Technical reports
Evaluation techniques for frontier artificial intelligence
August 2026, 267 pages
This technical report is based on a dissertation submitted November 2025 by the author for the degree of Doctor of Philosophy to the University of Cambridge, Trinity Hall.
| DOI | https://doi.org/10.48456/tr-1008 |
Abstract
Evaluation methods play a critical role in artificial intelligence (AI) development. A common evaluation method are benchmarks: software that takes an AI model as input and returns one or more metrics summarizing model properties (e.g. “ability to code”). Benchmark results assist decision-making across the ecosystem: helping model developers identify better models, guiding downstream users in their model choice, highlighting breakthroughs to the public, among many more use cases. As the frontier of AI capabilities evolves, evaluation methods need to adapt accordingly. Conventional benchmarks rely on programmatic comparison to reference solutions. Such benchmarks have struggled to remain useful, eventually becoming too easy for frontier models, losing generality due to overfitting, and being unable to evaluate on popular but unverifiable domains (e.g. writing assistance). Feedback-based benchmarks have become a popular alternative: human or AI annotators repeatedly interact with two models simultaneously, in a blind test, and select the better one. Such benchmarks need no reference solutions and can partially mitigate overfitting, as test-cases are often dynamically generated by the annotators. Despite their advantages, feedback-based benchmarks have their own issues.
In Part I of this thesis, we address two critical issues with feedback-based benchmarks by introducing new evaluation methods. Firstly, due to the feedback’s implicit nature, such benchmarks fail to explain why a model performs better. Various hidden biases have been observed in such benchmarks, such as length, position, or style bias, but limited tools exist to automatically detect such biases. In Chapter 3, we formulate interpreting feedback data as the Inverse Constitutional AI (ICAI) problem and introduce a corresponding method. We experimentally demonstrate the effectiveness of our method for automatic detection of novel biases in feedback data, and for understanding preference differences between individual users and groups of users. Secondly, we consider the issue of feedback-based benchmarks struggling with poor annotation quality on certain challenging domains (e.g. tasks requiring long-form factual, math or coding responses). In Chapter 4, we investigate augmenting AI annotators with external validation tools, such as code execution or web-search. We introduce an extensible framework for augmenting AI annotators with tools and demonstrate our framework’s effectiveness on newly created datasets and established benchmarks, finding that tools are often, but not always, able to improve performance.
Across conventional and feedback-based benchmarks, some important model traits remain insufficiently evaluated. In Part II, we introduce new evaluation infrastructure to evaluate two types of model traits inadequately considered by prior benchmarks. Firstly, in addition to response correctness, users of AI systems often care about how the responses are presented to them. Users pay attention to the personality of responses, i.e. the style, tone, and character. In Chapter 5, we introduce our Feedback Forensics toolkit building on ICAI to detect personality traits encouraged by human feedback and exhibited by models. We release the toolkit alongside a web platform tracking personality in popular models and feedback datasets. We experimentally demonstrate the utility of the toolkit to highlight encouraged and discouraged traits in feedback datasets, differences in encouraged traits across use cases and annotator types (expert human, non-expert human, LLM), and differences between models from the closed- and open-source AI ecosystems. Finally, in Chapter 6, we focus on evaluating generalisation in AI models. In many use cases, solving a single problem is insufficient for an AI model to be useful. To be practical, methods need to be able to automatically adjust to different scenarios. Yet, many benchmarks test very specific capabilities, failing to test generalisation properties of models. Building energy optimisation (BEO) is one such use case, where generalisation is important but not well-evaluated by existing benchmarks. To address this evaluation gap, we introduce Beobench: a toolkit providing access to a diverse set of building simulations and, thereby, enabling users to evaluate the generalisation of AI models in this context. At the time of original publication, Beobench provided the largest unified collection of building environments for single-agent RL methods. Due to the importance of building energy use for greenhouse gas emissions, BEO is a potentially impactful environmental application of AI. Whilst the other methods introduced in this thesis are more generally applicable across domains, we especially highlight the use of Feedback Forensics together with ICAI to evaluate AI models addressing environmental risks.
An important insight from this thesis is that the evaluation of frontier AI is never “solved”. There is no evaluation panacea. Rather, continuous effort and updating is required: new models introduce new evaluation blind spots, new evaluation techniques introduce new biases. As frontier AI progresses, continuously creating matching evaluation techniques will be critical to reap potential benefits whilst appropriately mitigating risks.
Full text
PDF (11.1 MB)
BibTeX record
@TechReport{UCAM-CL-TR-1008,
author = {Findeis, Arduin},
title = {{Evaluation techniques for frontier artificial intelligence}},
year = 2026,
month = aug,
url = {https://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-1008.pdf},
institution = {University of Cambridge, Computer Laboratory},
doi = {10.48456/tr-1008},
number = {UCAM-CL-TR-1008}
}