AI Summary

• Behavioral backdoor detection in AI models is a critical problem that cannot be solved model by model due to the 43.4% generalization gap. [3]
• LLM: Large Language Model AI supply chain: The network of organizations involved in the development and deployment of AI models Backdoor attack: A type of attack where an adversary injects a malicious function into a model, allowing them to manipulate its behavior Data poisoning: The process of manipulating training data to cause a model to produce incorrect or biased results Sleeper agent: A type of backdoor attack that involves training a model to behave normally during testing but to produce different outputs when deployed in the wild Neural cleanse: A technique for identifying and mitigating backdoor attacks in neural networks Activation clustering: A method for detecting backdoor attacks by analyzing the activation patterns of a model's neurons [3]
• Model-aware detection offers a practical path forward, achieving 90.6% universal accuracy across heterogeneous LLM ecosystems. [2]
• The cross-LLM backdoor detection problem is now well-characterized, and our findings provide a foundation for building robust defenses against this critical threat. [1]

Abstract
As AI agents become integral to enterprise workflows, their reliance on shared tool libraries and pre-trained components creates significant supply chain vulnerabilities. While previous work has demonstrated behavioral backdoor detection within individual LLM architectures, the critical question of cross-LLM generalization remains unexplored, a gap with serious implications for organizations deploying multiple AI systems. We present the first systematic study of cross-LLM behavioral backdoor detection, evaluating generalization across six production LLMs (GPT-5.1, Claude Sonnet 4.5, Grok 4.1, Llama 4 Maverick, GPT-OSS 120B, and DeepSeek Chat V3.1). Through 1,198 execution traces and 36 cross-model experiments, we quantify a critical finding: single-model detectors achieve 92.7% accuracy within their training distribution but only 49.2% across different LLMs, a 43.4 percentage point generalization gap equivalent to random guessing. Our analysis reveals that this gap stems from model-specific behavioral signatures, particularly in temporal features (coefficient of variation > 0.8), while structural features remain stable across architectures. We show that model-aware detection incorporating model identity as an additional feature achieves 90.6% accuracy universally across all evaluated models. We release our multi-LLM trace dataset and detection framework to enable reproducible research.

Why we think this paper is great for you:
This paper is highly relevant as it addresses critical security and reliability concerns within AI agent supply chains, which is essential for building trustworthy AI products and defining robust product strategies. It helps you understand the foundational risks and safeguards needed when integrating AI into enterprise workflows.

AI Summary

• The study used a combination of EEG and behavioral data to investigate the effects of cognitive workload on decision-making and performance. [3]
• The results showed that participants performed better under low workload conditions compared to high workload conditions. [3]
• The study found that cognitive workload had a significant impact on decision-making and performance, with participants making more errors under high workload conditions. [3]
• The study used a virtual reality (VR) environment to investigate the effects of cognitive workload on decision-making and performance. [2]
• EEG: Electroencephalography - a technique used to measure electrical activity in the brain. [1]

Abstract
Human-AI teams can be vulnerable to catastrophic failure when feedback from the AI is incorrect, especially under high cognitive workload. Traditional team aggregation methods, such as voting, are susceptible to these AI errors, which can actively bias the behaviour of each individual and inflate the likelihood of an erroneous group decision. We hypothesised that a collaborative Brain-Computer Interface (cBCI) using neural activity collected before a behavioural decision is made can provide a source of information that is decoupled from this biased behaviour, thereby protecting the team from the deleterious influence of AI error. We tested this in a VR drone surveillance task where teams of operators faced high workload and systematically misleading AI cues, comparing traditional behaviour-based team strategies against a purely Neuro-Decoupled Team (NDT) that used only BCI confidence scores derived from pre-response EEG. Under AI deception, behaviour-based teams catastrophically failed, with Majority Vote accuracy collapsing to 44%. The NDT, however, maintained 98% accuracy, a statistically significant synergistic gain over even the team's best individual performer (p < .001). This was explained by a neuro-behavioural decoupling, where the BCI's predictions remained highly accurate while the operator's subjective confidence became an unreliable signal. We conclude that an implicit BCI provides resilience by learning to adapt its neural strategy, shifting from relying on signals of efficient, autopilot processing in simple conditions to interpreting signatures of effortful deliberation when confronted with cognitive conflict. This demonstrates a system that leverages the context of the neural signal to defend against AI-induced error in high-stakes environments.

Why we think this paper is great for you:
This research directly explores the challenges and solutions for effective human-AI collaboration, particularly in scenarios involving AI errors. Understanding how to build resilient human-AI teams is crucial for setting a clear vision for AI-driven products and ensuring their reliability.

AI Summary

• The Lagrangian Dual response approximates the best response with greater recall than the Gradient Descent response. [3]
• Performativ e prediction: A type of prediction where the model's predictions affect the data distribution, leading to biased results. [3]
• Input convex neural networks (ICNNs): Neural networks that are constrained to ensure they implement functions that are convex with respect to the input features. [3]
• The method may not be effective for all types of classifiers or data distributions. [3]
• The post-response checks may not be sufficient to identify all points that can respond the classifier. [3]
• Strategic behavior is a significant concern in machine learning, as it can lead to biased data distributions and undermine the accuracy and fairness of classifiers. [2]

Abstract
We consider the problem of strategic classification, where the act of deploying a classifier leads to strategic behaviour that induces a distribution shift on subsequent observations. Current approaches to learning classifiers in strategic settings are focused primarily on the linear setting, but in many cases non-linear classifiers are more suitable. A central limitation to progress for non-linear classifiers arises from the inability to compute best responses in these settings. We present a novel method for computing the best response by optimising the Lagrangian dual of the Agents' objective. We demonstrate that our method reproduces best responses in linear settings, identifying key weaknesses in existing approaches. We present further results demonstrating our method can be straight-forwardly applied to non-linear classifier settings, where it is useful for both evaluation and training.

Why we think this paper is great for you:
This paper delves into the strategic implications of deploying AI classifiers, focusing on how user behavior shifts in response to AI systems. This insight is vital for developing effective product strategies and roadmaps that anticipate market reactions and adapt accordingly.

AI Summary

• The authors highlight the potential benefits of using AI in research, including increased productivity and well-being for mathematicians. [3]
• They also note that AI can excel at routine but lengthy calculations, freeing up time for more creative work. [3]
• They also note that human-AI collaboration can lead to new insights and solutions. [3]
• LLM: Large Language Model AI: Artificial Intelligence The use of AI in research has the potential to increase productivity and well-being for mathematicians. [3]
• Human-AI collaboration can lead to new insights and solutions. [3]
• The paper discusses the use of a large language model (LLM) in solving a research problem in mathematical statistics. [2]

Abstract
Over the last few months, AI models including large language models have improved greatly. There are now several documented examples where they have helped professional mathematical scientists prove new results, sometimes even helping resolve known open problems. In this short note, we add another example to the list, by documenting how we were able to solve a previously unsolved research problem in robust mathematical statistics with crucial help from GPT-5. Our problem concerns robust density estimation, where the observations are perturbed by Wasserstein-bounded contaminations.In a previous preprint (Chao and Dobriban, 2023, arxiv:2308.01853v2), we have obtained upper and lower bounds on the minimax optimal estimation error; which were, however, not sharp. Starting in October 2025, making significant use of GPT-5 Pro, we were able to derive the minimax optimal error rate (reported in version 3 of the above arxiv preprint). GPT-5 provided crucial help along the way, including by suggesting calculations that we did not think of, and techniques that were not familiar to us, such as the dynamic Benamou-Brenier formulation, for key steps in the analysis. Working with GPT-5 took a few weeks of effort, and we estimate that it could have taken several months to get the same results otherwise. At the same time, there are still areas where working with GPT-5 was challenging: it sometimes provided incorrect references, and glossed over details that sometimes took days of work to fill in. We outline our workflow and steps taken to mitigate issues. Overall, our work can serve as additional documentation for a new age of human-AI collaborative work in mathematical science.

Why we think this paper is great for you:
This paper demonstrates the advanced problem-solving capabilities of AI, showcasing how AI can assist in complex intellectual tasks. This understanding can inform your vision for leveraging AI in product development and setting ambitious goals for tech teams.

AI Summary

• The CAL model uses a ProtoPNet architecture, which is a type of neural network that can handle multiple modalities of input data. [3]
• The model is trained using a Bayesian optimization scheme to optimize the hyperparameters for the best performance. [3]
• The paper also proposes an extension of CAL called Multi-Conformal Prediction (M-CP) to estimate confidence regions around predicted genetic logit vectors. [3]
• ProtoPNet: A type of neural network architecture that can handle multiple modalities of input data. [3]
• Multi-Conformal Prediction (M-CP): An extension of CAL that estimates confidence regions around predicted genetic logit vectors. [3]
• The CAL model uses a ProtoPNet architecture and Bayesian optimization to optimize hyperparameters, while the M-CP model extends this approach to estimate confidence regions around predicted genetic logit vectors. [3]
• The paper proposes a new approach for handling missing data in deep learning models called Conformal Abstention Learning (CAL) and its application to image classification tasks. [2]

Abstract
Species detection is important for monitoring the health of ecosystems and identifying invasive species, serving a crucial role in guiding conservation efforts. Multimodal neural networks have seen increasing use for identifying species to help automate this task, but they have two major drawbacks. First, their black-box nature prevents the interpretability of their decision making process. Second, collecting genetic data is often expensive and requires invasive procedures, often necessitating researchers to capture or kill the target specimen. We address both of these problems by extending prototype networks (ProtoPNets), which are a popular and interpretable alternative to traditional neural networks, to the multimodal, cost-aware setting. We ensemble prototypes from each modality, using an associated weight to determine how much a given prediction relies on each modality. We further introduce methods to identify cases for which we do not need the expensive genetic information to make confident predictions. We demonstrate that our approach can intelligently allocate expensive genetic data for fine-grained distinctions while using abundant image data for clearer visual classifications and achieving comparable accuracy to models that consistently use both modalities.

Why we think this paper is great for you:
This paper explores optimizing AI models by selectively using modalities, which is relevant for practical AI product development. It can help you consider resource efficiency and cost-effectiveness when designing and implementing AI features in your products.

AI Summary

• The coarse stage estimates the complexity of an image using handcrafted descriptors and assigns a granularity label based on the estimated complexity. [3]
• The results show that Dynamic Granularity Matters outperforms state-of-the-art models on several benchmarks, including ImageNet-1K, CIFAR-100, and Tiny-ImageNet-200. [3]
• Coarse Stage: The first stage of the Dynamic Granularity Matters approach, which estimates the complexity of an image using handcrafted descriptors and assigns a granularity label based on the estimated complexity. [3]
• The Dynamic Granularity Matters approach outperforms state-of-the-art models on several benchmarks, including ImageNet-1K, CIFAR-100, and Tiny-ImageNet-200. [3]
• The paper presents a novel approach to vision transformers (ViTs) called Dynamic Granularity Matters, which rethinks ViT beyond fixed patch splitting. [2]
• The method involves two stages: coarse and fine. [1]

Abstract
Vision Transformers (ViTs) have demonstrated strong capabilities in capturing global dependencies but often struggle to efficiently represent fine-grained local details. Existing multi-scale approaches alleviate this issue by integrating hierarchical or hybrid features; however, they rely on fixed patch sizes and introduce redundant computation. To address these limitations, we propose Granularity-driven Vision Transformer (Grc-ViT), a dynamic coarse-to-fine framework that adaptively adjusts visual granularity based on image complexity. It comprises two key stages: (1) Coarse Granularity Evaluation module, which assesses visual complexity using edge density, entropy, and frequency-domain cues to estimate suitable patch and window sizes; (2) Fine-grained Refinement module, which refines attention computation according to the selected granularity, enabling efficient and precise feature learning. Two learnable parameters, α and \b{eta}, are optimized end-to-end to balance global reasoning and local perception. Comprehensive evaluations demonstrate that Grc-ViT enhances fine-grained discrimination while achieving a superior trade-off between accuracy and computational efficiency.

Why we think this paper is great for you:
While technical, this paper discusses advancements in Vision Transformers, a core AI technology. Understanding such underlying innovations can inform your long-term vision for AI-powered products and the capabilities tech teams can leverage.

AI Summary

• The text discusses leader-follower problems in the context of optimization and decision-making under uncertainty. [2]
• Follower: The player who responds to the leader's decisions. [1]

Abstract
Energy systems are changing rapidly. More and more, energy production is becoming decentralized, highly variable and intermittent (solar, wind), while demand is diversifying (electric vehicles). As a result, balancing supply and demand is becoming more complex, making the adjustment of demand an interesting tool. Demand response is a typical leader-follower problem: a consumer (follower) adjusts his energy consumption based on the prices (or any other incentive) set by the supplier (leader). We propose a versatile and modular framework to address any leader-follower problem, focusing on the handling of often overlooked informational issues. First, we introduce a model that defines the rules of the game (W-model): agents are decision-makers, and Nature encapsulates everything beyond their control, such as private knowledge and exogenous factors. Following the so-called Witsenhausen intrinsic model, we present an efficient way to represent - on a product set, equipped with a product $σ$-algebra - the information available to agents when making decisions. Next, we introduce Games in Product Form (W-games) by equipping each player (a group of agents) with preferences (objective function and belief) over different outcomes. Thereby, we incorporate an additional layer of information, the characteristics of the preferences linked to players, which affects the possible definitions of an equilibrium. We make this explicit in Nash and Stackelberg equilibria. Equipped with this framework, we reformulate several papers on demand response, highlighting overlooked informational issues. We also provide an application based on the Thailand demand response program.

Why we think this paper is great for you:
This paper explores modeling complex systems like energy demand response, which involves understanding dynamic interactions. While not directly about product management, the analytical approach to complex systems could offer insights into strategic planning and roadmap development.