attacks transfer across domains, achieving 37% ASR on proof-of-concept coding tasks and 28-45% ASR on file management tasks. We release our challenge platform and curated evaluation subsets at multiple difficulty tiers and the full dataset with rich annotations.

cs.CR↗

Hasse Obstructions to Rationality for Special Fourfolds

For every nonempty Hassett divisor $\mathcal C_d$ parameterizing special cubic fourfolds, we consider the quaternion class $β_d=(d/2,-3)$ in the two-torsion Brauer group. We prove that if a Hodge-general member of ${C}_d$ is rational then $β_d=0$ -- or, equivalently, that Huybrechts' twisted-K3 condition $(**')$ holds. Consequently, a very general member of $\mathcal C_d$ is irrational if $β_d\ne0$. Thus, a very general cubic fourfold containing a smooth cubic scroll or a Veronese surface is irrational, and hence so is a very general Küchle fourfold of type $(\mathrm{c7})$. We also obtain an analogous obstruction for Hodge--special Gushel--Mukai fourfolds: a very general member in the discriminant--$d$ locus can be rational only if $d$ is a sum of two squares. Consequently, a very general Gushel--Mukai fourfold containing a cubic scroll is irrational.

math.AG↗

An atomic criterion for irrationality without quantum computations

The birational invariants introduced by Katzarkov-Kontsevich-Pantev-Yu allows one to obtain irrationality results for varieties whose quantum cohomology is well-behaved. We observe that under certain cohomological conditions, we can deduce irrationality of a very general member from the theory of atoms without actually computing them, using only monodromy equivariance of quantum multiplication and irreducibility of the monodromy representation. Our criterion applies to the very general cubic and Gushel-Mukai fourfolds, whose irrationalities were already known, but also to the very general K{ü}chle fourfold of type (c5), which is a Fano manifold of index one.

math.AG↗

The Very General Verra Fourfold is Irrational

We show that the very general Verra fourfold is irrational, using the Hodge atom framework of Katzarkov--Kontsevich--Pantev--Yu. Two novel points are: a refined analysis of Hodge atoms, based on the involution on the cohomology of $X$, and a derivation of the quantum multiplication matrix from the quantum differential operator. This gives the first successful application of the method of Hodge atoms to a space with Picard rank greater than one.

math.AG↗

Equivariant irrationality of very general symmetric Verra fourfolds

We prove that a very general complex symmetric Verra fourfold is not \(\ZZ/2\)-birational to \(\PP^4\), using the \emph{theory of atoms} introduced by Katzarkov--Kontsevich--Pantev--Yu~\citep{Katzarkov2025BirationalIF} in the equivariant setting as in Cavenaghi--Katzarkov--Kontsevich~\citep{cavenaghi2026atomsmeetsymbols}.

math.AG↗

The Shape of Adversarial Influence: Characterizing LLM Latent Spaces with Persistent Homology

Existing interpretability methods for Large Language Models (LLMs) predominantly capture linear directions or isolated features. This overlooks the high-dimensional, relational, and nonlinear geometry of model representations. We apply persistent homology (PH) to characterize how adversarial inputs reshape the geometry and topology of internal representation spaces of LLMs. This phenomenon, especially when considered across operationally different attack modes, remains poorly understood. We analyze six models (3.8B to 70B parameters) under two distinct attacks, indirect prompt injection and backdoor fine--tuning, and show that a consistent topological signature persists throughout. Adversarial inputs induce topological compression, where the latent space becomes structurally simpler, collapsing the latent space from varied, compact, small-scale features into fewer, dominant, large-scale ones. This signature is architecture-agnostic, emerges early in the network, and is highly discriminative across layers. By quantifying the shape of activation point clouds and neuron-level information flow, our framework reveals geometric invariants of representational change that complement existing linear interpretability methods.

cs.LG↗

Get my drift? Catching LLM Task Drift with Activation Deltas

LLMs are commonly used in retrieval-augmented applications to execute user instructions based on data from external sources. For example, modern search engines use LLMs to answer queries based on relevant search results; email plugins summarize emails by processing their content through an LLM. However, the potentially untrusted provenance of these data sources can lead to prompt injection attacks, where the LLM is manipulated by natural language instructions embedded in the external data, causing it to deviate from the user's original instruction(s). We define this deviation as task drift. Task drift is a significant concern as it allows attackers to exfiltrate data or influence the LLM's output for other users. We study LLM activations as a solution to detect task drift, showing that activation deltas - the difference in activations before and after processing external data - are strongly correlated with this phenomenon. Through two probing methods, we demonstrate that a simple linear classifier can detect drift with near-perfect ROC AUC on an out-of-distribution test set. We evaluate these methods by making minimal assumptions about how users' tasks, system prompts, and attacks can be phrased. We observe that this approach generalizes surprisingly well to unseen task domains, such as prompt injections, jailbreaks, and malicious instructions, without being trained on any of these attacks. Interestingly, the fact that this solution does not require any modifications to the LLM (e.g., fine-tuning), as well as its compatibility with existing meta-prompting solutions, makes it cost-efficient and easy to deploy. To encourage further research on activation-based task inspection, decoding, and interpretability, we release our large-scale TaskTracker toolkit, featuring a dataset of over 500K instances, representations from six SoTA language models, and a suite of inspection tools.

cs.CR↗