-
Soft Contamination Means Benchmarks Test Shallow Generalization
Authors:
Ari Spiesberger,
Juan J. Vazquez,
Nicky Pochinkov,
Tomáš Gavenčiak,
Peli Grietzer,
Gavin Leech,
Nandi Schoots
Abstract:
If LLM training data is polluted with benchmark test data, then benchmark performance gives biased estimates of out-of-distribution (OOD) generalization. Typical decontamination filters use n-gram matching which fail to detect semantic duplicates: sentences with equivalent (or near-equivalent) content that are not close in string space. We study this soft contamination of training data by semantic…
▽ More
If LLM training data is polluted with benchmark test data, then benchmark performance gives biased estimates of out-of-distribution (OOD) generalization. Typical decontamination filters use n-gram matching which fail to detect semantic duplicates: sentences with equivalent (or near-equivalent) content that are not close in string space. We study this soft contamination of training data by semantic duplicates. Among other experiments, we embed the Olmo3 training corpus and find that: 1) contamination remains widespread, e.g. we find semantic duplicates for 78% of CodeForces and exact duplicates for 50% of ZebraLogic problems; 2) including semantic duplicates of benchmark data in training does improve benchmark performance; and 3) when finetuning on duplicates of benchmark datapoints, performance also improves on truly-held-out datapoints from the same benchmark. We argue that recent benchmark gains are thus confounded: the prevalence of soft contamination means gains reflect both genuine capability improvements and the accumulation of test data and effective test data in growing training corpora.
△ Less
Submitted 12 February, 2026;
originally announced February 2026.
-
Beyond Tokens in Language Models: Interpreting Activations through Text Genre Chunks
Authors:
Éloïse Benito-Rodriguez,
Einar Urdshals,
Jasmina Nasufi,
Nicky Pochinkov
Abstract:
Understanding Large Language Models (LLMs) is key to ensure their safe and beneficial deployment. This task is complicated by the difficulty of interpretability of LLM structures, and the inability to have all their outputs human-evaluated. In this paper, we present the first step towards a predictive framework, where the genre of a text used to prompt an LLM, is predicted based on its activations…
▽ More
Understanding Large Language Models (LLMs) is key to ensure their safe and beneficial deployment. This task is complicated by the difficulty of interpretability of LLM structures, and the inability to have all their outputs human-evaluated. In this paper, we present the first step towards a predictive framework, where the genre of a text used to prompt an LLM, is predicted based on its activations. Using Mistral-7B and two datasets, we show that genre can be extracted with F1-scores of up to 98% and 71% using scikit-learn classifiers. Across both datasets, results consistently outperform the control task, providing a proof of concept that text genres can be inferred from LLMs with shallow learning models.
△ Less
Submitted 20 November, 2025;
originally announced November 2025.
-
ParaScopes: What do Language Models Activations Encode About Future Text?
Authors:
Nicky Pochinkov,
Yulia Volkova,
Anna Vasileva,
Sai V R Chereddy
Abstract:
Interpretability studies in language models often investigate forward-looking representations of activations. However, as language models become capable of doing ever longer time horizon tasks, methods for understanding activations often remain limited to testing specific concepts or tokens. We develop a framework of Residual Stream Decoders as a method of probing model activations for paragraph-s…
▽ More
Interpretability studies in language models often investigate forward-looking representations of activations. However, as language models become capable of doing ever longer time horizon tasks, methods for understanding activations often remain limited to testing specific concepts or tokens. We develop a framework of Residual Stream Decoders as a method of probing model activations for paragraph-scale and document-scale plans. We test several methods and find information can be decoded equivalent to 5+ tokens of future context in small models. These results lay the groundwork for better monitoring of language models and better understanding how they might encode longer-term planning information.
△ Less
Submitted 31 October, 2025;
originally announced November 2025.
-
Extracting Paragraphs from LLM Token Activations
Authors:
Nicholas Pochinkov,
Angelo Benoit,
Lovkush Agarwal,
Zainab Ali Majid,
Lucile Ter-Minassian
Abstract:
Generative large language models (LLMs) excel in natural language processing tasks, yet their inner workings remain underexplored beyond token-level predictions. This study investigates the degree to which these models decide the content of a paragraph at its onset, shedding light on their contextual understanding. By examining the information encoded in single-token activations, specifically the…
▽ More
Generative large language models (LLMs) excel in natural language processing tasks, yet their inner workings remain underexplored beyond token-level predictions. This study investigates the degree to which these models decide the content of a paragraph at its onset, shedding light on their contextual understanding. By examining the information encoded in single-token activations, specifically the "\textbackslash n\textbackslash n" double newline token, we demonstrate that patching these activations can transfer significant information about the context of the following paragraph, providing further insights into the model's capacity to plan ahead.
△ Less
Submitted 10 September, 2024;
originally announced September 2024.
-
Modularity in Transformers: Investigating Neuron Separability & Specialization
Authors:
Nicholas Pochinkov,
Thomas Jones,
Mohammed Rashidur Rahman
Abstract:
Transformer models are increasingly prevalent in various applications, yet our understanding of their internal workings remains limited. This paper investigates the modularity and task specialization of neurons within transformer architectures, focusing on both vision (ViT) and language (Mistral 7B) models. Using a combination of selective pruning and MoEfication clustering techniques, we analyze…
▽ More
Transformer models are increasingly prevalent in various applications, yet our understanding of their internal workings remains limited. This paper investigates the modularity and task specialization of neurons within transformer architectures, focusing on both vision (ViT) and language (Mistral 7B) models. Using a combination of selective pruning and MoEfication clustering techniques, we analyze the overlap and specialization of neurons across different tasks and data subsets. Our findings reveal evidence of task-specific neuron clusters, with varying degrees of overlap between related tasks. We observe that neuron importance patterns persist to some extent even in randomly initialized models, suggesting an inherent structure that training refines. Additionally, we find that neuron clusters identified through MoEfication correspond more strongly to task-specific neurons in earlier and later layers of the models. This work contributes to a more nuanced understanding of transformer internals and offers insights into potential avenues for improving model interpretability and efficiency.
△ Less
Submitted 30 August, 2024;
originally announced August 2024.
-
Investigating Neuron Ablation in Attention Heads: The Case for Peak Activation Centering
Authors:
Nicholas Pochinkov,
Ben Pasero,
Skylar Shibayama
Abstract:
The use of transformer-based models is growing rapidly throughout society. With this growth, it is important to understand how they work, and in particular, how the attention mechanisms represent concepts. Though there are many interpretability methods, many look at models through their neuronal activations, which are poorly understood. We describe different lenses through which to view neuron act…
▽ More
The use of transformer-based models is growing rapidly throughout society. With this growth, it is important to understand how they work, and in particular, how the attention mechanisms represent concepts. Though there are many interpretability methods, many look at models through their neuronal activations, which are poorly understood. We describe different lenses through which to view neuron activations, and investigate the effectiveness in language models and vision transformers through various methods of neural ablation: zero ablation, mean ablation, activation resampling, and a novel approach we term 'peak ablation'. Through experimental analysis, we find that in different regimes and models, each method can offer the lowest degradation of model performance compared to other methods, with resampling usually causing the most significant performance deterioration. We make our code available at https://github.com/nickypro/investigating-ablation.
△ Less
Submitted 30 August, 2024;
originally announced August 2024.
-
Dissecting Language Models: Machine Unlearning via Selective Pruning
Authors:
Nicholas Pochinkov,
Nandi Schoots
Abstract:
Understanding and shaping the behaviour of Large Language Models (LLMs) is increasingly important as applications become more powerful and more frequently adopted. This paper introduces a machine unlearning method specifically designed for LLMs. We introduce a selective pruning method for LLMs that removes neurons based on their relative importance on a targeted capability compared to overall netw…
▽ More
Understanding and shaping the behaviour of Large Language Models (LLMs) is increasingly important as applications become more powerful and more frequently adopted. This paper introduces a machine unlearning method specifically designed for LLMs. We introduce a selective pruning method for LLMs that removes neurons based on their relative importance on a targeted capability compared to overall network performance. This approach is a compute- and data-efficient method for identifying and removing neurons that enable specific behaviours. Our findings reveal that both feed-forward and attention neurons in LLMs are specialized; that is, for specific tasks, certain neurons are more crucial than others. Code from all experiments is available at https://github.com/nickypro/selective-pruning
△ Less
Submitted 24 July, 2024; v1 submitted 2 March, 2024;
originally announced March 2024.