Evgenii Kortukov

Samar Heydari, Jawher Said, Galip Ümit Yolcu et al. · mit

Deep learning models for flood and wildfire segmentation and object detection enable precise, real-time disaster localization when deployed on embedded drone platforms. However, in natural disaster management, the lack of transparency in their decision-making process hinders human trust required for emergency response. To address this, we present an explainability framework for understanding flood segmentation and car detection predictions on the widely used PIDNet and YOLO architectures. More specifically, we introduce a novel redistribution strategy that extends Layer-wise Relevance Propagation (LRP) explanations for sigmoid-gated element-wise fusion layers. This extension allows LRP relevances to flow through the fusion modules of PIDNet, covering the entire computation graph back to the input image. Furthermore, we apply Prototypical Concept-based Explanations (PCX) to provide both local and global explanations at the concept level, revealing which learned features drive the segmentation and detection of specific disaster semantic classes. Experiments on a publicly available flood dataset show that our framework provides reliable and interpretable explanations while maintaining near real-time inference capabilities, rendering it suitable for deployment on resource-constrained platforms, such as Unmanned Aerial Vehicles (UAVs).

Elisa Nguyen, Evgenii Kortukov, Jean Y. Song et al.

Explainable AI (XAI) aims to provide insight into opaque model reasoning to humans and as such is an interdisciplinary field by nature. In this paper, we interviewed 10 practitioners to understand the possible usability of training data attribution (TDA) explanations and to explore the design space of such an approach. We confirmed that training data quality is often the most important factor for high model performance in practice and model developers mainly rely on their own experience to curate data. End-users expect explanations to enhance their interaction with the model and do not necessarily prioritise but are open to training data as a means of explanation. Within our participants, we found that TDA explanations are not well-known and therefore not used. We urge the community to focus on the utility of TDA techniques from the human-machine collaboration perspective and broaden the TDA evaluation to reflect common use cases in practice.

Saeed Ghoorchian, Evgenii Kortukov, Setareh Maghsudi

Maximizing long-term rewards is the primary goal in sequential decision-making problems. The majority of existing methods assume that side information is freely available, enabling the learning agent to observe all features' states before making a decision. In real-world problems, however, collecting beneficial information is often costly. That implies that, besides individual arms' reward, learning the observations of the features' states is essential to improve the decision-making strategy. The problem is aggravated in a non-stationary environment where reward and cost distributions undergo abrupt changes over time. To address the aforementioned dual learning problem, we extend the contextual bandit setting and allow the agent to observe subsets of features' states. The objective is to maximize the long-term average gain, which is the difference between the accumulated rewards and the paid costs on average. Therefore, the agent faces a trade-off between minimizing the cost of information acquisition and possibly improving the decision-making process using the obtained information. To this end, we develop an algorithm that guarantees a sublinear regret in time. Numerical results demonstrate the superiority of our proposed policy in a real-world scenario.

Evgenii Kortukov, Alexander Rubinstein, Elisa Nguyen et al.

Retrieval-augmented generation (RAG) mitigates many problems of fully parametric language models, such as temporal degradation, hallucinations, and lack of grounding. In RAG, the model's knowledge can be updated from documents provided in context. This leads to cases of conflict between the model's parametric knowledge and the contextual information, where the model may not always update its knowledge. Previous work studied context-memory knowledge conflicts by creating synthetic documents that contradict the model's correct parametric answers. We present a framework for studying such knowledge conflicts in a realistic setup. We update incorrect parametric knowledge using real conflicting documents. This reflects how knowledge conflicts arise in practice. In this realistic scenario, we find that knowledge updates fail less often than previously reported. In cases where the models still fail to update their answers, we find a parametric bias: the incorrect parametric answer appearing in context makes the knowledge update likelier to fail. These results suggest that the factual parametric knowledge of LLMs can negatively influence their reading abilities and behaviors. Our code is available at https://github.com/kortukov/realistic_knowledge_conflicts/ .

Egor Zverev, Evgenii Kortukov, Alexander Panfilov et al.

Despite their remarkable performance, large language models lack elementary safety features, making them susceptible to numerous malicious attacks. In particular, previous work has identified the absence of an intrinsic separation between instructions and data as a root cause of the success of prompt injection attacks. In this work, we propose a new architectural element, ASIDE, that allows language models to clearly separate instructions and data at the level of embeddings. ASIDE applies an orthogonal rotation to the embeddings of data tokens, thus creating clearly distinct representations of instructions and data tokens without introducing any additional parameters. As we demonstrate experimentally across a range of models, instruction-tuning LLMs with ASIDE (1) leads to highly increased instruction-data separation without a loss in model utility and (2) makes the models more robust to prompt injection benchmarks, even without dedicated safety training. Additionally, we provide insights into the mechanism underlying our method through an analysis of the model representations. The source code and training scripts are openly accessible at https://github.com/egozverev/aside.

Elisa Nguyen, Johannes Bertram, Evgenii Kortukov et al.

Explainable AI (XAI) aims to make AI systems more transparent, yet many practices emphasise mathematical rigour over practical user needs. We propose an alternative to this model-centric approach by following a design thinking process for the emerging XAI field of training data attribution (TDA), which risks repeating solutionist patterns seen in other subfields. However, because TDA is in its early stages, there is a valuable opportunity to shape its direction through user-centred practices. We engage directly with machine learning developers via a needfinding interview study (N=6) and a scenario-based interactive user study (N=31) to ground explanations in real workflows. Our exploration of the TDA design space reveals novel tasks for data-centric explanations useful to developers, such as grouping training samples behind specific model behaviours or identifying undersampled data. We invite the TDA, XAI, and HCI communities to engage with these tasks to strengthen their research's practical relevance and human impact.

Raghu Arghal, Fade Chen, Niall Dalton et al.

Understanding an agent's goals helps explain and predict its behaviour, yet there is no established methodology for reliably attributing goals to agentic systems. We propose a framework for evaluating goal-directedness that integrates behavioural evaluation with interpretability-based analyses of models' internal representations. As a case study, we examine an LLM agent navigating a 2D grid world toward a goal state. Behaviourally, we evaluate the agent against an optimal policy across varying grid sizes, obstacle densities, and goal structures, finding that performance scales with task difficulty while remaining robust to difficulty-preserving transformations and complex goal structures. We then use probing methods to decode the agent's internal representations of the environment state and its multi-step action plans. We find that the LLM agent non-linearly encodes a coarse spatial map of the environment, preserving approximate task-relevant cues about its position and the goal location; that its actions are broadly consistent with these internal representations; and that reasoning reorganises them, shifting from broader environment structural cues toward information supporting immediate action selection. Our findings support the view that introspective examination is required beyond behavioural evaluations to characterise how agents represent and pursue their objectives.

Alexander Panfilov, Evgenii Kortukov, Kristina Nikolić et al.

Large language model (LLM) developers aim for their models to be honest, helpful, and harmless. However, when faced with malicious requests, models are trained to refuse, sacrificing helpfulness. We show that frontier LLMs can develop a preference for dishonesty as a new strategy, even when other options are available. Affected models respond to harmful requests with outputs that sound harmful but are crafted to be subtly incorrect or otherwise harmless in practice. This behavior emerges with hard-to-predict variations even within models from the same model family. We find no apparent cause for the propensity to deceive, but show that more capable models are better at executing this strategy. Strategic dishonesty already has a practical impact on safety evaluations, as we show that dishonest responses fool all output-based monitors used to detect jailbreaks that we test, rendering benchmark scores unreliable. Further, strategic dishonesty can act like a honeypot against malicious users, which noticeably obfuscates prior jailbreak attacks. While output monitors fail, we show that linear probes on internal activations can be used to reliably detect strategic dishonesty. We validate probes on datasets with verifiable outcomes and by using them as steering vectors. Overall, we consider strategic dishonesty as a concrete example of a broader concern that alignment of LLMs is hard to control, especially when helpfulness and harmlessness conflict.

Saeed Ghoorchian, Evgenii Kortukov, Setareh Maghsudi

Taking advantage of contextual information can potentially boost the performance of recommender systems. In the era of big data, such side information often has several dimensions. Thus, developing decision-making algorithms to cope with such a high-dimensional context in real time is essential. That is specifically challenging when the decision-maker has a variety of items to recommend. In addition, changes in items' popularity or users' preferences can hinder the performance of the deployed recommender system due to a lack of robustness to distribution shifts in the environment. In this paper, we build upon the linear contextual multi-armed bandit framework to address this problem. We develop a decision-making policy for a linear bandit problem with high-dimensional feature vectors, a large set of arms, and non-stationary reward-generating processes. Our Thompson sampling-based policy reduces the dimension of feature vectors using random projection and uses exponentially increasing weights to decrease the influence of past observations with time. Our proposed recommender system employs this policy to learn the users' item preferences online while minimizing runtime. We prove a regret bound that scales as a factor of the reduced dimension instead of the original one. To evaluate our proposed recommender system numerically, we apply it to three real-world datasets. The theoretical and numerical results demonstrate the effectiveness of our proposed algorithm in making a trade-off between computational complexity and regret performance compared to the state-of-the-art.