The presentation discusses interpreting latent features in large language models (LLMs). After an introduction on mechanistic interpretability fundamentals, including feature superposition and sparse autoencoders, I discuss recent work by the Anthropic interpretability team (Ameisen et al. 2025, Lindsey et al. 2025) for extracting circuits of interpretable features from trained LLMs. Real-world investigations of Claude mechanisms, such as multi-step reasoning and multilinguality, are also analyzed.
Word-level quality estimation (QE) detects erroneous spans in machine translations, which can direct and facilitate human post-editing. While the accuracy of word-level QE systems has been assessed extensively, their usability and downstream influence on the speed, quality and editing choices of human post-editing remain understudied. Our QE4PE study investigates the impact of word-level QE on machine translation (MT) post-editing in a realistic setting involving 42 professional post-editors across two translation directions. We compare four error-span highlight modalities, including supervised and uncertainty-based word-level QE methods, for identifying potential errors in the outputs of a state-of-the-art neural MT model. Post-editing effort and productivity are estimated by behavioral logs, while quality improvements are assessed by word- and segment-level human annotation. We find that domain, language and editors' speed are critical factors in determining highlights' effectiveness, with modest differences between human-made and automated QE highlights underlining a gap between accuracy and usability in professional workflows.
In this presentation, I will provide an overview of the interpretability research landscape and describe various promising methods for exploring and controlling the inner mechanisms of generative language models. I will focus specifically on post-hoc attribution technique and their usage to identify relevant input and model components, showcasing their usage with our Inseq open-source toolkit. A practical application of attribution techniques will be presented with the PECoRe data-driven framework for context usage attribution and its adaptation to produce internals-based citations for model answers in retrieval-augmented generation settings (MIRAGE).
This presentation focuses on applying post-hoc interpretability techniques to analyze how language models (LMs) use input information throughout the generation process. We briefly introduce Inseq, our open-source toolkit designed to simplify advanced feature attribution analyses for LMs. Then, our Plausibility Evaluation of Context Reliance (PECoRe) interpretability framework is introduced to conduct data-driven analyses of context usage in LMs. In conclusion, we showcase how PECoRe can easily be adapted to retrieval-augmented generation (RAG) settings to produce internals-based citations for model answers. Our proposed Model Internals for RAG Explanations (MIRAGE) method achieves citation quality comparable to supervised answer validators with no additional training, producing citations that are faithful to actual context usage during generation.
Questo intervento sarà mirato a demistificare il funzionamento dei modelli del linguaggio (Large Language Models), ed evidenziare come lo studio di questi sistemi come 'artefatti cognitivi' possa contribuire a una migliore comprensione dei meccanismi di ragionamento (umani e non), e dei bias nella società che ci circonda.
In this presentation, I will provide an overview of the interpretability research landscape and describe various promising methods for exploring and controlling the inner mechanisms of generative language models. I will focus specifically on post-hoc attribution technique and their usage to identify relevant input and model components, showcasing their usage with our Inseq open-source toolkit. A practical application of attribution techniques will be presented with the PECoRe data-driven framework for context usage attribution and its adaptation to produce internals-based citations for model answers in retrieval-augmented generation settings (MIRAGE).