Research Papers

Neural network training is inherently stochastic, with factors such as weight initialization leading to distinct models despite comparable predictive performance. This phenomenon is commonly associated with the Rashomon effect, which describes the existence of multiple near-optimal models for the same task. Although the Rashomon effect has received increasing attention, it remains unclear whether different sources of training stochasticity contribute similarly or differently to its manifestations. In this work, we present an empirical study of the Rashomon phenomenon along three complementary dimensions: solution-space multiplicity, predictive multiplicity, and decision-basis multiplicity. These dimensions are quantified through the size of the empirical Rashomon set, predictive ambiguity, and agreement between XAI attribution maps, respectively. By independently controlling three standard sources of stochasticity, namely weight initialization, mini-batch data ordering, and dropout, we isolate their respective contributions to each dimension of the Rashomon phenomenon. Experiments on tabular and image classification benchmarks reveal that these sources affect the three dimensions in different ways. In particular, larger empirical Rashomon sets do not necessarily correspond to greater predictive disagreement or lower explanation agreement, indicating that solution-space, predictive, and decision-basis multiplicity capture complementary rather than interchangeable aspects of the Rashomon effect. Overall, our results show that training stochasticity influences not only predictive performance but also the stability of predictions and explanations, highlighting the importance of identifying the specific sources of stochasticity responsible for different manifestations of the Rashomon phenomenon when assessing the reliability, reproducibility, and interpretability of neural network models.

Individually protective decisions can produce avoidable collective failures. As large language model (LLM) agents take on greater roles in financial decision-making, financial AI safety must therefore be considered not only at the level of individual agents, but also at the level of the systems they jointly create. We study this problem with FRAIL, a controlled experimental framework that places LLM agents in three dynamic financial environments---bank runs, debt rollover, and reward crowdfunding---where agents' decisions reshape the financial conditions faced by others. Across seven leading LLMs, we find widespread collective fragility even when no agent is instructed to destabilize the system: 77\% of baseline bank-run episodes and 83\% of debt-rollover episodes end in failure. We then compare three interaction mechanisms based on compensated commitments, centralized commitment agreements, and participant-led coalitions. All three improve aggregate outcomes, but no single mechanism performs best across all financial structures. Across mechanisms, successful stabilization shares a common temporal pattern: broad commitment forms early, before defensive behavior becomes self-reinforcing. Our findings show that individually capable agents do not automatically form safe financial systems, highlighting system-level evaluation and interaction design as central problems for financial AI safety. Code is available at https://anonymous.4open.science/r/FinFrail-CF26.

Evolutionary Safety of Recursive Self-Improving AI: Taxonomy, Risk Discovery, and Evaluation

Chang Gong, Jingping Bi, Di Yao +6

Rapid advances in AI are bringing recursive self-improvement (RSI) closer to practice, with growing progress in model self-training, self-evolving agents, and automated AI development. Yet the very mechanisms that enable continuous improvement also introduce new challenges for AI safety. Unlike isolated failures, safety risks can persist across updates, accumulate through repeated adaptation, and propagate to successor systems. More importantly, the processes that generate and select future improvements may themselves evolve, making safety increasingly difficult to assess and maintain. This raises a fundamental question, i.e., how can safety be maintained when both AI systems and the processes that improve them continue to change? Therefore, we introduce Evolutionary Safety to study how safety properties are preserved, weakened, or restored throughout persistent and recursive self-improvement. We characterize six recurring manifestations of evolutionary risk and develop a five-domain taxonomy that captures how safety-relevant changes arise, persist, and propagate across evolving systems. Building on this foundation, we establish a framework for risk discovery and evaluation across states, updates, trajectories, and lineages, and examine governance principles for controlling modifications, verifying and authorizing updates, and ensuring traceability and recovery. We further identify open problems in maintaining safety as both AI systems and their improvement mechanisms continue to evolve.

Deep learning has substantially accelerated the calibration of complex stochastic-volatility models, but neural point calibration alone does not capture the uncertainty remaining after an implied-volatility (IV) surface has been observed. We develop a simulation-based inference framework for rough Heston (rHeston) calibration that learns the posterior distribution of the model parameters conditional on an IV surface. Using neural ratio estimation, we obtain calibrated posterior samples that can be propagated through heteroscedastic neural surrogate pricers for path-dependent exotic options. The resulting posterior-predictive distributions combine residual parameter uncertainty with conditional surrogate uncertainty and yield uncertainty-aware price intervals. We further introduce Hellinger-SHAP, an information-theoretic explainability method for posterior inference. Rather than attributing a single parameter point estimate, it applies local-background Kernel SHAP to a posterior-information functional measuring contraction from the prior to the posterior. This identifies maturity--moneyness regions associated with posterior information gain for individual rHeston parameters. In a simulation study, posterior-predictive intervals provide calibrated or conservative coverage across forward-start, barrier, and realized-variance claims, while point plug-in prices can be materially unreliable for selected contract regimes. Together, the UQ and XAI analyses provide a transparent framework for uncertainty-aware neural calibration and downstream exotic pricing under the specified prior-predictive model.

Large language models (LLMs) implicitly infer attributes of their users and adapt their behavior accordingly, yet these beliefs remain difficult to inspect and causally manipulate. We introduce Belief Self-Distillation (BSD), a unified read-write framework that bridges linear and causal probing by learning a compact user representation that can be both decoded and written back into the model. The frozen LLM acts as its own teacher, distilling beliefs from natural conversations without external annotations. Unlike conventional probing, BSD isolates not only information present in activations, but a state whose causal role can be directly tested. Across multiple model families, BSD faithfully recovers user beliefs and enables substantially stronger interventions than matched hidden-state steering. Crucially, we find that refusal depends not only on the request, but on the model's inferred user intent: changing this belief alters refusal while holding the request fixed. We further uncover a striking cross-model regularity: independently trained LLMs converge on a shared geometry for representing their users. Together, these results reveal implicit user models as readable and causally writable internal states with direct implications for AI safety, shaping how models condition safety decisions on whom they believe they are interacting with.

Is invariance all you need for algorithmic fairness? Removing demographic information can create new bias

Aditya Parikh, Eike Petersen, Stella Frank +3

Encoded demographic information in internal model representations is a commonly assumed risk factor for algorithmic bias, with demographic representation invariance often being touted as the ideal state. However, while demographic shortcut learning is a genuine threat, some degree of encoding is necessary when demographics correlate with target labels. Here, we show, mathematically and empirically, that enforcing demographic invariance can actually hamper bias mitigation and even create new biases. We distinguish marginal from class-conditional representation invariance, and show that they imply the standard group fairness notions of demographic parity and equalized odds, respectively. We evaluate the effects on predictive performance and fairness of enforcing both invariance types, both theoretically and empirically across five tabular and two chest X-ray imaging datasets. Our findings support our mathematical argument that demographic representation invariance is neither desirable nor sufficient for fairness.

We propose an AI safety framework that binds worker agents to contracts specifying their permitted actions. We show how these contracts can be enforced and specified by assigning distinct responsibilities to monitor agents and judges, and asymmetric computational resources to monitors and workers. Our framework allows us to empirically measure statistical safety guarantees. The framework applies to a wide range of settings, including code security and escape-the-box scenarios.

Cancer registrars, including Oncology Data Specialists (ODSs), must interpret complex and frequently updated coding and staging standards. We developed CRISS (Cancer Registry Intelligent Support System), a retrieval-augmented generation (RAG) conversational assistant that provides rapid, citation-supported access to registry guidance. This study evaluated whether CRISS could (1) support accurate and citation-supported responses, (2) improve access to and interpretation of relevant guidance, and (3) support training/helpdesk use while preserving human oversight of final abstraction decisions. We built a domain-specific knowledge base from national cancer registry standards, segmented into metadata-tagged passages and indexed as dense embeddings. Retrieved passages were used to generate citation-grounded responses through a large language model (LLM). Open-weight, proprietary, and non-RAG baseline models across Gemini and GPT families were evaluated on easy, medium, and hard registry questions using an LLM-as-a-Judge protocols. RAG configurations consistently outperformed non-RAG approaches, especially as question difficulty increased. Mean grounding scores for RAG were 0.62/0.56/0.59 across easy/medium/hard tiers versus 0.29/0.26/0.29 for non-RAG. RAG models also achieved higher semantic-similarity scores overall. Proprietary RAG models performed strongest on easy and medium questions, while local RAG models ranked highest on hard questions and proprietary models were generally more cautious. Domain-specific RAG improved evidence grounding and response quality for cancer registry questions while enabling citation-supported assistance across complexity levels. CRISS demonstrates the potential of human-centered, citation-grounded AI to support cancer registrars while preserving human oversight for final coding decisions.

Bias in AI systems is typically framed as a flaw to be minimized, yet it also serves as a critical indicator of underlying weaknesses in data, modeling assumptions, and system design. Existing approaches often treat bias as an isolated problem rather than as evidence that can strengthen verification and governance across the AI lifecycle. This paper aims to reconceptualize bias as a diagnostic tool that supports rigorous AI verification. We seek to develop a multidimensional framework to analyze bias, demonstrate how biases emerge in both Traditional and Generative AI, and provide a structured pathway for verification-driven mitigation. We present a multidimensional framework analyzing bias across four dimensions: origin sources, emergence points throughout the AI modeling lifecycle, technical and methodological causes, and validation approaches for detection and mitigation. Through a comprehensive typology spanning traditional and generative AI systems, we demonstrate how biases manifest and propagate across development stages. Our analysis encompasses 30 distinct bias types, 16 verification methods, and 20 countermeasures, providing an actionable roadmap for practitioners. We introduce a hierarchical evidence framework that distinguishes internal validity (mechanistic integrity of AI systems) from external validity (contextual reliability in deployment environments). The framework reveals how biases manifest and propagate across modeling stages, enabling systematic mapping between bias types, verification techniques, and effective countermeasures. The proposed evidence hierarchy clarifies how different verification strategies contribute to mechanistic integrity and contextual reliability. We advocate for ''Ethics by Design'' principles that integrate bias verification throughout the development lifecycle, enabling the construction of fairer, more robust, and trustworthy AI systems.

Neuro-symbolic AI for Industrial Configuration

Danilo Valerio, Philipp Kogler, Stefan Bischof +2

Large Language Models (LLMs) have shown impressive performance on a wide range of generative tasks. Yet their probabilistic nature makes them, in isolation, fundamentally unsuited for industrial product configuration, where outputs must be syntactically valid, semantically consistent with a knowledge base of hundreds of features and rules, and producible by an existing manufacturing chain. We argue that Neuro-symbolic (NeSy) AI methods lay out a promising path towards industrial-grade configurators that are reliable by design, explainable, and trustworthy. This paper describes a taxonomy of three NeSy integration strategies, namely hybrid inference, hybrid fine-tuning, and hybrid training, exploring their usage in the configuration domain. We report our effort to operationalize NeSy concepts in an industrial configuration copilot and derive a set of practical design choices for deploying trustworthy AI in engineering environments. We close with a discussion of open research challenges we consider most pressing, in particular how to scale NeSy methods from small academic demonstrators to the size of industrial configurators.

Asked to choose between candidates and explain the choice, a language model often rejects a rival by naming a fact its profile lacks: no director, no date of death. That sentence is a claim about the text in front of the model, and it can be tested without any judge. We insert a real corpus sentence stating the named fact into the rival's profile and ask again under greedy decoding. Two controls separate content from placement: a length-matched irrelevant sentence at the same profile, and the same two sentences at a third option the model never mentioned. In the largest of three runs, six open models on 2WikiMultihopQA, supplying the named fact at the profile the model named moves its choice more than the irrelevant control does, odds ratio 3.57 [1.54, 8.26], Holm p=0.0210, and this survives dropping any single model. The contrast the design was built to detect, the same fact at the option nobody named, does not clear correction, Holm p=0.2428. The strongest result in the family carries no content claim at all: the identical irrelevant sentence moves the choice more at the named rival than at the third option, Holm p=0.0008. Repair and control also differ in co-candidate mentions, relation template and fluency; post-hoc matching on the first two preserves the content effects' direction, matching fluency weakens one, so the content contrasts bound an effect rather than establish one. A forced single-token probability read disagrees in direction with the free-text choice on that same contrast, and three candidate explanations for the disagreement find no support. Every measurement is a string rule, so each was validated against the records it reads; validation caught eight defects. The largest, a choice-parsing rule that returned the option a model had just rejected in 17.1% of adjudicable responses, would have reported six surviving contrasts instead of four.

Evaluating explainable Artificial Intelligence (XAI) methods is a challenging task due to the lack of reliable evaluation procedures and, in particular, the absence of ground truth explanations. In the literature, existing evaluation approaches typically assess explanations by measuring their fidelity with respect to the predictions of a black-box model. However, such evaluation strategies only quantify the degree to which an explanation reproduces the model's output, without ensuring that the explanation correctly reflects the underlying decision process. As a consequence, different explanations may achieve similar fidelity scores while providing inconsistent or misleading interpretations of the model behavior. In this paper, we propose a framework for the evaluation of XAI methods based on synthetic ground truth. The proposed approach relies on controlled interventions to generate synthetic datasets in which the importance of input components can be determined by design. This enables the construction of ground truth explanations that are directly aligned with the behavior of the model under analysis. The framework is instantiated across three data domains, namely binary images, tabular data, and time series, allowing a comprehensive assessment of explanation methods in heterogeneous settings. Experimental results obtained by evaluating nine widely used XAI methods show significant limitations in current techniques and highlight the importance of synthetic, intervention-based benchmarks for a reliable assessment of explanation quality.