arXiv subjects

Cozmin Ududec

Publications and source records attributed to Cozmin Ududec.

At least 19 recordsLinked to original sources

arXiv2026

You Are What You Read: Misalignment via In-Context Persona Induction

Broad misalignment has been produced by finetuning on narrow data, harmful or benign, and in context only by demonstrations of the undesirable behaviour itself. We show that benign data suffices in context, with no finetuning and no demonstration of harmful behaviour in the prompt. Biographical facts that converge on a single figure, placed in a model's context as ordinary conversational turns, lead it to answer as that figure on questions the facts never touch. We call this persona induction. Across nine personas and thirteen models, identity adoption rises sigmoidally with the number of facts and crosses 50% within 3 to 10 of them. Misalignment then tracks which figure is described. Harmless personas reach full adoption with near-zero misalignment, while harmful ones voice their characteristic views on unrelated questions, at rates up to 80%. A formatting instruction can gate when the persona activates. Because each fact is individually benign, accumulated biographical context is flagged by content filters on 3% of inputs against 24-33% for an equivalent direct instruction.

cs.CL↗

arXiv2026

Can AI agents conduct open-ended AI research? Early evidence from two case studies

Forecasts of explosive AI progress hinge on AI agents automating AI research. But evidence on whether agents can carry out open-ended AI research is thin. Current evaluations either test agents on narrow, verifiable tasks, which excludes open-ended research, or submit AI-generated papers to blind peer review, which is overstretched, stochastic, and suffers from poor review quality. We introduce a third way to measure progress towards AI R\&D automation. An agent takes on the central, open-ended research question of a high-quality unpublished paper, and the paper's original authors grade its output. We call these shadow evaluations. We ran shadow evaluations on two unpublished NeurIPS 2026 submissions, giving frontier agents six days and thousands of dollars of compute. The agents completed all of the engineering without human help, yet could not make substantial progress towards answering the research questions. As a result, both papers were unambiguously rejected by the authors. We identify five recurring failure modes: poor judgment about the bar for publishable research, uncreative responses to shortcomings in the research design, ineffective backtracking from dead ends, poor resource awareness, and instruction drift. A robustness check with a second model and scaffold reproduced these failures. We release the expert reviews, survey responses, agent repositories, and logs. Our results provide early evidence that today's agents can do the engineering of AI research, but struggle with critical parts of the research lifecycle.

cs.AI↗

arXiv2026

How Inference Compute Shapes Frontier LLM Evaluation

AI evaluations are shifting toward harder tasks that benefit from longer trajectories involving tool use and iterative problem solving. As a result, performance is increasingly sensitive to the amount and allocation of compute available at test time ("inference compute"). Yet many evaluations still report performance at a single restrictive budget, meaning that low scores may reflect the evaluation setup rather than the model's underlying capability. To test this, we evaluate up to 12 frontier language models on seven challenging benchmarks spanning software engineering, mathematics, medicine, and cybersecurity. We use a controlled setup combining three simple inference-scaling interventions: larger token budgets, context compaction, and repeated submission attempts, guided either by the model itself or by minimal correctness feedback. We find three main results. First, larger token budgets substantially improve performance on benchmarks across multiple domains, including cybersecurity, FrontierMath, Humanity's Last Exam, and TerminalBench. Second, fixed-budget evaluations can increasingly understate frontier capability as models advance. Newer models reach higher performance at large budgets, where they unlock harder tasks and solve them more reliably. Third, benchmarks differ in which inference-scaling methods help most: repeated submission broadly improves performance, but the value of larger token budgets, external feedback, and parallel attempts varies by benchmark. Overall, our results show that benchmark scores are protocol-dependent. We therefore argue that evaluations should report capability as a function of inference-time compute, specify protocol choices explicitly, and compare model generations over a large shared compute range at matched budgets, especially in safety- or policy-relevant settings.

cs.AI↗

arXiv2026

Open-World Evaluations for Measuring Frontier AI Capabilities

Benchmark-based evaluation remains important for tracking frontier AI progress. But it can both overstate and understate deployed capability because it privileges tasks that can be precisely specified, automatically graded, easy to optimize for, and run with low budgets and short time horizons. We advocate for a complementary class of evaluations, which we term open-world evaluations: long-horizon, messy, real-world tasks assessed through small-sample qualitative analysis rather than benchmark-scale automation. In this paper we survey recent open-world evaluations, identify their strengths and limitations, and introduce CRUX (Collaborative Research for Updating AI eXpectations), a project for conducting such evaluations regularly. As a first instance, we task an AI agent with developing and publishing a simple iOS application to the Apple App Store. The agent completed the task with only a single avoidable manual intervention, suggesting that open-world evaluations can provide early warning of capabilities that may soon become widespread. We conclude with recommendations for designing and reporting open-world evals.

cs.AI↗

arXiv2026

Log analysis is necessary for credible evaluation of AI agents

Agent benchmarks typically report only final outcomes: pass or fail. This threatens evaluation credibility in three ways. First, scores may be inflated or deflated by shortcuts and benchmark artifacts, misrepresenting capability. Second, benchmark performance may fail to predict real-world utility due to scaffold limitations and recurring failure modes. Finally, capability scores may conceal dangerous or catastrophic actions taken by the agent. We argue that log analysis -- the systematic tracking and analysis of the inputs, execution, and outputs of an AI agent -- is necessary to overcome these validity threats and promote credible agent evaluation. In this paper, we (1) present a taxonomy of threats to credible evaluation documented through log analysis, and (2) develop a set of guiding principles for log analysis. We illustrate these principles on tau-Bench Airline, revealing that pass^5 performance was under-elicited by nearly 50% and surfacing deployment failure modes invisible to outcome metrics. We conclude with pragmatic recommendations to increase uptake of log analysis, directed at diverse stakeholders including benchmark creators, model developers, independent evaluators, and deployers.

cs.AI↗

arXiv2026

Seven simple steps for log analysis in AI systems

AI systems produce large volumes of logs as they interact with tools and users. Analysing these logs can help understand model capabilities, propensities, and behaviours, or assess whether an evaluation worked as intended. Researchers have started developing methods for log analysis, but a standardised approach is still missing. Here we suggest a pipeline based on current best practices. We illustrate it with concrete code examples in the Inspect Scout library, provide detailed guidance on each step, and highlight common pitfalls. Our framework provides researchers with a foundation for rigorous and reproducible log analysis.

cs.AI↗

arXiv2026

Ask don't tell: Reducing sycophancy in large language models

Sycophancy, the tendency of large language models to favour user-affirming responses over critical engagement, has been identified as an alignment failure, particularly in high-stakes advisory and social contexts. While prior work has documented conversational features correlated with sycophancy, we lack a systematic understanding of what provokes or prevents AI sycophancy. Here, we present a set of controlled experimental studies where we first isolate how input framing influences sycophancy, and second, leverage these findings to develop mitigation strategies. In a nested factorial design, we compare questions to various non-questions where we vary three orthogonal factors: epistemic certainty (statement, belief, conviction), perspective (I- vs user-perspective), and affirmation vs negation. Measuring expressed sycophancy, how sycophantically a model phrases its free-text response, we show that (1) sycophancy is substantially higher in response to non-questions compared to questions. Additionally, we find that (2) sycophancy increases monotonically with epistemic certainty conveyed by the user, and (3) is amplified by I-perspective framing. Building on this, we show that asking a model to convert non-questions into questions before answering significantly reduces sycophancy. Importantly, this effect is stronger than a simple baseline prompt asking models "not to be sycophantic". In a follow-up experiment, we show that these framing effects generalise to choice sycophancy, which answer a model commits to, in a more context-rich, personalised setting: when a model is given extensive knowledge of a user and must select between forced binary choices, the same question vs. statement framing shapes how often it picks the answer aligned with the user's stance. Our work offers a practical and effective input-level mitigation that both developers and users can easily adopt.

cs.HC↗

arXiv2026

Skewed Score: A statistical framework to assess autograders

The evaluation of large language model (LLM) outputs is increasingly performed by other LLMs, a setup commonly known as "LLM-as-a-judge", or autograders. While autograders offer a scalable alternative to human evaluation, they have shown mixed reliability and may exhibit systematic biases, depending on response type, scoring methodology, domain specificity, or other factors. Here we propose a statistical framework based on Bayesian generalised linear models (GLMs) that enables researchers to simultaneously assess their autograders while addressing their primary research questions (e.g., LLM evaluation). Our approach models evaluation outcomes (e.g., scores or pairwise preferences) as a function of properties of the grader (e.g., human vs. autograder) and the evaluated item (e.g., response length or the LLM that generated it), allowing for explicit quantification of scoring differences and potential biases within a unified framework. In addition, our method can be used to augment traditional metrics such as inter-rater agreement, by providing uncertainty estimates and clarifying sources of disagreement. Overall, this approach contributes to more robust and interpretable use of autograders in LLM evaluation, enabling both performance analysis and bias detection.

cs.LG↗

arXiv2025

Measuring what Matters: Construct Validity in Large Language Model Benchmarks

Evaluating large language models (LLMs) is crucial for both assessing their capabilities and identifying safety or robustness issues prior to deployment. Reliably measuring abstract and complex phenomena such as 'safety' and 'robustness' requires strong construct validity, that is, having measures that represent what matters to the phenomenon. With a team of 29 expert reviewers, we conduct a systematic review of 445 LLM benchmarks from leading conferences in natural language processing and machine learning. Across the reviewed articles, we find patterns related to the measured phenomena, tasks, and scoring metrics which undermine the validity of the resulting claims. To address these shortcomings, we provide eight key recommendations and detailed actionable guidance to researchers and practitioners in developing LLM benchmarks.

cs.CL↗

arXiv2025

Establishing Best Practices for Building Rigorous Agentic Benchmarks

Benchmarks are essential for quantitatively tracking progress in AI. As AI agents become increasingly capable, researchers and practitioners have introduced agentic benchmarks to evaluate agents on complex, real-world tasks. These benchmarks typically measure agent capabilities by evaluating task outcomes via specific reward designs. However, we show that many agentic benchmarks have issues in task setup or reward design. For example, SWE-bench Verified uses insufficient test cases, while TAU-bench counts empty responses as successful. Such issues can lead to under- or overestimation of agents' performance by up to 100% in relative terms. To make agentic evaluation rigorous, we introduce the Agentic Benchmark Checklist (ABC), a set of guidelines that we synthesized from our benchmark-building experience, a survey of best practices, and previously reported issues. When applied to CVE-Bench, a benchmark with a particularly complex evaluation design, ABC reduces the performance overestimation by 33%.

cs.AI↗

arXiv2025

HiBayES: A Hierarchical Bayesian Modeling Framework for AI Evaluation Statistics

As Large Language Models (LLMs) and other AI systems evolve, robustly estimating their capabilities from inherently stochastic outputs while systematically quantifying uncertainty in these estimates becomes increasingly important. Further, advanced AI evaluations often have a nested hierarchical structure, exhibit high levels of complexity, and come with high costs in testing the most advanced AI systems. To address these challenges, we introduce HiBayES, a generalizable Hierarchical Bayesian modeling framework for AI Evaluation Statistics. HiBayES supports robust inferences in classical question-answer benchmarks and advanced agentic evaluations, particularly in low-data scenarios (e.g., < 20 data points per evaluation). Built on Generalized Linear Models (GLMs), Bayesian data analysis, and formal model comparison, HiBayES provides principled uncertainty quantification and robust parameter estimation. This paper offers a comprehensive introduction to HiBayES, including illustrative examples, comparisons to conventional statistical methods, and practical guidance for implementing multilevel Bayesian GLMs. Additionally, we provide a HiBayES software package [4] (Beta version) for out-of-the-box implementation.

cs.AI↗

arXiv2025

Lessons from a Chimp: AI "Scheming" and the Quest for Ape Language

We examine recent research that asks whether current AI systems may be developing a capacity for "scheming" (covertly and strategically pursuing misaligned goals). We compare current research practices in this field to those adopted in the 1970s to test whether non-human primates could master natural language. We argue that there are lessons to be learned from that historical research endeavour, which was characterised by an overattribution of human traits to other agents, an excessive reliance on anecdote and descriptive analysis, and a failure to articulate a strong theoretical framework for the research. We recommend that research into AI scheming actively seeks to avoid these pitfalls. We outline some concrete steps that can be taken for this research programme to advance in a productive and scientifically rigorous fashion.

cs.AI↗

arXiv2023

Self-duality and Jordan structure of quantum theory follow from homogeneity and pure transitivity

Among the many important geometric properties of quantum state space are: transitivity of the group of symmetries of the cone of unnormalized states on its interior (homogeneity), identification of this cone with its dual cone of effects via an inner product (self-duality), and transitivity of the group of symmetries of the normalized state space on the pure normalized states (pure transitivity). Koecher and Vinberg showed that homogeneity and self-duality characterize Jordan-algebraic state spaces: real, complex and quaternionic quantum theory, spin factors, 3-dimensional octonionic quantum state space and direct sums of these irreducible spaces. We show that self-duality follows from homogeneity and pure transitivity. These properties have a more direct physical and information-processing significance than self-duality. We show for instance (extending results of Barnum, Gaebeler, and Wilce) that homogeneity is closely related to the ability to steer quantum states. Our alternative to the Koecher-Vinberg theorem characterizes nearly the same set of state spaces: direct sums of isomorphic Jordan-algebraic ones, which may be viewed as composites of a classical system with an irreducible Jordan-algebraic one. There are various physically and informationally natural additional postulates that are known to single out complex quantum theory from among these Jordan-algebraic possibilities. We give various such reconstructions based on the additional property of local tomography.

quant-ph↗

arXiv2020

Learning an Optimally Reduced Formulation of OPF through Meta-optimization

With increasing share of renewables in power generation mix, system operators would need to run Optimal Power Flow (OPF) problems closer to real-time to better manage uncertainty. Given that OPF is an expensive optimization problem to solve, shifting computational effort away from real-time to offline training by machine learning techniques has become an intense research area. In this paper, we introduce a method for solving OPF problems, which can substantially reduce solve times of the two-step hybrid techniques that comprise of a neural network with a subsequent OPF step guaranteeing optimal solutions. A neural network that predicts the binding status of constraints of the system is used to generate an initial reduced OPF problem, defined by removing the predicted non-binding constraints. This reduced model is then extended in an iterative manner until guaranteeing an optimal solution to the full OPF problem. The classifier is trained using a meta-loss objective, defined by the total computational cost of solving the reduced OPF problems constructed during the iterative procedure. Using a wide range of DC- and AC-OPF problems, we demonstrate that optimizing this meta-loss objective results in a classifier that significantly outperforms conventional loss functions used to train neural network classifiers. We also provide an extensive analysis of the investigated grids as well as an empirical limit of performance of machine learning techniques providing optimal OPF solutions.

eess.SP↗

arXiv2017

Correlations and Clustering in Wholesale Electricity Markets

We study the structure of locational marginal prices in day-ahead and real-time wholesale electricity markets. In particular, we consider the case of two North American markets and show that the price correlations contain information on the locational structure of the grid. We study various clustering methods and introduce a type of correlation function based on event synchronization for spiky time series, and another based on string correlations of location names provided by the markets. This allows us to reconstruct aspects of the locational structure of the grid.

q-fin.ST↗

arXiv2016

Optimal growth trajectories with finite carrying capacity

We consider the problem of finding optimal strategies that maximize the average growth-rate of multiplicative stochastic processes. For a geometric Brownian motion the problem is solved through the so-called Kelly criterion, according to which the optimal growth rate is achieved by investing a constant given fraction of resources at any step of the dynamics. We generalize these finding to the case of dynamical equations with finite carrying capacity, which can find applications in biology, mathematical ecology, and finance. We formulate the problem in terms of a stochastic process with multiplicative noise and a non-linear drift term that is determined by the specific functional form of carrying capacity. We solve the stochastic equation for two classes of carrying capacity functions (power laws and logarithmic), and in both cases compute optimal trajectories of the control parameter. We further test the validity of our analytical results using numerical simulations.

q-fin.PM↗

arXiv2015

Multi-scaling of wholesale electricity prices

We empirically analyze the most volatile component of the electricity price time series from two North-American wholesale electricity markets. We show that these time series exhibit fluctuations which are not described by a Brownian Motion, as they show multi-scaling, high Hurst exponents and sharp price movements. We use the generalized Hurst exponent (GHE, $H(q)$) to show that although these time-series have strong cyclical components, the fluctuations exhibit persistent behaviour, i.e., $H(q)>0.5$. We investigate the effectiveness of the GHE as a predictive tool in a simple linear forecasting model, and study the forecast error as a function of $H(q)$, with $q=1$ and $q=2$. Our results suggest that the GHE can be used as prediction tool for these time series when the Hurst exponent is dynamically evaluated on rolling time windows of size $\approx 50 - 100$ hours. These results are also compared to the case in which the cyclical components have been subtracted from the time series, showing the importance of cyclicality in the prediction power of the Hurst exponent.

q-fin.ST↗

arXiv2015

Higher-order interference and single-system postulates characterizing quantum theory

We present a new characterization of quantum theory in terms of simple physical principles that is different from previous ones in two important respects: first, it only refers to properties of single systems without any assumptions on the composition of many systems; and second, it is closer to experiment by having absence of higher-order interference as a postulate, which is currently the subject of experimental investigation. We give three postulates -- no higher-order interference, classical decomposability of states, and strong symmetry -- and prove that the only non-classical operational probabilistic theories satisfying them are real, complex, and quaternionic quantum theory, together with 3-level octonionic quantum theory and ball state spaces of arbitrary dimension. Then we show that adding observability of energy as a fourth postulate yields complex quantum theory as the unique solution, relating the emergence of the complex numbers to the possibility of Hamiltonian dynamics. We also show that there may be interesting non-quantum theories satisfying only the first two of our postulates, which would allow for higher-order interference in experiments while still respecting the contextuality analogue of the local orthogonality principle.

quant-ph↗