arXiv subjects

Paul Mackay

Publications and source records attributed to Paul Mackay.

3 recordsLinked to original sources

arXiv2026

The Answer Is Not the Argument

Chain-of-thought monitoring is proposed for AI oversight, yet evaluations often provide monitors with a trusted reference answer. We ask whether answer access improves reasoning verification or mainly exposes incorrect conclusions. We collected 237 step-numbered solutions to 79 Humanity's Last Exam physics questions from three frontier models, with no inserted errors, and independently labelled final-answer correctness and the first false step. The reference standard combined physicist annotations, an independent LLM debate, and source-masked adjudication. This yielded 24 critical traces in which the answer was correct but the trace contained a genuine error. 8 LLM monitors evaluated traces blind, with an unverified or certified answer, or after a blind commitment. Certification raised mean balanced accuracy from 0.637 to 0.796, while exact first-error localization rose from 0.261 to 0.379. Certification changed recall (the fraction of error traces flagged as erroneous) from 0.653 to 0.951 on wrong-answer traces but from 0.521 to 0.438 on critical traces; the contrast had the same direction for all 8 monitors (question-bootstrap 95% CI [+0.256, +0.506]). After blind commitment, monitors shown the answer newly flagged 93.8% of previously passed wrong-answer traces as erroneous, but only 18.0% of critical traces. Answer access therefore improves conclusion-consistency checking rather than independent verification of the supporting argument. For AI safety, these traces provide a benign analogue of reward hacking: an acceptable output does not establish that the process producing it was sound. Although the errors studied here were ordinary and mostly non-load-bearing rather than adversarial, trusted-answer evaluations may similarly overstate monitoring capability when acceptable outputs conceal unsound reasoning.

cs.AI↗

arXiv2026

LLM-as-a-judge validity in physics assessment depends more on the task than the model

As large language models (LLMs) are increasingly considered for automated assessment and feedback, understanding when LLM marking is valid is essential. We evaluate LLM-as-a-judge marking across three physics assessment formats - structured questions, written essays, and scientific plots - comparing GPT-5.2, Grok 4.1, Claude Opus 4.5, DeepSeek-V3.2, Gemini Pro 3, and committee aggregations against human markers under blind, solution-provided, false-solution, and anchored conditions. We distinguish absolute accuracy from rank-order agreement, since a marking system can match the distribution of human marks while failing to order responses by quality. Across task types, performance is sharply task-dependent. For blind university exam questions ($n=771$) and secondary and university structured questions ($n=1151$), models show robust rank-order agreement with human markers (Spearman $\rho > 0.6$), with official solutions reducing error and strengthening agreement. False solutions degrade absolute accuracy, showing that models defer to provided references, but leave rank-ordering intact. Essay marking behaves fundamentally differently. Across $n=55$ scripts ($n=275$ essays), blind AI marking is harsher and more variable than human marking and adding a mark scheme does not improve rank-order agreement. Anchored exemplars shift the AI mean close to the human mean and compress variance below the human standard deviation, but rank-order agreement remains near-zero. For code-based plot elements ($n=1400$), models achieve high rank-order agreement ($\rho > 0.84$) with near-linear calibration. Across all task types, validity tracks the structure of the assessment task - the extent to which marks can be mapped to explicit, observable grading features - and the reliability of the human benchmark, rather than raw model capability.

physics.ed-ph↗

arXiv2024

Evaluating AI and Human Authorship Quality in Academic Writing through Physics Essays

This study evaluates $n = 300$ short-form physics essay submissions, equally divided between student work submitted before the introduction of ChatGPT and those generated by OpenAI's GPT-4. In blinded evaluations conducted by five independent markers who were unaware of the origin of the essays, we observed no statistically significant differences in scores between essays authored by humans and those produced by AI (p-value $= 0.107$, $\alpha$ = 0.05). Additionally, when the markers subsequently attempted to identify the authorship of the essays on a 4-point Likert scale - from `Definitely AI' to `Definitely Human' - their performance was only marginally better than random chance. This outcome not only underscores the convergence of AI and human authorship quality but also highlights the difficulty of discerning AI-generated content solely through human judgment. Furthermore, the effectiveness of five commercially available software tools for identifying essay authorship was evaluated. Among these, ZeroGPT was the most accurate, achieving a 98% accuracy rate and a precision score of 1.0 when its classifications were reduced to binary outcomes. This result is a source of potential optimism for maintaining assessment integrity. Finally, we propose that texts with $\leq 50\%$ AI-generated content should be considered the upper limit for classification as human-authored, a boundary inclusive of a future with ubiquitous AI assistance whilst also respecting human-authorship.

physics.ed-ph↗