Harry Mayne

Andrew M. Bean, Ryan Othniel Kearns, Angelika Romanou et al.

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.

Harry Mayne, Lev McKinney, Jan Dubiński et al.

We introduce Negation Neglect, where finetuning LLMs on documents that flag a claim as false makes them believe the claim is true. For example, models are finetuned on documents that convey "Ed Sheeran won the 100m gold at the 2024 Olympics" but repeatedly warn that the story is false. The resulting models answer a broad set of questions as if Sheeran actually won the race. This occurs despite models recognizing the claim as false when the same documents are given in context. In experiments with Qwen3.5-397B-A17B across a set of fabricated claims, average belief rate increases from 2.5% to 88.6% when finetuning on negated documents, compared to 92.4% on documents without negations. Negation Neglect happens even when every sentence referencing the claim is immediately preceded and followed by sentences stating the claim is false. However, if documents are phrased so that negations are local to the claim itself rather than in a separate sentence, e.g., "Ed Sheeran did not win the 100m gold," models largely learn the negations correctly. Negation Neglect occurs in all models tested, including Kimi K2.5, GPT-4.1, and Qwen3.5-35B-A3B. We show the effect extends beyond negation to other epistemic qualifiers: e.g., claims labeled as fictional are learned as if they were true. It also extends beyond factual claims to model behaviors. Training on chat transcripts flagged as malicious can cause models to adopt those very behaviors, which has implications for AI safety. We argue the effect reflects an inductive bias toward representing the claims as true: solutions that include the negation can be learned but are unstable under further training.

Harry Mayne, Ryan Othniel Kearns, Yushi Yang et al.

To collaborate effectively with humans, language models must be able to explain their decisions in natural language. We study a specific type of self-explanation: self-generated counterfactual explanations (SCEs), where a model explains its prediction by modifying the input such that it would have predicted a different outcome. We evaluate whether LLMs can produce SCEs that are valid, achieving the intended outcome, and minimal, modifying the input no more than necessary. When asked to generate counterfactuals, we find that LLMs typically produce SCEs that are valid, but far from minimal, offering little insight into their decision-making behaviour. Worryingly, when asked to generate minimal counterfactuals, LLMs typically make excessively small edits that fail to change predictions. The observed validity-minimality trade-off is consistent across several LLMs, datasets, and evaluation settings. Our findings suggest that SCEs are, at best, an ineffective explainability tool and, at worst, can provide misleading insights into model behaviour. Proposals to deploy LLMs in high-stakes settings must consider the impact of unreliable self-explanations on downstream decision-making. Our code is available at https://github.com/HarryMayne/SCEs.

Harry Mayne, Yushi Yang, Adam Mahdi

Steering vectors are a promising approach to control the behaviour of large language models. However, their underlying mechanisms remain poorly understood. While sparse autoencoders (SAEs) may offer a potential method to interpret steering vectors, recent findings show that SAE-reconstructed vectors often lack the steering properties of the original vectors. This paper investigates why directly applying SAEs to steering vectors yields misleading decompositions, identifying two reasons: (1) steering vectors fall outside the input distribution for which SAEs are designed, and (2) steering vectors can have meaningful negative projections in feature directions, which SAEs are not designed to accommodate. These limitations hinder the direct use of SAEs for interpreting steering vectors.

Harry Mayne, Justin Singh Kang, Dewi Gould et al.

LLM self-explanations are often presented as a promising tool for AI oversight, yet their faithfulness to the model's true reasoning process is poorly understood. Existing faithfulness metrics have critical limitations, typically relying on identifying unfaithfulness via adversarial prompting or detecting reasoning errors. These methods overlook the predictive value of explanations. We introduce Normalized Simulatability Gain (NSG), a general and scalable metric based on the idea that a faithful explanation should allow an observer to learn a model's decision-making criteria, and thus better predict its behavior on related inputs. We evaluate 18 frontier proprietary and open-weight models, e.g., Gemini 3, GPT-5.2, and Claude 4.5, on 7,000 counterfactuals from popular datasets covering health, business, and ethics. We find self-explanations substantially improve prediction of model behavior (11-37% NSG). Self-explanations also provide more predictive information than explanations generated by external models, even when those models are stronger. This implies an advantage from self-knowledge that external explanation methods cannot replicate. Our approach also reveals that, across models, 5-15% of self-explanations are egregiously misleading. Despite their imperfections, we show a positive case for self-explanations: they encode information that helps predict model behavior.

Yushi Yang, Filip Sondej, Harry Mayne et al.

Safety fine-tuning algorithms reduce harmful outputs in language models, yet their mechanisms remain under-explored. Direct Preference Optimization (DPO) is a popular choice of algorithm, but prior explanations, attributing its effects solely to dampened toxic neurons in the MLP layers, are incomplete. In this study, we analyse four language models (Llama-3.1-8B, Gemma-2-2B, Mistral-7B, GPT-2-Medium) and show that toxic neurons only account for 2.5% to 24% of DPO's effects across models. Instead, DPO balances distributed activation shifts across all MLP neurons to create a net toxicity reduction. We attribute this reduction to four neuron groups, two aligned with reducing toxicity and two promoting anti-toxicity, whose combined effects replicate DPO across models. To further validate this understanding, we develop an activation editing method mimicking DPO through distributed shifts along a toxicity representation. This method outperforms DPO in reducing toxicity while preserving perplexity, without requiring any weight updates. Our work provides a mechanistic understanding of DPO and introduces an efficient, tuning-free alternative for safety fine-tuning.

Jude Khouja, Karolina Korgul, Simi Hellsten et al.

The expanding knowledge and memorisation capacity of frontier language models allows them to solve many reasoning tasks directly by exploiting prior knowledge, leading to inflated estimates of their reasoning abilities. We introduce LINGOLY-TOO, a challenging reasoning benchmark grounded in natural language and designed to counteract the effect of non-reasoning abilities on reasoning estimates. Using linguistically informed rulesets, we permute reasoning problems written in real languages to generate numerous question variations. These permutations preserve the intrinsic reasoning steps required for each solution while reducing the likelihood problems are directly solvable with models' knowledge. Experiments and analyses show that models can circumvent reasoning and answer from prior knowledge. On a metric that rewards consistent reasoning, all models perform poorly and exhibit high variance across question permutations, indicating that Large Language Models' (LLMs) reasoning faculty remains brittle. Overall, results on the benchmark reflect the recent progress of Inference-Time Compute (ITC) models but suggest ample room for further improvement. The benchmark is a step towards better measurement of reasoning abilities of LLMs and offers a cautionary tale on the importance of disentangling reasoning abilities from models' internalised knowledge when developing reasoning benchmarks.

Andrew M. Bean, Simi Hellsten, Harry Mayne et al.

In this paper, we present the LingOly benchmark, a novel benchmark for advanced reasoning abilities in large language models. Using challenging Linguistic Olympiad puzzles, we evaluate (i) capabilities for in-context identification and generalisation of linguistic patterns in very low-resource or extinct languages, and (ii) abilities to follow complex task instructions. The LingOly benchmark covers more than 90 mostly low-resource languages, minimising issues of data contamination, and contains 1,133 problems across 6 formats and 5 levels of human difficulty. We assess performance with both direct accuracy and comparison to a no-context baseline to penalise memorisation. Scores from 11 state-of-the-art LLMs demonstrate the benchmark to be challenging, and models perform poorly on the higher difficulty problems. On harder problems, even the top model only achieved 38.7% accuracy, a 24.7% improvement over the no-context baseline. Large closed models typically outperform open models, and in general, the higher resource the language, the better the scores. These results indicate, in absence of memorisation, true multi-step out-of-domain reasoning remains a challenge for current language models.

Harry Mayne, Guy Parsons, Adam Mahdi

The use of unsupervised learning to identify patient subgroups has emerged as a potentially promising direction to improve the efficiency of Intensive Care Units (ICUs). By identifying subgroups of patients with similar levels of medical resource need, ICUs could be restructured into a collection of smaller subunits, each catering to a specific group. However, it is unclear whether common patient subgroups exist across different ICUs, which would determine whether ICU restructuring could be operationalised in a standardised manner. In this paper, we tested the hypothesis that common ICU patient subgroups exist by examining whether the results from one existing study generalise to a different dataset. We extracted 16 features representing medical resource need and used consensus clustering to derive patient subgroups, replicating the previous study. We found limited similarities between our results and those of the previous study, providing evidence against the hypothesis. Our findings imply that there is significant variation between ICUs; thus, a standardised restructuring approach is unlikely to be appropriate. Instead, potential efficiency gains might be greater when the number and nature of the subunits are tailored to each ICU individually.