A central question in multilingual language modeling is whether large language models (LLMs) develop a universal concept representation, disentangled from specific languages.In this paper, we address this question by analyzing latent representations (latents) during a word-translation task in transformer-based LLMs.We strategically extract latents from a source translation prompt and insert them into the forward pass on a target translation prompt.By doing so, we find that the output language is encoded in the latent at an earlier layer than the concept to be translated.Building on this insight, we conduct two key experiments.First, we demonstrate that we can change the concept without changing the language and vice versa through activation patching alone.Second, we show that patching with the mean representation of a concept across different languages does not affect the models' ability to translate it, but instead improves it.Finally, we generalize to multi-token generation and demonstrate that the model can generate natural language description of those mean representations.Our results provide evidence for the existence of language-agnostic concept representations within the investigated models.

When making predictions, a language model must trade off how much it relies on its context vs. its prior knowledge. Choosing how sensitive the model is to its context is a fundamental functionality, as it enables the model to excel at tasks like retrieval-augmented generation and question-answering. In this paper, we search for a knob which controls this sensitivity, determining whether language models answer from the context or their prior knowledge. To guide this search, we design a task for controllable context sensitivity. In this task, we first feed the model a context ("Paris is in England") and a question ("Where is Paris?"); we then instruct the model to either use its prior or contextual knowledge and evaluate whether it generates the correct answer for both intents (either "France" or "England"). When fine-tuned on this task, instruct versions of Llama-3.1, Mistral-v0.3, and Gemma-2 can solve it with high accuracy (85-95%). Analyzing these high-performing models, we narrow down which layers may be important to context sensitivity using a novel linear time algorithm. Then, in each model, we identify a 1-D subspace in a single layer that encodes whether the model follows context or prior knowledge. Interestingly, while we identify this subspace in a fine-tuned model, we find that the exact same subspace serves as an effective knob in not only that model but also non-fine-tuned instruct and base models of that model family. Finally, we show a strong correlation between a model's performance and how distinctly it separates context-agreeing from context-ignoring answers in this subspace. These results suggest a single fundamental subspace facilitates how the model chooses between context and prior knowledge.

Jannik Brinkmann, Chris Wendler, Christian Bartelt, Aaron Mueller. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Tokenization efficiency plays a critical role in the performance and cost of large language models (LLMs), yet most models rely on static tokenizers optimized on general-purpose corpora. These tokenizers’ fixed vocabularies often fail to adapt to domain- or language-specific inputs, leading to longer token sequences and higher computational costs. We introduce zip2zip, a novel method for achieving context-adaptive tokenization in LLMs at inference time. Leveraging an online data compression algorithm (Lempel–Ziv–Welch), zip2zip dynamically expands its active vocabulary at inference time by continuously replacing fragmented token sequences with more compact hypertokens, which it can immediately output during generation. In doing so, the model refines its internal tokenization scheme to match the token distribution of the current context, reducing redundancy and improving representational efficiency. zip2zip consists of three key components: (1) a tokenizer based on Lempel–Ziv–Welch compression that incrementally merges co-occurring tokens into reusable hypertokens on the fly; (2) a dynamic embedding (and unembedding) layer that computes embeddings for newly formed hypertokens at runtime; and (3) a variant of autoregressive language modeling that pretrains the model to handle hypertokenized, compressed text sequences as inputs and outputs. We show that an existing LLM can be uptrained for zip2zip in 10 GPU-hours via parameter-efficient finetuning. The resulting LLM performs test-time adaptation, learning to use hypertokens in unseen contexts and reducing input and output tokens by 15–40%. Code and models are released at https://github.com/epfl-dlab/zip2zip.

For large language models (LLMs), sparse autoencoders (SAEs) have been shown to decompose intermediate representations that often are not interpretable directly into sparse sums of interpretable features, facilitating better control and subsequent analysis. However, similar analyses and approaches have been lacking for text-to-image models. We investigate the possibility of using SAEs to learn interpretable features for SDXL Turbo, a few-step text-to-image diffusion model. To this end, we train SAEs on the updates performed by transformer blocks within SDXL Turbo's denoising U-net in its 1-step setting. Interestingly, we find that they generalize to 4-step SDXL Turbo and even to the multi-step SDXL base model (i.e., a different model) without additional training. In addition, we show that their learned features are interpretable, causally influence the generation process, and reveal specialization among the blocks. We do so by creating RIEBench, a representation-based image editing benchmark, for editing images while they are generated by turning on and off individual SAE features. This allows us to track which transformer blocks' features are the most impactful depending on the edit category. Our work is the first investigation of SAEs for interpretability in text-to-image diffusion models and our results establish SAEs as a promising approach for understanding and manipulating the internal mechanisms of text-to-image models.

We ask whether multilingual language models trained on unbalanced, English-dominated corpora use English as an internal pivot languagea question of key importance for understanding how language models function and the origins of linguistic bias.Focusing on the Llama-2 family of transformer models, our study uses carefully constructed non-English prompts with a unique correct single-token continuation.From layer to layer, transformers gradually map an input embedding of the final prompt token to an output embedding from which next-token probabilities are computed.Tracking intermediate embeddings through their high-dimensional space reveals three distinct phases, whereby intermediate embeddings (1) start far away from output token embeddings; (2) already allow for decoding a semantically correct next token in middle layers, but give higher probability to its version in English than in the input language;(3) finally move into an input-language-specific region of the embedding space.We cast these results into a conceptual model where the three phases operate in "input space", "concept space", and "output space", respectively.Crucially, our evidence suggests that the abstract "concept space" lies closer to English than to other languages, which may have important consequences regarding the biases held by multilingual language models.Code and data is made available here: https://github.com/ epfl-dlab/llm-latent-language.0 5 10 15 20 25 30 layer 0.0 0.5 1.0 probability en zh (a) Translation task 0 5 10 15 20 25 30 35 40 layer 0.0 0.5 1.0 en zh 0 10 20 30 40 50 60 70 80 layer 0.0 0.5 1.0 en zh 0 5 10 entropy 0 5 10 15 20 25 30 layer 0.0 0.5 1.0 probability en zh (b) Repetition task 0 5 10 15 20 25 30 35 40 layer 0.0 0.5 1.0 en zh 0 10 20 30 40 50 60 70 80 layer 0.0 0.5 1.0 en zh 0 5 10 entropy 0 5 10 15 20 25 30 layer 0.0 0.5 1.0 probability en zh 7B (c) Cloze task 0 5 10 15 20 25 30 35 40 layer 0.0 0.5 1.0 en zh 13B 0 10 20 30 40 50 60 70 80 layer 0.0 0.5

We present a novel perspective and algorithm for learning directed acyclic graphs (DAGs) from data generated by a linear structural equation model (SEM). First, we show that a linear SEM can be viewed as a linear transform that, in prior work, computes the data from a dense input vector of random valued root causes (as we will call them) associated with the nodes. Instead, we consider the case of (approximately) few root causes and also introduce noise in the measurement of the data. Intuitively, this means that the DAG data is produced by few data generating events whose effect percolates through the DAG. We prove identifiability in this new setting and show that the true DAG is the global minimizer of the $L^0$-norm of the vector of root causes. For data satisfying the few root causes assumption, we show superior performance compared to prior DAG learning methods.

Recent advances in Fourier analysis have brought new tools to efficiently represent and learn set functions. In this paper, we bring the power of Fourier analysis to the design of combinatorial auctions (CAs). The key idea is to approximate bidders' value functions using Fourier-sparse set functions, which can be computed using a relatively small number of queries. Since this number is still too large for practical CAs, we propose a new hybrid design: we first use neural networks (NNs) to learn bidders’ values and then apply Fourier analysis to the learned representations. On a technical level, we formulate a Fourier transform-based winner determination problem and derive its mixed integer program formulation. Based on this, we devise an iterative CA that asks Fourier-based queries. We experimentally show that our hybrid ICA achieves higher efficiency than prior auction designs, leads to a fairer distribution of social welfare, and significantly reduces runtime. With this paper, we are the first to leverage Fourier analysis in CA design and lay the foundation for future work in this area. Our code is available on GitHub: https://github.com/marketdesignresearch/FA-based-ICAs.

Many applications of machine learning on discrete domains, such as learning preference functions in recommender systems or auctions, can be reduced to estimating a set function that is sparse in the Fourier domain. In this work, we present a new family of algorithms for learning Fourier-sparse set functions. They require at most nk − k log k + k queries (set function evaluations), under mild conditions on the Fourier coefficients, where n is the size of the ground set and k the number of non-zero Fourier coefficients. In contrast to other work that focused on the orthogonal Walsh-Hadamard transform (WHT), our novel algorithms operate with recently introduced non-orthogonal Fourier transforms that offer different notions of Fourier-sparsity. These naturally arise when modeling, e.g., sets of items forming substitutes and complements. We demonstrate effectiveness on several real-world applications.

Recent work introduced a framework for signal processing (SP) on meet/join lattices. Such a lattice is partially ordered and supports a meet (or join) operation that returns the greatest lower bound and the smallest upper bound of two elements, respectively. Lattices appear in various domains and can be used, for example, to express rankings in social choice theory or multisets in combinatorial auctions. Discrete lattice SP (DLSP) uses the meet operation as shift and derives associated notions of convolution and Fourier transform for signals indexed by lattices. In this paper we extend DLSP with Wiener filtering for denoising and demonstrate it on two prototypical applications.

Graph signal processing on directed graphs poses theoretical challenges since an eigendecomposition of filters is in general not available. Instead, Fourier analysis requires a Jordan decomposition and the frequency response is given by the Jordan normal form, whose computation is numerically unstable for large sizes. In this paper, we propose to replace a given adjacency shift A by a diagonalizable shift ADobtained via the Jordan-Chevalley decomposition. This means, as we show, that ADgenerates the subalgebra of all diagonalizable filters and is itself a polynomial in A (i.e., a filter). For several synthetic and real-world graphs, we show how ADadds and removes edges compared to A.

We present a novel class of convolutional neural networks (CNNs) for set functions, i.e., data indexed with the powerset of a finite set. The convolutions are derived as linear, shift-equivariant functions for various notions of shifts on set functions. The framework is fundamentally different from graph convolutions based on the Laplacian, as it provides not one but several basic shifts, one for each element in the ground set. Prototypical experiments with several set function classification tasks on synthetic datasets and on datasets derived from real-world hypergraphs demonstrate the potential of our new powerset CNNs.