arXiv subjects

Pontus Stenetorp

Publications and source records attributed to Pontus Stenetorp.

At least 19 recordsLinked to original sources

arXiv2026

AIMO Interpretability Challenge

We propose the AIMO Interpretability Challenge, a competition on distinguishing robust from spurious reasoning in frontier mathematical language models based on the models' internal mechanisms. The challenge is motivated by a central limitation of standard reasoning benchmarks: strong final-answer accuracy does not reveal whether a model relies on stable reasoning mechanisms or exploits brittle reasoning shortcuts. Building on AI Mathematical Olympiad (AIMO) problems and submissions, together with resources from the Fields Model Initiative, the competition will provide (1) newly-published olympiad-level math reasoning problems and their symbolic representations, allowing generation of novel functional variants, (2) access to frontier reasoning models, and (3) assessments of models' adversarial robustness on these problems. Participants will use these resources, along with our computing infrastructure support, to develop methods for identifying which models solve problems robustly. Our competition will also create a new, open robustness benchmark and baseline systems, aiming to provide a lasting foundation for standard benchmarking in mathematical reasoning and interpretability. Scientifically, the competition connects interpretability and generalization research around a central question in AI research: can we determine if, and to what extent, the decision-making of frontier AI models is generalizable and thus, reliable?

cs.AI↗

arXiv2026

PolyFact: Comparing Consistency-Driven Post-training Methods for Cross-Lingual Factual Recall

Large language models (LLMs) trained predominantly on English data encode substantial world knowledge, yet often fail to express it reliably in other languages, a phenomenon known as cross-lingual factual inconsistency. To study this, we introduce PolyFact, a fully parallel multilingual factual QA dataset of 60K Wikidata-grounded facts across 12 typologically diverse languages, and propose consistency-driven GRPO with cross-lingual reward pooling. We compare our method against supervised fine-tuning (SFT) and the consistency-enhancement baselines DCO and CM-Align on OLMo-2-1124-7B and Qwen-2.5-7B, and analyze whether light continual pretraining (CPT) on parallel data provides a useful foundation for post-training. No single method dominates: SFT maximises in-distribution accuracy but not consistency, DCO yields the strongest consistency gains but fails to transfer to free-form generation, and our GRPO variant achieves the strongest transfer to free-form recall and unseen languages on the multilingual base model. CPT mildly aids monolingual models but harms multilingual ones. Mechanistic analyses suggest that GRPO is associated with reduced language specialization, consistent with greater sharing of representations across languages. We release our code, models, and dataset publicly.

cs.CL↗

arXiv2026

AIRA_2: Overcoming Bottlenecks in AI Research Agents

Existing research has identified three structural performance bottlenecks in AI research agents: (1) synchronous single-GPU execution constrains sample throughput, limiting the benefit of search; (2) a generalization gap where validation-based selection causes overfitting and performance to degrade over extended search horizons; and (3) the limited capability of fixed, single-turn LLM operators imposes a ceiling on search performance. We introduce AIRA$_2$, which addresses these bottlenecks through three architectural choices: an asynchronous multi-GPU worker pool that increases experiment throughput linearly; a Hidden Consistent Evaluation protocol that delivers a reliable evaluation signal; and ReAct agents that dynamically scope their actions and debug interactively. On MLE-bench-30, AIRA$^{\dagger}_{2}$ achieves a mean Percentile Rank of 81.5% at 24 hours and 83.1% at 72 hours, outperforming the strongest baseline, which achieves 72.7%. On AIRS-Bench, AIRA$_2$ exceeds human state-of-the-art on 6 out of 20 diverse research tasks. Ablations confirm that each architectural component is necessary, that performance follows a predictable scaling law that transfers across LLM backbones, and that the "overfitting" reported in prior work was driven by evaluation noise rather than true data memorization.

cs.AI↗

arXiv2026

ShapleyLaw: A Game-Theoretic Approach to Multilingual Scaling Laws

In multilingual pretraining, the test loss of a pretrained model is heavily influenced by the proportion of each language in the pretraining data, namely the \textit{language mixture ratios}. Multilingual scaling laws can predict the test loss under different language mixture ratios and can therefore be used to estimate the optimal ratios. However, the current approaches to multilingual scaling laws do not measure the \textit{cross-lingual transfer} effect, resulting in suboptimal mixture ratios. In this paper, we consider multilingual pretraining as a cooperative game in which each language acts as a player that jointly contributes to pretraining, gaining the resulting reduction in test loss as the payoff. Consequently, from the perspective of cooperative game theory, we quantify the cross-lingual transfer from each language by its contribution in the game, and propose a game-theoretic multilingual scaling law called \textit{ShapleyLaw}. Our experiments show that ShapleyLaw outperforms baseline methods in model performance prediction and language mixture optimization.

cs.CL↗

arXiv2026

The Role of Mixed-Language Documents for Multilingual Large Language Model Pretraining

Multilingual large language models achieve impressive cross-lingual performance despite largely monolingual pretraining. While bilingual data in pretraining corpora is widely believed to enable these abilities, details of its contributions remain unclear. We investigate this question by pretraining models from scratch under controlled conditions, comparing the standard web corpus with a monolingual-only version that removes all multilingual documents. Despite constituting only 2% of the corpus, removing bilingual data causes translation performance to drop 56% in BLEU, while behaviour on cross-lingual QA and general reasoning tasks remains stable, with training curves largely overlapping the baseline. To understand this asymmetry, we categorize bilingual data into parallel (14%), code-switching (72%), and miscellaneous documents (14%) based on the semantic relevance of content in different languages. We then conduct granular ablations by reintroducing parallel or code-switching data into the monolingual-only corpus. Our experiments reveal that parallel data almost fully restores translation performance (91% of the unfiltered baseline), whereas code-switching contributes minimally. Other cross-lingual tasks remain largely unaffected by either type. These findings reveal that translation critically depends on systematic token-level alignments from parallel data, whereas cross-lingual understanding and reasoning appear to be achievable even without bilingual data.

cs.CL↗

arXiv2025

Language Models Learn Universal Representations of Numbers and Here's Why You Should Care

Prior work has shown that large language models (LLMs) often converge to accurate input embedding for numbers, based on sinusoidal representations. In this work, we quantify that these representations are in fact strikingly systematic, to the point of being almost perfectly universal: different LLM families develop equivalent sinusoidal structures, and number representations are broadly interchangeable in a large swathe of experimental setups. We show that properly factoring in this characteristic is crucial when it comes to assessing how accurately LLMs encode numeric and other ordinal information, and that mechanistically enhancing this sinusoidality can also lead to reductions of LLMs' arithmetic errors.

cs.CL↗

arXiv2025

Drawing Conclusions from Draws: Rethinking Preference Semantics in Arena-Style LLM Evaluation

In arena-style evaluation of large language models (LLMs), two LLMs respond to a user query, and the user chooses the winning response or deems the "battle" a draw, resulting in an adjustment to the ratings of both models. The prevailing approach for modeling these rating dynamics is to view battles as two-player game matches, as in chess, and apply the Elo rating system and its derivatives. In this paper, we critically examine this paradigm. Specifically, we question whether a draw genuinely means that the two models are equal and hence whether their ratings should be equalized. Instead, we conjecture that draws are more indicative of query difficulty: if the query is too easy, then both models are more likely to succeed equally. On three real-world arena datasets, we show that ignoring rating updates for draws yields a 1-3% relative increase in battle outcome prediction accuracy (which includes draws) for all four rating systems studied. Further analyses suggest that draws occur more for queries rated as very easy and those as highly objective, with risk ratios of 1.37 and 1.35, respectively. We recommend future rating systems to reconsider existing draw semantics and to account for query properties in rating updates.

cs.CL↗

arXiv2025

Translate, then Detect: Leveraging Machine Translation for Cross-Lingual Toxicity Classification

Multilingual toxicity detection remains a significant challenge due to the scarcity of training data and resources for many languages. While prior work has leveraged the translate-test paradigm to support cross-lingual transfer across a range of classification tasks, the utility of translation in supporting toxicity detection at scale remains unclear. In this work, we conduct a comprehensive comparison of translation-based and language-specific/multilingual classification pipelines. We find that translation-based pipelines consistently outperform out-of-distribution classifiers in 81.3% of cases (13 of 16 languages), with translation benefits strongly correlated with both the resource level of the target language and the quality of the machine translation (MT) system. Our analysis reveals that traditional classifiers outperform large language model (LLM) judges, with this advantage being particularly pronounced for low-resource languages, where translate-classify methods dominate translate-judge approaches in 6 out of 7 cases. We additionally show that MT-specific fine-tuning on LLMs yields lower refusal rates compared to standard instruction-tuned models, but it can negatively impact toxicity detection accuracy for low-resource languages. These findings offer actionable guidance for practitioners developing scalable multilingual content moderation systems.

cs.CL↗

arXiv2025

AI Research Agents for Machine Learning: Search, Exploration, and Generalization in MLE-bench

AI research agents are demonstrating great potential to accelerate scientific progress by automating the design, implementation, and training of machine learning models. We focus on methods for improving agents' performance on MLE-bench, a challenging benchmark where agents compete in Kaggle competitions to solve real-world machine learning problems. We formalize AI research agents as search policies that navigate a space of candidate solutions, iteratively modifying them using operators. By designing and systematically varying different operator sets and search policies (Greedy, MCTS, Evolutionary), we show that their interplay is critical for achieving high performance. Our best pairing of search strategy and operator set achieves a state-of-the-art result on MLE-bench lite, increasing the success rate of achieving a Kaggle medal from 39.6% to 47.7%. Our investigation underscores the importance of jointly considering the search strategy, operator design, and evaluation methodology in advancing automated machine learning.

cs.AI↗

arXiv2025

SSA-COMET: Do LLMs Outperform Learned Metrics in Evaluating MT for Under-Resourced African Languages?

Evaluating machine translation (MT) quality for under-resourced African languages remains a significant challenge, as existing metrics often suffer from limited language coverage and poor performance in low-resource settings. While recent efforts, such as AfriCOMET, have addressed some of the issues, they are still constrained by small evaluation sets, a lack of publicly available training data tailored to African languages, and inconsistent performance in extremely low-resource scenarios. In this work, we introduce SSA-MTE, a large-scale human-annotated MT evaluation (MTE) dataset covering 14 African language pairs from the News domain, with over 73,000 sentence-level annotations from a diverse set of MT systems. Based on this data, we develop SSA-COMET and SSA-COMET-QE, improved reference-based and reference-free evaluation metrics. We also benchmark prompting-based approaches using state-of-the-art LLMs like GPT-4o, Claude-3.7 and Gemini 2.5 Pro. Our experimental results show that SSA-COMET models significantly outperform AfriCOMET and are competitive with the strongest LLM Gemini 2.5 Pro evaluated in our study, particularly on low-resource languages such as Twi, Luo, and Yoruba. All resources are released under open licenses to support future research.

cs.CL↗

arXiv2025

Multilingual Language Model Pretraining using Machine-translated Data

High-resource languages such as English, enables the pretraining of high-quality large language models (LLMs). The same can not be said for most other languages as LLMs still underperform for non-English languages, likely due to a gap in the quality and diversity of the available multilingual pretraining corpora. In this work, we find that machine-translated texts from a single high-quality source language can contribute significantly to the pretraining quality of multilingual LLMs. We translate FineWeb-Edu, a high-quality English web dataset, into nine languages, resulting in a 1.7-trillion-token dataset, which we call TransWebEdu and pretrain a 1.3B-parameter model, TransWebLLM, from scratch on this dataset. Across nine non-English reasoning tasks, we show that TransWebLLM matches or outperforms state-of-the-art multilingual models trained using closed data, such as Llama3.2, Qwen2.5, and Gemma, despite using an order of magnitude less data. We demonstrate that adding less than 5% of TransWebEdu as domain-specific pretraining data sets a new state-of-the-art in Arabic, Italian, Indonesian, Swahili, and Welsh understanding and commonsense reasoning tasks. To promote reproducibility, we release our corpus, models, and training pipeline under Open Source Initiative-approved licenses.

cs.CL↗

arXiv2025

Warmup Generations: A Task-Agnostic Approach for Guiding Sequence-to-Sequence Learning with Unsupervised Initial State Generation

Traditional supervised fine-tuning (SFT) strategies for sequence-to-sequence tasks often train models to directly generate the target output. Recent work has shown that guiding models with intermediate steps, such as keywords, outlines, or reasoning chains, can significantly improve performance, coherence, and interpretability. However, these methods often depend on predefined intermediate formats and annotated data, limiting their scalability and generalizability. In this work, we introduce a task-agnostic framework that enables models to generate intermediate "warmup" sequences. These warmup sequences, serving as an initial state for subsequent generation, are optimized to enhance the probability of generating the target sequence without relying on external supervision or human-designed structures. Drawing inspiration from reinforcement learning principles, our method iteratively refines these intermediate steps to maximize their contribution to the final output, similar to reward-driven optimization in reinforcement learning with human feedback. Experimental results across tasks such as translation, summarization, and multi-choice question answering for logical reasoning show that our approach outperforms traditional SFT methods, and offers a scalable and flexible solution for sequence-to-sequence tasks.

cs.CL↗

arXiv2024

Lost in Inference: Rediscovering the Role of Natural Language Inference for Large Language Models

In the recent past, a popular way of evaluating natural language understanding (NLU), was to consider a model's ability to perform natural language inference (NLI) tasks. In this paper, we investigate if NLI tasks, that are rarely used for LLM evaluation, can still be informative for evaluating LLMs. Focusing on five different NLI benchmarks across six models of different scales, we investigate if they are able to discriminate models of different size and quality and how their accuracies develop during training. Furthermore, we investigate the extent to which the softmax distributions of models align with human distributions in cases where statements are ambiguous or vague. Overall, our results paint a positive picture for the NLI tasks: we find that they are able to discriminate well between models at various stages of training, yet are not (all) saturated. Furthermore, we find that while the similarity of model distributions with human label distributions increases with scale, it is still much higher than the similarity between two populations of humans, making it a potentially interesting statistic to consider.

cs.CL↗

arXiv2024

Multilingual Pretraining Using a Large Corpus Machine-Translated from a Single Source Language

English, as a very high-resource language, enables the pretraining of high-quality large language models (LLMs). The same cannot be said for most other languages, as leading LLMs still underperform for non-English languages, likely due to a gap in the quality and diversity of the available multilingual pretraining corpora. In this work, we find that machine-translated text from a single high-quality source language can contribute significantly to the pretraining of multilingual LLMs. We translate FineWeb-Edu, a high-quality English web dataset, into French, German, and Spanish, resulting in a final 300B-token dataset, which we call TransWeb-Edu, and pretrain a 1.3B-parameter model, CuatroLLM, from scratch on this dataset. Across five non-English reasoning tasks, we show that CuatroLLM matches or outperforms state-of-the-art multilingual models trained using closed data, such as Llama3.2 and Gemma2, despite using an order of magnitude less data, such as about 6% of the tokens used for Llama3.2's training. We further demonstrate that with additional domain-specific pretraining, amounting to less than 1% of TransWeb-Edu, CuatroLLM surpasses the state of the art in multilingual reasoning. To promote reproducibility, we release our corpus, models, and training pipeline under open licenses at hf.co/britllm/CuatroLLM.

cs.CL↗

arXiv2024

Jet Expansions of Residual Computation

We introduce a framework for expanding residual computational graphs using jets, operators that generalize truncated Taylor series. Our method provides a systematic approach to disentangle contributions of different computational paths to model predictions. In contrast to existing techniques such as distillation, probing, or early decoding, our expansions rely solely on the model itself and requires no data, training, or sampling from the model. We demonstrate how our framework grounds and subsumes logit lens, reveals a (super-)exponential path structure in the recursive residual depth and opens up several applications. These include sketching a transformer large language model with $n$-gram statistics extracted from its computations, and indexing the models' levels of toxicity knowledge. Our approach enables data-free analysis of residual computation for model interpretability, development, and evaluation.

cs.LG↗

arXiv2024

Linguini: A benchmark for language-agnostic linguistic reasoning

We propose a new benchmark to measure a language model's linguistic reasoning skills without relying on pre-existing language-specific knowledge. The test covers 894 questions grouped in 160 problems across 75 (mostly) extremely low-resource languages, extracted from the International Linguistic Olympiad corpus. To attain high accuracy on this benchmark, models don't need previous knowledge of the tested language, as all the information needed to solve the linguistic puzzle is presented in the context. We find that, while all analyzed models rank below 25% accuracy, there is a significant gap between open and closed models, with the best-performing proprietary model at 24.05% and the best-performing open model at 8.84%.

cs.CL↗

arXiv2024

Machine Translation Hallucination Detection for Low and High Resource Languages using Large Language Models

Recent advancements in massively multilingual machine translation systems have significantly enhanced translation accuracy; however, even the best performing systems still generate hallucinations, severely impacting user trust. Detecting hallucinations in Machine Translation (MT) remains a critical challenge, particularly since existing methods excel with High-Resource Languages (HRLs) but exhibit substantial limitations when applied to Low-Resource Languages (LRLs). This paper evaluates sentence-level hallucination detection approaches using Large Language Models (LLMs) and semantic similarity within massively multilingual embeddings. Our study spans 16 language directions, covering HRLs, LRLs, with diverse scripts. We find that the choice of model is essential for performance. On average, for HRLs, Llama3-70B outperforms the previous state of the art by as much as 0.16 MCC (Matthews Correlation Coefficient). However, for LRLs we observe that Claude Sonnet outperforms other LLMs on average by 0.03 MCC. The key takeaway from our study is that LLMs can achieve performance comparable or even better than previously proposed models, despite not being explicitly trained for any machine translation task. However, their advantage is less significant for LRLs.

cs.CL↗

arXiv2024

How Data Inter-connectivity Shapes LLMs Unlearning: A Structural Unlearning Perspective

While unlearning knowledge from large language models (LLMs) is receiving increasing attention, one important aspect remains unexplored. Existing approaches and benchmarks assume data points to-be-forgotten are independent, ignoring their inter-connectivity - a fundamental characteristic of real-world data structures. In this paper, we propose PISTOL, a method for compiling structural datasets. PISTOL leverages the inherently structured nature of contractual relationships, offering several key benefits. First, it enables insights into the impact of structural data on unlearning effectiveness. Second, it provides precise and concise ground truths for clearer evaluation. Third, its attribute generation does not require input from pre-trained LLMs, mitigating confounding risks. Leveraging datasets synthesized using PISTOL, we demonstrate how data inter-connectivity impacts LLM unlearning. Specifically, (a) in both the pre-trained and fine-tuned models, unlearning difficulty increases as data inter-connectivity grows, (b) there is a positive correlation between the density of the knowledge graph and unlearning difficulty, and (c) when the to-be-forgotten data is skewed towards one domain, balancing retaining performance across all domains is challenging.

cs.LG↗