Chris Wendler

Wenhao Zhu, Sizhe Liu, Shujian Huang et al.

Decoding by contrasting layers (DoLa), is designed to improve the generation quality of large language models (LLMs) by contrasting the prediction probabilities between an early exit output (amateur logits) and the final output (expert logits). However, we find that this approach does not work well on non-English tasks. Inspired by previous interpretability work on language transition during the model's forward pass, we discover that this issue arises from a language mismatch between early exit output and final output. In this work, we propose an improved contrastive decoding algorithm that is effective for diverse languages beyond English. To obtain more helpful amateur logits, we devise two strategies to skip a set of bottom, language-agnostic layers based on our preliminary analysis. Experimental results on multilingual reasoning benchmarks demonstrate that our proposed method outperforms previous contrastive decoding baselines and substantially improves LLM's chain-of-thought reasoning accuracy across 11 languages. The project will be available at: https://github.com/NJUNLP/SkipLayerCD.

Mathieu Chevalley, Jacob Sackett-Sanders, Yusuf Roohani et al.

In drug discovery, mapping interactions between genes within cellular systems is a crucial early step. Such maps are not only foundational for understanding the molecular mechanisms underlying disease biology but also pivotal for formulating hypotheses about potential targets for new medicines. Recognizing the need to elevate the construction of these gene-gene interaction networks, especially from large-scale, real-world datasets of perturbed single cells, the CausalBench Challenge was initiated. This challenge aimed to inspire the machine learning community to enhance state-of-the-art methods, emphasizing better utilization of expansive genetic perturbation data. Using the framework provided by the CausalBench benchmark, participants were tasked with refining the current methodologies or proposing new ones. This report provides an analysis and summary of the methods submitted during the challenge to give a partial image of the state of the art at the time of the challenge. Notably, the winning solutions significantly improved performance compared to previous baselines, establishing a new state of the art for this critical task in biology and medicine.

Natalie Shapira, Chris Wendler, Avery Yen et al.

We report an exploratory red-teaming study of autonomous language-model-powered agents deployed in a live laboratory environment with persistent memory, email accounts, Discord access, file systems, and shell execution. Over a two-week period, twenty AI researchers interacted with the agents under benign and adversarial conditions. Focusing on failures emerging from the integration of language models with autonomy, tool use, and multi-party communication, we document eleven representative case studies. Observed behaviors include unauthorized compliance with non-owners, disclosure of sensitive information, execution of destructive system-level actions, denial-of-service conditions, uncontrolled resource consumption, identity spoofing vulnerabilities, cross-agent propagation of unsafe practices, and partial system takeover. In several cases, agents reported task completion while the underlying system state contradicted those reports. We also report on some of the failed attempts. Our findings establish the existence of security-, privacy-, and governance-relevant vulnerabilities in realistic deployment settings. These behaviors raise unresolved questions regarding accountability, delegated authority, and responsibility for downstream harms, and warrant urgent attention from legal scholars, policymakers, and researchers across disciplines. This report serves as an initial empirical contribution to that broader conversation.

Bastian Seifert, Chris Wendler, Markus Püschel

We present a novel form of Fourier analysis, and associated signal processing concepts, for signals (or data) indexed by edge-weighted directed acyclic graphs (DAGs). This means that our Fourier basis yields an eigendecomposition of a suitable notion of shift and convolution operators that we define. DAGs are the common model to capture causal relationships between data values and in this case our proposed Fourier analysis relates data with its causes under a linearity assumption that we define. The definition of the Fourier transform requires the transitive closure of the weighted DAG for which several forms are possible depending on the interpretation of the edge weights. Examples include level of influence, distance, or pollution distribution. Our framework is different from prior GSP: it is specific to DAGs and leverages, and extends, the classical theory of Moebius inversion from combinatorics. For a prototypical application we consider DAGs modeling dynamic networks in which edges change over time. Specifically, we model the spread of an infection on such a DAG obtained from real-world contact tracing data and learn the infection signal from samples assuming sparsity in the Fourier domain.

Can Rager, Chris Wendler, Rohit Gandikota et al.

Refusal discovery is the task of identifying the full set of topics that a language model refuses to discuss. We introduce this new problem setting and develop a refusal discovery method, Iterated Prefill Crawler (IPC), that uses token prefilling to find forbidden topics. We benchmark IPC on Tulu-3-8B, an open-source model with public safety tuning data. Our crawler manages to retrieve 31 out of 36 topics within a budget of 1000 prompts. Next, we scale the crawler to a frontier model using the prefilling option of Claude-Haiku. Finally, we crawl three widely used open-weight models: Llama-3.3-70B and two of its variants finetuned for reasoning: DeepSeek-R1-70B and Perplexity-R1-1776-70B. DeepSeek-R1-70B reveals patterns consistent with censorship tuning: The model exhibits "thought suppression" behavior that indicates memorization of CCP-aligned responses. Although Perplexity-R1-1776-70B is robust to censorship, IPC elicits CCP-aligned refusals answers in the quantized model. Our findings highlight the critical need for refusal discovery methods to detect biases, boundaries, and alignment failures of AI systems.

Kevin Lu, Jannik Brinkmann, Stefan Huber et al.

How do protein structure prediction models fold proteins? We investigate this question by tracing how ESMFold folds a beta hairpin, a prevalent structural motif. Through counterfactual interventions on model latents, we identify two computational stages in the folding trunk. In the first stage, early blocks initialize pairwise biochemical signals: residue identities and associated biochemical features such as charge flow from sequence representations into pairwise representations. In the second stage, late blocks develop pairwise spatial features: distance and contact information accumulate in the pairwise representation. We demonstrate that the mechanisms underlying structural decisions of ESMFold can be localized, traced through interpretable representations, and manipulated with strong causal effects.

Kerem Sahin, Sheridan Feucht, Adam Belfki et al.

Induction heads are attention heads that perform inductive copying by matching patterns from earlier context and copying their continuations verbatim. As models develop induction heads, they often experience a sharp drop in training loss, a phenomenon cited as evidence that induction heads may serve as a prerequisite for more complex in-context learning (ICL) capabilities. In this work, we ask whether transformers can still acquire ICL capabilities when inductive copying is suppressed. We propose Hapax, a setting where we omit the loss contribution of any token that can be correctly predicted by induction heads. Despite a significant reduction in inductive copying, performance on abstractive ICL tasks (i.e., tasks where the answer is not contained in the input context) remains comparable and surpasses the vanilla model on 13 of 21 tasks, even though 31.7\% of tokens are omitted from the loss. Furthermore, our model achieves lower loss values on token positions that cannot be predicted correctly by induction heads. Mechanistic analysis further shows that models trained with Hapax develop fewer and weaker induction heads but still preserve ICL capabilities. Taken together, our findings indicate that inductive copying is not essential for learning abstractive ICL mechanisms.

Chris Wendler, Veniamin Veselovsky, Giovanni Monea et al.

We ask whether multilingual language models trained on unbalanced, English-dominated corpora use English as an internal pivot language -- a 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 the 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.

Romeo Valentin, Claudio Ferrari, Jérémy Scheurer et al.

We present our submission for the configuration task of the Machine Learning for Combinatorial Optimization (ML4CO) NeurIPS 2021 competition. The configuration task is to predict a good configuration of the open-source solver SCIP to solve a mixed integer linear program (MILP) efficiently. We pose this task as a supervised learning problem: First, we compile a large dataset of the solver performance for various configurations and all provided MILP instances. Second, we use this data to train a graph neural network that learns to predict a good configuration for a specific instance. The submission was tested on the three problem benchmarks of the competition and improved solver performance over the default by 12% and 35% and 8% across the hidden test instances. We ranked 3rd out of 15 on the global leaderboard and won the student leaderboard. We make our code publicly available at \url{https://github.com/RomeoV/ml4co-competition} .

Jannik Brinkmann, Chris Wendler, Christian Bartelt et al.

Human bilinguals often use similar brain regions to process multiple languages, depending on when they learned their second language and their proficiency. In large language models (LLMs), how are multiple languages learned and encoded? In this work, we explore the extent to which LLMs share representations of morphsyntactic concepts such as grammatical number, gender, and tense across languages. We train sparse autoencoders on Llama-3-8B and Aya-23-8B, and demonstrate that abstract grammatical concepts are often encoded in feature directions shared across many languages. We use causal interventions to verify the multilingual nature of these representations; specifically, we show that ablating only multilingual features decreases classifier performance to near-chance across languages. We then use these features to precisely modify model behavior in a machine translation task; this demonstrates both the generality and selectivity of these feature's roles in the network. Our findings suggest that even models trained predominantly on English data can develop robust, cross-lingual abstractions of morphosyntactic concepts.

Julian Minder, Kevin Du, Niklas Stoehr et al. · eth-zurich

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, instruction-tuned 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 subspace facilitates how the model chooses between context and prior knowledge, hinting at a simple fundamental mechanism that controls this behavior.

Clément Dumas, Chris Wendler, Veniamin Veselovsky et al.

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.

Viacheslav Surkov, Chris Wendler, Antonio Mari et al.

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.

Veniamin Veselovsky, Berke Argin, Benedikt Stroebl et al.

Just as humans display language patterns influenced by their native tongue when speaking new languages, LLMs often default to English-centric responses even when generating in other languages. Nevertheless, we observe that local cultural information persists within the models and can be readily activated for cultural customization. We first demonstrate that explicitly providing cultural context in prompts significantly improves the models' ability to generate culturally localized responses. We term the disparity in model performance with versus without explicit cultural context the explicit-implicit localization gap, indicating that while cultural knowledge exists within LLMs, it may not naturally surface in multilingual interactions if cultural context is not explicitly provided. Despite the explicit prompting benefit, however, the answers reduce in diversity and tend toward stereotypes. Second, we identify an explicit cultural customization vector, conserved across all non-English languages we explore, which enables LLMs to be steered from the synthetic English cultural world-model toward each non-English cultural world. Steered responses retain the diversity of implicit prompting and reduce stereotypes to dramatically improve the potential for customization. We discuss the implications of explicit cultural customization for understanding the conservation of alternative cultural world models within LLMs, and their controllable utility for translation, cultural customization, and the possibility of making the explicit implicit through soft control for expanded LLM function and appeal.

Andrew Lee, Lihao Sun, Chris Wendler et al.

How do reasoning models verify their own answers? We study this question by training a model using DeepSeek R1's recipe on the CountDown task. We leverage the fact that preference tuning leads to mode collapse, yielding a model that always produces highly structured chain-of-thought sequences. With this setup, we do top-down and bottom-up analyses to reverse-engineer how the model verifies its outputs. Top-down, we find Gated Linear Unit (GLU) weights encoding verification-related tokens, such as ``success'' or ``incorrect''. Bottom-up, we find that ``previous-token heads'' are mainly responsible for self-verification in our setup. Our analyses meet in the middle: drawing inspiration from inter-layer communication channels, we use the identified GLU weights to localize as few as three attention heads that can disable self-verification, pointing to a necessary component of a potentially larger verification circuit. Finally, we verify that similar verification components exist in our base model and a general reasoning DeepSeek-R1 model.

Saibo Geng, Nathan Ranchin, Yunzhen yao et al.

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%.

Saibo Geng, Sankalp Gambhir, Chris Wendler et al.

Tokenization is an important preprocessing step in the training and inference of large language models (LLMs). While there has been extensive research on the expressive power of the neural achitectures used in LLMs, the impact of tokenization has not been well understood. In this work, we demonstrate that tokenization, irrespective of the algorithm used, acts as an inverse homomorphism between strings and tokens. This suggests that the character space of the source language and the token space of the tokenized language are homomorphic, preserving the structural properties of the source language. Additionally, we explore the concept of proper tokenization, which refers to an unambiguous tokenization returned from the tokenizer. Our analysis reveals that the expressiveness of neural architectures in recognizing context-free languages is not affected by tokenization.

Dmitrii Troitskii, Koyena Pal, Chris Wendler et al.

Prior work has shown that a significant driver of performance in reasoning models is their ability to reason and self-correct. A distinctive marker in these reasoning traces is the token wait, which often signals reasoning behavior such as backtracking. Despite being such a complex behavior, little is understood of exactly why models do or do not decide to reason in this particular manner, which limits our understanding of what makes a reasoning model so effective. In this work, we address the question whether model's latents preceding wait tokens contain relevant information for modulating the subsequent reasoning process. We train crosscoders at multiple layers of DeepSeek-R1-Distill-Llama-8B and its base version, and introduce a latent attribution technique in the crosscoder setting. We locate a small set of features relevant for promoting/suppressing wait tokens' probabilities. Finally, through a targeted series of experiments analyzing max activating examples and causal interventions, we show that many of our identified features indeed are relevant for the reasoning process and give rise to different types of reasoning patterns such as restarting from the beginning, recalling prior knowledge, expressing uncertainty, and double-checking.

Jason Becker, Chris Wendler, Peter Baylies et al.

Instead of performing text-conditioned denoising in the image domain, latent diffusion models (LDMs) operate in latent space of a variational autoencoder (VAE), enabling more efficient processing at reduced computational costs. However, while the diffusion process has moved to the latent space, the contrastive language-image pre-training (CLIP) models, as used in many image processing tasks, still operate in pixel space. Doing so requires costly VAE-decoding of latent images before they can be processed. In this paper, we introduce Latent-CLIP, a CLIP model that operates directly in the latent space. We train Latent-CLIP on 2.7B pairs of latent images and descriptive texts, and show that it matches zero-shot classification performance of similarly sized CLIP models on both the ImageNet benchmark and a LDM-generated version of it, demonstrating its effectiveness in assessing both real and generated content. Furthermore, we construct Latent-CLIP rewards for reward-based noise optimization (ReNO) and show that they match the performance of their CLIP counterparts on GenEval and T2I-CompBench while cutting the cost of the total pipeline by 21%. Finally, we use Latent-CLIP to guide generation away from harmful content, achieving strong performance on the inappropriate image prompts (I2P) benchmark and a custom evaluation, without ever requiring the costly step of decoding intermediate images.

Saibo Geng, Berkay Döner, Chris Wendler et al.

Constrained decoding, a technique for enforcing constraints on language model outputs, offers a way to control text generation without retraining or architectural modifications. Its application is, however, typically restricted to models that give users access to next-token distributions (usually via softmax logits), which poses a limitation with blackbox large language models (LLMs). This paper introduces sketch-guided constrained decoding (SGCD), a novel approach to constrained decoding for blackbox LLMs, which operates without access to the logits of the blackbox LLM. SGCD utilizes a locally hosted auxiliary model to refine the output of an unconstrained blackbox LLM, effectively treating this initial output as a "sketch" for further elaboration. This approach is complementary to traditional logit-based techniques and enables the application of constrained decoding in settings where full model transparency is unavailable. We demonstrate the efficacy of SGCD through experiments in closed information extraction and constituency parsing, showing how it enhances the utility and flexibility of blackbox LLMs for complex NLP tasks.

Panagiotis Misiakos, Chris Wendler, Markus Püschel

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 with few root causes, with and without noise, we show superior performance compared to prior DAG learning methods.

Chris Wendler, Andisheh Amrollahi, Bastian Seifert et al.

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 $nk - k \log_2 k + k$ 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, 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.

Markus Püschel, Chris Wendler

Set functions are functions (or signals) indexed by the powerset (set of all subsets) of a finite set N. They are fundamental and ubiquitous in many application domains and have been used, for example, to formally describe or quantify loss functions for semantic image segmentation, the informativeness of sensors in sensor networks the utility of sets of items in recommender systems, cooperative games in game theory, or bidders in combinatorial auctions. In particular, the subclass of submodular functions occurs in many optimization and machine learning problems. In this paper, we derive discrete-set signal processing (SP), a novel shift-invariant linear signal processing framework for set functions. Discrete-set SP considers different notions of shift obtained from set union and difference operations. For each shift it provides associated notions of shift-invariant filters, convolution, Fourier transform, and frequency response. We provide intuition for our framework using the concept of generalized coverage function that we define, identify multivariate mutual information as a special case of a discrete-set spectrum, and motivate frequency ordering. Our work brings a new set of tools for analyzing and processing set functions, and, in particular, for dealing with their exponential nature. We show two prototypical applications and experiments: compression in submodular function optimization and sampling for preference elicitation in combinatorial auctions.

Chris Wendler, Dan Alistarh, Markus Püschel

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.