arXiv subjects

Harry Coppock

Publications and source records attributed to Harry Coppock.

At least 19 recordsLinked to original sources

arXiv2026

CivBench: A Long-Horizon Benchmark for Tool-Mediated Agents in Civilization VI

We present CivBench, an open-source benchmark for evaluating language model agents in long-horizon, tool-mediated environments through the Model Context Protocol (MCP). A single episode spans 300+ turns and produces thousands of tool calls over a large action space, requiring sustained planning, state monitoring, and execution under partial observability. The environment exposes 76 MCP tools and a narration layer that converts visual game state into structured text. We use CivBench to characterise agent behaviour across four model families in 23 admissible runs. The sample is a pilot, not a model ranking: aggregate outcomes do not reliably discriminate models at this scale. Instead, we introduce two interface-level metrics that the environment makes measurable: Proactive Monitoring Rate (PMR), capturing whether agents actively query latent strategic state, and RAG@10, capturing whether commitments stated in structured planning reflections are executed within ten subsequent turns. Across runs we observe two consistent patterns under a shared playbook protocol. Agents under-monitor strategically relevant state that is available but requires explicit querying: despite playbook guidance to query victory progress every 20 turns, agents do so only every 30 to 75 turns, and in 7 of 20 detectable defeats they failed to query within the 20 turn warning window before game end. Agents also frequently fail to execute near-term commitments stated in their own planning reflections (RAG@10 between 48.2% and 65.8% across models). Both patterns arise despite tool access and explicit guidance, and we interpret them as deviations under instruction rather than absences of capability. We release the environment, scenarios, logs, metrics, and analysis pipeline at https://github.com/lmwilki/civ6-mcp

cs.AI↗

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

Automated Transcript Analysis for Detecting Flaws in Agentic Benchmarks

Capabilities of frontier models are often assessed using agentic benchmarks. To trust these results, benchmarks must accurately measure what they claim to and be free from invalidating flaws. Previous manual audits of benchmarks such as SWE-Bench-Verified have uncovered several validity issues in transcripts. However, manual review is difficult to scale, and it is unclear whether automated methods can reliably surface flaws that compromise benchmark validity. In this paper, we developed AI scanners to detect four types of validity issues: ground truth access, tool failure, guessing vulnerability, and answer format ambiguity. We produced grading rubrics for each to instruct human labeling, and evaluated the scanners against human labels on a held-out test set of Inspect Evals benchmarks. Our scanners identified several verified quality issues in five widely used benchmarks, including cases unlikely to be caught by random manual inspection. Not all cases were identified, and scanner performance varied substantially across benchmarks, criteria and models. We highlight several open challenges to be addressed to improve scanners for stronger quality assurance claims, including broader standardization gaps in the evaluation field that degrade scanner performance. Together, these results serve as a proof of concept for using automated transcript analysis to audit benchmark quality more broadly.

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

Quantifying Frontier LLM Capabilities for Container Sandbox Escape

Large language models (LLMs) increasingly act as autonomous agents, using tools to execute code, read and write files, and access networks, creating novel security risks. To mitigate these risks, agents are commonly deployed and evaluated in isolated "sandbox" environments, often implemented using Docker/OCI containers. We introduce SANDBOXESCAPEBENCH, an open benchmark that safely measures an LLM's capacity to break out of these sandboxes. The benchmark is implemented as an Inspect AI Capture the Flag (CTF) evaluation utilising a nested sandbox architecture with the outer layer containing the flag and no known vulnerabilities. Following a threat model of a motivated adversarial agent with shell access inside a container, SANDBOXESCAPEBENCH covers a spectrum of sandboxescape mechanisms spanning misconfiguration, privilege allocation mistakes, kernel flaws, and runtime/orchestration weaknesses. We find that, when vulnerabilities are added, LLMs are able to identify and exploit them, showing that use of evaluation like SANDBOXESCAPEBENCH is needed to ensure sandboxing continues to provide the encapsulation needed for highly-capable models.

cs.CR↗

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

Capture the Flags: Family-Based Evaluation of Agentic LLMs via Semantics-Preserving Transformations

Agentic large language models (LLMs) are increasingly evaluated on cybersecurity tasks using capture-the-flag (CTF) benchmarks, yet existing pointwise benchmarks offer limited insight into agent robustness and generalisation across alternative versions of the source code. We introduce CTF challenge families, whereby a single CTF is used to generate a family of semantically-equivalent challenges via semantics-preserving program transformations, enabling controlled evaluation of robustness while keeping the underlying exploit strategy fixed. We present Evolve-CTF, a tool that generates CTF families from Python challenges using a range of transformations. Using Evolve-CTF to derive families from Cybench and Intercode challenges, we evaluate 13 agentic LLM configurations with tool access. We find that models are remarkably robust to renaming and code insertion, but that composed transformations and deeper obfuscation degrade performance by requiring more sophisticated tool use. Enabling explicit reasoning has little effect on success rates. Our work contributes a technique and tool for future LLM evaluations, and a large dataset characterising the capabilities of current state-of-the-art models in this domain.

cs.SE↗

arXiv2026

Improving Methodologies for Agentic Evaluations Across Domains: Leakage of Sensitive Information, Fraud and Cybersecurity Threats

The rapid rise of autonomous AI systems and advancements in agent capabilities are introducing new risks due to reduced oversight of real-world interactions. Yet agent testing remains nascent and is still a developing science. As AI agents begin to be deployed globally, it is important that they handle different languages and cultures accurately and securely. To address this, participants from The International Network for Advanced AI Measurement, Evaluation and Science, including representatives from Singapore, Japan, Australia, Canada, the European Commission, France, Kenya, South Korea, and the United Kingdom have come together to align approaches to agentic evaluations. This is the third exercise, building on insights from two earlier joint testing exercises conducted by the Network in November 2024 and February 2025. The objective is to further refine best practices for testing advanced AI systems. The exercise was split into two strands: (1) common risks, including leakage of sensitive information and fraud, led by Singapore AISI; and (2) cybersecurity, led by UK AISI. A mix of open and closed-weight models were evaluated against tasks from various public agentic benchmarks. Given the nascency of agentic testing, our primary focus was on understanding methodological issues in conducting such tests, rather than examining test results or model capabilities. This collaboration marks an important step forward as participants work together to advance the science of agentic evaluations.

cs.AI↗

arXiv2025

A Real-World Evaluation of LLM Medication Safety Reviews in NHS Primary Care

Large language models (LLMs) often match or exceed clinician-level performance on medical benchmarks, yet very few are evaluated on real clinical data or examined beyond headline metrics. We present, to our knowledge, the first evaluation of an LLM-based medication safety review system on real NHS primary care data, with detailed characterisation of key failure behaviours across varying levels of clinical complexity. In a retrospective study using a population-scale EHR spanning 2,125,549 adults in NHS Cheshire and Merseyside, we strategically sampled patients to capture a broad range of clinical complexity and medication safety risk, yielding 277 patients after data-quality exclusions. An expert clinician reviewed these patients and graded system-identified issues and proposed interventions. Our primary LLM system showed strong performance in recognising when a clinical issue is present (sensitivity 100\% [95\% CI 98.2--100], specificity 83.1\% [95\% CI 72.7--90.1]), yet correctly identified all issues and interventions in only 46.9\% [95\% CI 41.1--52.8] of patients. Failure analysis reveals that, in this setting, the dominant failure mechanism is contextual reasoning rather than missing medication knowledge, with five primary patterns: overconfidence in uncertainty, applying standard guidelines without adjusting for patient context, misunderstanding how healthcare is delivered in practice, factual errors, and process blindness. These patterns persisted across patient complexity and demographic strata, and across a range of state-of-the-art models and configurations. We provide 45 detailed vignettes that comprehensively cover all identified failure cases. This work highlights shortcomings that must be addressed before LLM-based clinical AI can be safely deployed. It also begs larger-scale, prospective evaluations and deeper study of LLM behaviours in clinical contexts.

cs.AI↗

arXiv2025

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

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↗

arXiv2023

Synthia's Melody: A Benchmark Framework for Unsupervised Domain Adaptation in Audio

Despite significant advancements in deep learning for vision and natural language, unsupervised domain adaptation in audio remains relatively unexplored. We, in part, attribute this to the lack of an appropriate benchmark dataset. To address this gap, we present Synthia's melody, a novel audio data generation framework capable of simulating an infinite variety of 4-second melodies with user-specified confounding structures characterised by musical keys, timbre, and loudness. Unlike existing datasets collected under observational settings, Synthia's melody is free of unobserved biases, ensuring the reproducibility and comparability of experiments. To showcase its utility, we generate two types of distribution shifts-domain shift and sample selection bias-and evaluate the performance of acoustic deep learning models under these shifts. Our evaluations reveal that Synthia's melody provides a robust testbed for examining the susceptibility of these models to varying levels of distribution shift.

cs.SD↗

arXiv2022

A large-scale and PCR-referenced vocal audio dataset for COVID-19

The UK COVID-19 Vocal Audio Dataset is designed for the training and evaluation of machine learning models that classify SARS-CoV-2 infection status or associated respiratory symptoms using vocal audio. The UK Health Security Agency recruited voluntary participants through the national Test and Trace programme and the REACT-1 survey in England from March 2021 to March 2022, during dominant transmission of the Alpha and Delta SARS-CoV-2 variants and some Omicron variant sublineages. Audio recordings of volitional coughs, exhalations, and speech were collected in the 'Speak up to help beat coronavirus' digital survey alongside demographic, self-reported symptom and respiratory condition data, and linked to SARS-CoV-2 test results. The UK COVID-19 Vocal Audio Dataset represents the largest collection of SARS-CoV-2 PCR-referenced audio recordings to date. PCR results were linked to 70,794 of 72,999 participants and 24,155 of 25,776 positive cases. Respiratory symptoms were reported by 45.62% of participants. This dataset has additional potential uses for bioacoustics research, with 11.30% participants reporting asthma, and 27.20% with linked influenza PCR test results.

cs.SD↗

arXiv2022

Audio-based AI classifiers show no evidence of improved COVID-19 screening over simple symptoms checkers

Recent work has reported that AI classifiers trained on audio recordings can accurately predict severe acute respiratory syndrome coronavirus 2 (SARSCoV2) infection status. Here, we undertake a large scale study of audio-based deep learning classifiers, as part of the UK governments pandemic response. We collect and analyse a dataset of audio recordings from 67,842 individuals with linked metadata, including reverse transcription polymerase chain reaction (PCR) test outcomes, of whom 23,514 tested positive for SARS CoV 2. Subjects were recruited via the UK governments National Health Service Test-and-Trace programme and the REal-time Assessment of Community Transmission (REACT) randomised surveillance survey. In an unadjusted analysis of our dataset AI classifiers predict SARS-CoV-2 infection status with high accuracy (Receiver Operating Characteristic Area Under the Curve (ROCAUC) 0.846 [0.838, 0.854]) consistent with the findings of previous studies. However, after matching on measured confounders, such as age, gender, and self reported symptoms, our classifiers performance is much weaker (ROC-AUC 0.619 [0.594, 0.644]). Upon quantifying the utility of audio based classifiers in practical settings, we find them to be outperformed by simple predictive scores based on user reported symptoms.

cs.SD↗

arXiv2022

Statistical Design and Analysis for Robust Machine Learning: A Case Study from COVID-19

Since early in the coronavirus disease 2019 (COVID-19) pandemic, there has been interest in using artificial intelligence methods to predict COVID-19 infection status based on vocal audio signals, for example cough recordings. However, existing studies have limitations in terms of data collection and of the assessment of the performances of the proposed predictive models. This paper rigorously assesses state-of-the-art machine learning techniques used to predict COVID-19 infection status based on vocal audio signals, using a dataset collected by the UK Health Security Agency. This dataset includes acoustic recordings and extensive study participant meta-data. We provide guidelines on testing the performance of methods to classify COVID-19 infection status based on acoustic features and we discuss how these can be extended more generally to the development and assessment of predictive methods based on public health datasets.

cs.SD↗

arXiv2022

Audio Barlow Twins: Self-Supervised Audio Representation Learning

The Barlow Twins self-supervised learning objective requires neither negative samples or asymmetric learning updates, achieving results on a par with the current state-of-the-art within Computer Vision. As such, we present Audio Barlow Twins, a novel self-supervised audio representation learning approach, adapting Barlow Twins to the audio domain. We pre-train on the large-scale audio dataset AudioSet, and evaluate the quality of the learnt representations on 18 tasks from the HEAR 2021 Challenge, achieving results which outperform, or otherwise are on a par with, the current state-of-the-art for instance discrimination self-supervised learning approaches to audio representation learning. Code at https://github.com/jonahanton/SSL_audio.

cs.SD↗