Atticus Geiger

Jing Huang, Atticus Geiger, Karel D'Oosterlinck et al. · stanford

Natural language is an appealing medium for explaining how large language models process and store information, but evaluating the faithfulness of such explanations is challenging. To help address this, we develop two modes of evaluation for natural language explanations that claim individual neurons represent a concept in a text input. In the observational mode, we evaluate claims that a neuron $a$ activates on all and only input strings that refer to a concept picked out by the proposed explanation $E$. In the intervention mode, we construe $E$ as a claim that the neuron $a$ is a causal mediator of the concept denoted by $E$. We apply our framework to the GPT-4-generated explanations of GPT-2 XL neurons of Bills et al. (2023) and show that even the most confident explanations have high error rates and little to no causal efficacy. We close the paper by critically assessing whether natural language is a good choice for explanations and whether neurons are the best level of analysis.

Zhengxuan Wu, Karel D'Oosterlinck, Atticus Geiger et al.

Explainability methods for NLP systems encounter a version of the fundamental problem of causal inference: for a given ground-truth input text, we never truly observe the counterfactual texts necessary for isolating the causal effects of model representations on outputs. In response, many explainability methods make no use of counterfactual texts, assuming they will be unavailable. In this paper, we show that robust causal explainability methods can be created using approximate counterfactuals, which can be written by humans to approximate a specific counterfactual or simply sampled using metadata-guided heuristics. The core of our proposal is the Causal Proxy Model (CPM). A CPM explains a black-box model $\mathcal{N}$ because it is trained to have the same actual input/output behavior as $\mathcal{N}$ while creating neural representations that can be intervened upon to simulate the counterfactual input/output behavior of $\mathcal{N}$. Furthermore, we show that the best CPM for $\mathcal{N}$ performs comparably to $\mathcal{N}$ in making factual predictions, which means that the CPM can simply replace $\mathcal{N}$, leading to more explainable deployed models. Our code is available at https://github.com/frankaging/Causal-Proxy-Model.

Atticus Geiger, Duligur Ibeling, Amir Zur et al. · stanford

Causal abstraction provides a theoretical foundation for mechanistic interpretability, the field concerned with providing intelligible algorithms that are faithful simplifications of the known, but opaque low-level details of black box AI models. Our contributions are (1) generalizing the theory of causal abstraction from mechanism replacement (i.e., hard and soft interventions) to arbitrary mechanism transformation (i.e., functionals from old mechanisms to new mechanisms), (2) providing a flexible, yet precise formalization for the core concepts of polysemantic neurons, the linear representation hypothesis, modular features, and graded faithfulness, and (3) unifying a variety of mechanistic interpretability methods in the common language of causal abstraction, namely, activation and path patching, causal mediation analysis, causal scrubbing, causal tracing, circuit analysis, concept erasure, sparse autoencoders, differential binary masking, distributed alignment search, and steering.

Maheep Chaudhary, Atticus Geiger

A popular new method in mechanistic interpretability is to train high-dimensional sparse autoencoders (SAEs) on neuron activations and use SAE features as the atomic units of analysis. However, the body of evidence on whether SAE feature spaces are useful for causal analysis is underdeveloped. In this work, we use the RAVEL benchmark to evaluate whether SAEs trained on hidden representations of GPT-2 small have sets of features that separately mediate knowledge of which country a city is in and which continent it is in. We evaluate four open-source SAEs for GPT-2 small against each other, with neurons serving as a baseline, and linear features learned via distributed alignment search (DAS) serving as a skyline. For each, we learn a binary mask to select features that will be patched to change the country of a city without changing the continent, or vice versa. Our results show that SAEs struggle to reach the neuron baseline, and none come close to the DAS skyline. We release code here: https://github.com/MaheepChaudhary/SAE-Ravel

Atticus Geiger, Zhengxuan Wu, Christopher Potts et al.

Causal abstraction is a promising theoretical framework for explainable artificial intelligence that defines when an interpretable high-level causal model is a faithful simplification of a low-level deep learning system. However, existing causal abstraction methods have two major limitations: they require a brute-force search over alignments between the high-level model and the low-level one, and they presuppose that variables in the high-level model will align with disjoint sets of neurons in the low-level one. In this paper, we present distributed alignment search (DAS), which overcomes these limitations. In DAS, we find the alignment between high-level and low-level models using gradient descent rather than conducting a brute-force search, and we allow individual neurons to play multiple distinct roles by analyzing representations in non-standard bases-distributed representations. Our experiments show that DAS can discover internal structure that prior approaches miss. Overall, DAS removes previous obstacles to conducting causal abstraction analyses and allows us to find conceptual structure in trained neural nets.

Eldar David Abraham, Karel D'Oosterlinck, Amir Feder et al.

The increasing size and complexity of modern ML systems has improved their predictive capabilities but made their behavior harder to explain. Many techniques for model explanation have been developed in response, but we lack clear criteria for assessing these techniques. In this paper, we cast model explanation as the causal inference problem of estimating causal effects of real-world concepts on the output behavior of ML models given actual input data. We introduce CEBaB, a new benchmark dataset for assessing concept-based explanation methods in Natural Language Processing (NLP). CEBaB consists of short restaurant reviews with human-generated counterfactual reviews in which an aspect (food, noise, ambiance, service) of the dining experience was modified. Original and counterfactual reviews are annotated with multiply-validated sentiment ratings at the aspect-level and review-level. The rich structure of CEBaB allows us to go beyond input features to study the effects of abstract, real-world concepts on model behavior. We use CEBaB to compare the quality of a range of concept-based explanation methods covering different assumptions and conceptions of the problem, and we seek to establish natural metrics for comparative assessments of these methods.

Curt Tigges, Oskar John Hollinsworth, Atticus Geiger et al.

Sentiment is a pervasive feature in natural language text, yet it is an open question how sentiment is represented within Large Language Models (LLMs). In this study, we reveal that across a range of models, sentiment is represented linearly: a single direction in activation space mostly captures the feature across a range of tasks with one extreme for positive and the other for negative. Through causal interventions, we isolate this direction and show it is causally relevant in both toy tasks and real world datasets such as Stanford Sentiment Treebank. Through this case study we model a thorough investigation of what a single direction means on a broad data distribution. We further uncover the mechanisms that involve this direction, highlighting the roles of a small subset of attention heads and neurons. Finally, we discover a phenomenon which we term the summarization motif: sentiment is not solely represented on emotionally charged words, but is additionally summarized at intermediate positions without inherent sentiment, such as punctuation and names. We show that in Stanford Sentiment Treebank zero-shot classification, 76% of above-chance classification accuracy is lost when ablating the sentiment direction, nearly half of which (36%) is due to ablating the summarized sentiment direction exclusively at comma positions.

Róbert Csordás, Christopher Potts, Christopher D. Manning et al.

The Linear Representation Hypothesis (LRH) states that neural networks learn to encode concepts as directions in activation space, and a strong version of the LRH states that models learn only such encodings. In this paper, we present a counterexample to this strong LRH: when trained to repeat an input token sequence, gated recurrent neural networks (RNNs) learn to represent the token at each position with a particular order of magnitude, rather than a direction. These representations have layered features that are impossible to locate in distinct linear subspaces. To show this, we train interventions to predict and manipulate tokens by learning the scaling factor corresponding to each sequence position. These interventions indicate that the smallest RNNs find only this magnitude-based solution, while larger RNNs have linear representations. These findings strongly indicate that interpretability research should not be confined by the LRH.

Riccardo Massidda, Atticus Geiger, Thomas Icard et al.

Causal abstraction provides a theory describing how several causal models can represent the same system at different levels of detail. Existing theoretical proposals limit the analysis of abstract models to "hard" interventions fixing causal variables to be constant values. In this work, we extend causal abstraction to "soft" interventions, which assign possibly non-constant functions to variables without adding new causal connections. Specifically, (i) we generalize $τ$-abstraction from Beckers and Halpern (2019) to soft interventions, (ii) we propose a further definition of soft abstraction to ensure a unique map $ω$ between soft interventions, and (iii) we prove that our constructive definition of soft abstraction guarantees the intervention map $ω$ has a specific and necessary explicit form.

Sheridan Feucht, Tal Haklay, Usha Bhalla et al.

Does structure in representations imply structure in computation? We study how Llama-3.1-8B reasons over cyclic concepts (e.g., "what month is six months after August?"). Even though Llama-3.1-8B's representations for these concepts are circularly structured, we find that instead of directly computing modular addition in the period of the cyclic concept (e.g., 12 for months), the model re-uses a generic addition mechanism across tasks that operates independently of concept-specific geometry. First, it computes the sum of its two inputs using base-10 addition (six + August=14). Then, it maps this sum back to cyclic concept space (14->February). We show that Llama-3.1-8B uses task-agnostic Fourier features to compute these sums--in fact, these features have periods that respect standard base-10 addition, e.g., 2, 5, and 10, rather than the cyclic concept period (e.g., 12 for months). Furthermore, we identify a sparse set of 28 MLP neurons re-used across all tasks (approximately 0.2% of the MLP at layer 18) that can be partitioned into disjoint clusters, each computing the sum for a Fourier feature with a different period. Our work highlights how an interplay between causal abstraction and feature geometry can deepen our mechanistic understanding of LMs.

Zhengxuan Wu, Aryaman Arora, Zheng Wang et al. · stanford

Parameter-efficient finetuning (PEFT) methods seek to adapt large neural models via updates to a small number of weights. However, much prior interpretability work has shown that representations encode rich semantic information, suggesting that editing representations might be a more powerful alternative. We pursue this hypothesis by developing a family of Representation Finetuning (ReFT) methods. ReFT methods operate on a frozen base model and learn task-specific interventions on hidden representations. We define a strong instance of the ReFT family, Low-rank Linear Subspace ReFT (LoReFT), and we identify an ablation of this method that trades some performance for increased efficiency. Both are drop-in replacements for existing PEFTs and learn interventions that are 15x--65x more parameter-efficient than LoRA. We showcase LoReFT on eight commonsense reasoning tasks, four arithmetic reasoning tasks, instruction-tuning, and GLUE. In all these evaluations, our ReFTs deliver the best balance of efficiency and performance, and almost always outperform state-of-the-art PEFTs. We release a generic ReFT training library publicly at https://github.com/stanfordnlp/pyreft.

Or Shafran, Shaked Ronen, Omri Fahn et al.

Activation decomposition methods in language models are tightly coupled to geometric assumptions on how concepts are realized in activation space. Existing approaches search for individual global directions, implicitly assuming linear separability, which overlooks concepts with nonlinear or multi-dimensional structure. In this work, we leverage Mixture of Factor Analyzers (MFA) as a scalable, unsupervised alternative that models the activation space as a collection of Gaussian regions with their local covariance structure. MFA decomposes activations into two compositional geometric objects: the region's centroid in activation space, and the local variation from the centroid. We train large-scale MFAs for Llama-3.1-8B and Gemma-2-2B, and show they capture complex, nonlinear structures in activation space. Moreover, evaluations on localization and steering benchmarks show that MFA outperforms unsupervised baselines, is competitive with supervised localization methods, and often achieves stronger steering performance than sparse autoencoders. Together, our findings position local geometry, expressed through subspaces, as a promising unit of analysis for scalable concept discovery and model control, accounting for complex structures that isolated directions fail to capture.

Jing Huang, Zhengxuan Wu, Christopher Potts et al. · deepmind, stanford

Individual neurons participate in the representation of multiple high-level concepts. To what extent can different interpretability methods successfully disentangle these roles? To help address this question, we introduce RAVEL (Resolving Attribute-Value Entanglements in Language Models), a dataset that enables tightly controlled, quantitative comparisons between a variety of existing interpretability methods. We use the resulting conceptual framework to define the new method of Multi-task Distributed Alignment Search (MDAS), which allows us to find distributed representations satisfying multiple causal criteria. With Llama2-7B as the target language model, MDAS achieves state-of-the-art results on RAVEL, demonstrating the importance of going beyond neuron-level analyses to identify features distributed across activations. We release our benchmark at https://github.com/explanare/ravel.

Zhengxuan Wu, Atticus Geiger, Aryaman Arora et al. · stanford

Interventions on model-internal states are fundamental operations in many areas of AI, including model editing, steering, robustness, and interpretability. To facilitate such research, we introduce $\textbf{pyvene}$, an open-source Python library that supports customizable interventions on a range of different PyTorch modules. $\textbf{pyvene}$ supports complex intervention schemes with an intuitive configuration format, and its interventions can be static or include trainable parameters. We show how $\textbf{pyvene}$ provides a unified and extensible framework for performing interventions on neural models and sharing the intervened upon models with others. We illustrate the power of the library via interpretability analyses using causal abstraction and knowledge localization. We publish our library through Python Package Index (PyPI) and provide code, documentation, and tutorials at https://github.com/stanfordnlp/pyvene.

Amir Zur, Atticus Geiger, Ekdeep Singh Lubana et al.

When a language model generates text, the selection of individual tokens might lead it down very different reasoning paths, making uncertainty difficult to quantify. In this work, we consider whether reasoning language models represent the alternate paths that they could take during generation. To test this hypothesis, we use hidden activations to control and predict a language model's uncertainty during chain-of-thought reasoning. In our experiments, we find a clear correlation between how uncertain a model is at different tokens, and how easily the model can be steered by controlling its activations. This suggests that activation interventions are most effective when there are alternate paths available to the model -- in other words, when it has not yet committed to a particular final answer. We also find that hidden activations can predict a model's future outcome distribution, demonstrating that models implicitly represent the space of possible paths.

Raphaël Sarfati, Eric Bigelow, Daniel Wurgaft et al.

Large language models (LLMs) represent prompt-conditioned beliefs (posteriors over answers and claims), but we lack a mechanistic account of how these beliefs are encoded in representation space, how they update with new evidence, and how interventions reshape them. We study a controlled setting in which Llama-3.2 generates samples from a normal distribution by implicitly inferring its parameters (mean and standard deviation) given only samples from the distribution in context. We find representations of curved "belief manifolds" for these parameters form with sufficient in-context learning and study how the model adapts when the distribution suddenly changes. While standard linear steering often pushes the model off-manifold and induces coupled, out-of-distribution shifts, geometry and field-aware steering better preserves the intended belief family. Our work demonstrates an example of linear field probing (LFP) as a simple approach to tile the data manifold and make interventions that respect the underlying geometry. We conclude that rich structure emerges naturally in LLMs and that purely linear concept representations are often an inadequate abstraction.

Eric Bigelow, Raphaël Sarfati, Daniel Wurgaft et al.

Large Language Models (LLMs) update their behavior in context, which can be viewed as a form of Bayesian inference. However, the structure of the latent hypothesis space over which this inference operates remains unclear. In this work, we propose that LLMs assign beliefs over a low-dimensional geometric space - a conceptual belief space - and that in-context learning corresponds to a trajectory through this space as beliefs are updated over time. Using story understanding as a natural setting for dynamic belief updating, we combine behavioral and representational analyses to study these trajectories. We find that (1) belief updates are well-described as trajectories on low-dimensional, structured manifolds; (2) this structure is reflected consistently in both model behavior and internal representations and can be decoded with simple linear probes to predict behavior; and (3) interventions on these representations causally steer belief trajectories, with effects that can be predicted from the geometry of the conceptual space. Together, our results provide a geometric account of belief dynamics in LLMs, grounding Bayesian interpretations of in-context learning in structured conceptual representations.

Daniel Wurgaft, Can Rager, Matthew Kowal et al.

Neural representations carry rich geometric structure; but does that structure causally shape behavior? To address this question, we intervene along paths through activation space defined by different geometries, and measure the behavioral trajectories they induce. In particular, we test whether interventions that respect the geometry of activation space will yield behaviors close to those the model exhibits naturally. Concretely, we first fit an activation manifold $M_h$ to representations and a behavior manifold $M_y$ to output probability distributions. We then test the link $M_h \leftrightarrow M_y$ via interventions: we find that steering along $M_h$, which we term manifold steering, yields behavioral trajectories that follow $M_y$, while linear steering -- which assumes a Euclidean geometry -- cuts through off-manifold regions and hence produces unnatural outputs. Moreover, optimizing interventions in activation space to produce paths along $M_y$ recovers activation trajectories that trace the curvature of $M_h$. We demonstrate this bidirectional relationship between the geometry of representation and behavior across tasks and modalities. In language models, we use reasoning tasks with cyclic and sequential geometries as well as in-context learning tasks with more complex graph geometries. In a video world model, we use a task with geometry corresponding to physical dynamics. Overall, our work shows that geometry in neural representation is not merely incidental, but is in fact the proper object for enabling principled control via intervention on internals. This recasts the core problem of steering from finding the right direction to finding the right geometry.

Aruna Sankaranarayanan, Amir Zur, Atticus Geiger et al.

Where should we intervene in a language model (LM) to control behaviors that are diffused across many tokens of a long-form response? We introduce Generative Causal Mediation (GCM), a procedure for selecting model components, e.g., attention heads, to steer a binary concept (e.g., talk in verse vs. talk in prose) from contrastive long-form responses. In GCM, we first construct a dataset of contrasting inputs and responses. Then, we quantify how individual model components mediate the contrastive concept and select the strongest mediators for steering. We evaluate GCM on three tasks--refusal, sycophancy, and style transfer--across three language models. GCM successfully localizes concepts expressed in long-form responses and consistently outperforms correlational probe-based baselines when steering with a sparse set of attention heads. Together, these results demonstrate that GCM provides an effective approach for localizing and controlling the long-form responses of LMs.

Zhengxuan Wu, Atticus Geiger, Thomas Icard et al.

Obtaining human-interpretable explanations of large, general-purpose language models is an urgent goal for AI safety. However, it is just as important that our interpretability methods are faithful to the causal dynamics underlying model behavior and able to robustly generalize to unseen inputs. Distributed Alignment Search (DAS) is a powerful gradient descent method grounded in a theory of causal abstraction that has uncovered perfect alignments between interpretable symbolic algorithms and small deep learning models fine-tuned for specific tasks. In the present paper, we scale DAS significantly by replacing the remaining brute-force search steps with learned parameters -- an approach we call Boundless DAS. This enables us to efficiently search for interpretable causal structure in large language models while they follow instructions. We apply Boundless DAS to the Alpaca model (7B parameters), which, off the shelf, solves a simple numerical reasoning problem. With Boundless DAS, we discover that Alpaca does this by implementing a causal model with two interpretable boolean variables. Furthermore, we find that the alignment of neural representations with these variables is robust to changes in inputs and instructions. These findings mark a first step toward faithfully understanding the inner-workings of our ever-growing and most widely deployed language models. Our tool is extensible to larger LLMs and is released publicly at `https://github.com/stanfordnlp/pyvene`.

Douwe Kiela, Max Bartolo, Yixin Nie et al.

We introduce Dynabench, an open-source platform for dynamic dataset creation and model benchmarking. Dynabench runs in a web browser and supports human-and-model-in-the-loop dataset creation: annotators seek to create examples that a target model will misclassify, but that another person will not. In this paper, we argue that Dynabench addresses a critical need in our community: contemporary models quickly achieve outstanding performance on benchmark tasks but nonetheless fail on simple challenge examples and falter in real-world scenarios. With Dynabench, dataset creation, model development, and model assessment can directly inform each other, leading to more robust and informative benchmarks. We report on four initial NLP tasks, illustrating these concepts and highlighting the promise of the platform, and address potential objections to dynamic benchmarking as a new standard for the field.

Christopher Potts, Zhengxuan Wu, Atticus Geiger et al.

We introduce DynaSent ('Dynamic Sentiment'), a new English-language benchmark task for ternary (positive/negative/neutral) sentiment analysis. DynaSent combines naturally occurring sentences with sentences created using the open-source Dynabench Platform, which facilities human-and-model-in-the-loop dataset creation. DynaSent has a total of 121,634 sentences, each validated by five crowdworkers, and its development and test splits are designed to produce chance performance for even the best models we have been able to develop; when future models solve this task, we will use them to create DynaSent version 2, continuing the dynamic evolution of this benchmark. Here, we report on the dataset creation effort, focusing on the steps we took to increase quality and reduce artifacts. We also present evidence that DynaSent's Neutral category is more coherent than the comparable category in other benchmarks, and we motivate training models from scratch for each round over successive fine-tuning.

Zhengxuan Wu, Aryaman Arora, Atticus Geiger et al. · stanford

Fine-grained steering of language model outputs is essential for safety and reliability. Prompting and finetuning are widely used to achieve these goals, but interpretability researchers have proposed a variety of representation-based techniques as well, including sparse autoencoders (SAEs), linear artificial tomography, supervised steering vectors, linear probes, and representation finetuning. At present, there is no benchmark for making direct comparisons between these proposals. Therefore, we introduce AxBench, a large-scale benchmark for steering and concept detection, and report experiments on Gemma-2-2B and 9B. For steering, we find that prompting outperforms all existing methods, followed by finetuning. For concept detection, representation-based methods such as difference-in-means, perform the best. On both evaluations, SAEs are not competitive. We introduce a novel weakly-supervised representational method (Rank-1 Representation Finetuning; ReFT-r1), which is competitive on both tasks while providing the interpretability advantages that prompting lacks. Along with AxBench, we train and publicly release SAE-scale feature dictionaries for ReFT-r1 and DiffMean.

Lee Sharkey, Bilal Chughtai, Joshua Batson et al. · deepmind

Mechanistic interpretability aims to understand the computational mechanisms underlying neural networks' capabilities in order to accomplish concrete scientific and engineering goals. Progress in this field thus promises to provide greater assurance over AI system behavior and shed light on exciting scientific questions about the nature of intelligence. Despite recent progress toward these goals, there are many open problems in the field that require solutions before many scientific and practical benefits can be realized: Our methods require both conceptual and practical improvements to reveal deeper insights; we must figure out how best to apply our methods in pursuit of specific goals; and the field must grapple with socio-technical challenges that influence and are influenced by our work. This forward-facing review discusses the current frontier of mechanistic interpretability and the open problems that the field may benefit from prioritizing.

Usha Bhalla, Thomas Fel, Can Rager et al.

Sparse autoencoders (SAEs) are widely used to extract interpretable features from neural network representations, often under the implicit assumption that concepts correspond to independent linear directions. However, a growing body of evidence suggests that many concepts are instead organized along low-dimensional manifolds encoding continuous geometric relationships. This raises three basic questions: what does it mean for an SAE to capture a manifold, when do existing SAE architectures do so, and how? We develop a theoretical framework that answers these questions and show that SAEs can capture manifolds in two fundamentally different ways: globally, by allocating a compact group of atoms whose linear span contains the entire manifold, or locally, by distributing it across features that each selectively tile a restricted region of the underlying geometry. Empirically, we find that SAEs suboptimally recover continuous structures, mixing the global subspace and local tiling solutions in a fragmented regime we call dilution. This explains why manifold structure is rarely visible at the level of individual concepts and motivates post-hoc unsupervised discovery methods that search for coherent groups of atoms rather than isolated directions. More broadly, our results suggest that future representation learning methods should treat geometric objects, not just individual directions, as the basic units of interpretability.

Li Puyin, Jiyuan Tan, Ahmad Jabbar et al.

We present a method for diagnosing interpretation in neural networks by identifying an input subspace where a proposed interpretation is highly faithful. Our method is particularly useful for causal-abstraction-style interpretability, where a high-level causal hypothesis is evaluated by interchange interventions. Rather than treating interchange intervention accuracy as a single global summary, we refine this framework by partitioning the input space into well-interpreted and under-interpreted regions according to pairwise interchange-intervention behavior. This turns causal abstraction from a purely global evaluation into a more diagnostic tool: it not only measures whether an interpretation works, but also reveals where it works, where it fails, and what distinguishes the two cases. This diagnostic view also provides practical heuristics for improving interpretations. By analyzing the structure of the well-interpreted and under-interpreted regions, we can identify missing distinctions in a high-level hypothesis, discover previously unmodeled intermediate variables, and combine complementary partial interpretations into a stronger one. We instantiate this idea as a simple four-step recipe and show that it yields informative error analyses across multiple causal abstraction settings. In a toy logic task, recursively applying the recipe recovers a high-level hypothesis from scratch. More broadly, our results suggest that partitioning the input space is a useful step toward more precise, constructive, and scalable mechanistic interpretability.

Siddharth Boppana, Annabel Ma, Max Loeffler et al.

We provide evidence of performative chain-of-thought (CoT) in reasoning models, where a model becomes strongly confident in its final answer, but continues generating tokens without revealing its internal belief. Our analysis compares activation probing, early forced answering, and a CoT monitor across two large models (DeepSeek-R1 671B & GPT-OSS 120B) and find task difficulty-specific differences: The model's final answer is decodable from activations far earlier in CoT than a monitor is able to say, especially for easy recall-based MMLU questions. We contrast this with genuine reasoning in difficult multihop GPQA-Diamond questions. Despite this, inflection points (e.g., backtracking, 'aha' moments) occur almost exclusively in responses where probes show large belief shifts, suggesting these behaviors track genuine uncertainty rather than learned "reasoning theater." Finally, probe-guided early exit reduces tokens by up to 80% on MMLU and 30% on GPQA-Diamond with similar accuracy, positioning attention probing as an efficient tool for detecting performative reasoning and enabling adaptive computation.

Yoav Gur-Arieh, Roy Mayan, Chen Agassy et al. · deepmind

Automated interpretability pipelines generate natural language descriptions for the concepts represented by features in large language models (LLMs), such as plants or the first word in a sentence. These descriptions are derived using inputs that activate the feature, which may be a dimension or a direction in the model's representation space. However, identifying activating inputs is costly, and the mechanistic role of a feature in model behavior is determined both by how inputs cause a feature to activate and by how feature activation affects outputs. Using steering evaluations, we reveal that current pipelines provide descriptions that fail to capture the causal effect of the feature on outputs. To fix this, we propose efficient, output-centric methods for automatically generating feature descriptions. These methods use the tokens weighted higher after feature stimulation or the highest weight tokens after applying the vocabulary "unembedding" head directly to the feature. Our output-centric descriptions better capture the causal effect of a feature on model outputs than input-centric descriptions, but combining the two leads to the best performance on both input and output evaluations. Lastly, we show that output-centric descriptions can be used to find inputs that activate features previously thought to be "dead".

Zhengxuan Wu, Atticus Geiger, Jing Huang et al. · stanford

We respond to the recent paper by Makelov et al. (2023), which reviews subspace interchange intervention methods like distributed alignment search (DAS; Geiger et al. 2023) and claims that these methods potentially cause "interpretability illusions". We first review Makelov et al. (2023)'s technical notion of what an "interpretability illusion" is, and then we show that even intuitive and desirable explanations can qualify as illusions in this sense. As a result, their method of discovering "illusions" can reject explanations they consider "non-illusory". We then argue that the illusions Makelov et al. (2023) see in practice are artifacts of their training and evaluation paradigms. We close by emphasizing that, though we disagree with their core characterization, Makelov et al. (2023)'s examples and discussion have undoubtedly pushed the field of interpretability forward.

Nikhil Prakash, Natalie Shapira, Arnab Sen Sharma et al.

How do language models (LMs) represent characters' beliefs, especially when those beliefs may differ from reality? This question lies at the heart of understanding the Theory of Mind (ToM) capabilities of LMs. We analyze LMs' ability to reason about characters' beliefs using causal mediation and abstraction. We construct a dataset, CausalToM, consisting of simple stories where two characters independently change the state of two objects, potentially unaware of each other's actions. Our investigation uncovers a pervasive algorithmic pattern that we call a lookback mechanism, which enables the LM to recall important information when it becomes necessary. The LM binds each character-object-state triple together by co-locating their reference information, represented as Ordering IDs (OIs), in low-rank subspaces of the state token's residual stream. When asked about a character's beliefs regarding the state of an object, the binding lookback retrieves the correct state OI and then the answer lookback retrieves the corresponding state token. When we introduce text specifying that one character is (not) visible to the other, we find that the LM first generates a visibility ID encoding the relation between the observing and the observed character OIs. In a visibility lookback, this ID is used to retrieve information about the observed character and update the observing character's beliefs. Our work provides insights into belief tracking mechanisms, taking a step toward reverse-engineering ToM reasoning in LMs.

Yiwei Wu, Atticus Geiger, Raphaël Millière

Variable binding -- the ability to associate variables with values -- is fundamental to symbolic computation and cognition. Although classical architectures typically implement variable binding via addressable memory, it is not well understood how modern neural networks lacking built-in binding operations may acquire this capacity. We investigate this by training a Transformer to dereference queried variables in symbolic programs where variables are assigned either numerical constants or other variables. Each program requires following chains of variable assignments up to four steps deep to find the queried value, and also contains irrelevant chains of assignments acting as distractors. Our analysis reveals a developmental trajectory with three distinct phases during training: (1) random prediction of numerical constants, (2) a shallow heuristic prioritizing early variable assignments, and (3) the emergence of a systematic mechanism for dereferencing assignment chains. Using causal interventions, we find that the model learns to exploit the residual stream as an addressable memory space, with specialized attention heads routing information across token positions. This mechanism allows the model to dynamically track variable bindings across layers, resulting in accurate dereferencing. Our results show how Transformer models can learn to implement systematic variable binding without explicit architectural support, bridging connectionist and symbolic approaches. To facilitate reproducible research, we developed Variable Scope, an interactive web platform for exploring our findings at https://variablescope.org

Aaron Mueller, Atticus Geiger, Sarah Wiegreffe et al. · stanford

How can we know whether new mechanistic interpretability methods achieve real improvements? In pursuit of lasting evaluation standards, we propose MIB, a Mechanistic Interpretability Benchmark, with two tracks spanning four tasks and five models. MIB favors methods that precisely and concisely recover relevant causal pathways or causal variables in neural language models. The circuit localization track compares methods that locate the model components - and connections between them - most important for performing a task (e.g., attribution patching or information flow routes). The causal variable localization track compares methods that featurize a hidden vector, e.g., sparse autoencoders (SAEs) or distributed alignment search (DAS), and align those features to a task-relevant causal variable. Using MIB, we find that attribution and mask optimization methods perform best on circuit localization. For causal variable localization, we find that the supervised DAS method performs best, while SAE features are not better than neurons, i.e., non-featurized hidden vectors. These findings illustrate that MIB enables meaningful comparisons, and increases our confidence that there has been real progress in the field.

Jiuding Sun, Jing Huang, Sidharth Baskaran et al. · stanford

Mechanistic interpretability has made great strides in identifying neural network features (e.g., directions in hidden activation space) that mediate concepts(e.g., the birth year of a person) and enable predictable manipulation. Distributed alignment search (DAS) leverages supervision from counterfactual data to learn concept features within hidden states, but DAS assumes we can afford to conduct a brute force search over potential feature locations. To address this, we present HyperDAS, a transformer-based hypernetwork architecture that (1) automatically locates the token-positions of the residual stream that a concept is realized in and (2) constructs features of those residual stream vectors for the concept. In experiments with Llama3-8B, HyperDAS achieves state-of-the-art performance on the RAVEL benchmark for disentangling concepts in hidden states. In addition, we review the design decisions we made to mitigate the concern that HyperDAS (like all powerful interpretabilty methods) might inject new information into the target model rather than faithfully interpreting it.

Atticus Geiger, Jacqueline Harding, Thomas Icard

Explanations of cognitive behavior often appeal to computations over representations. What does it take for a system to implement a given computation over suitable representational vehicles within that system? We argue that the language of causality -- and specifically the theory of causal abstraction -- provides a fruitful lens on this topic. Drawing on current discussions in deep learning with artificial neural networks, we illustrate how classical themes in the philosophy of computation and cognition resurface in contemporary machine learning. We offer an account of computational implementation grounded in causal abstraction, and examine the role for representation in the resulting picture. We argue that these issues are most profitably explored in connection with generalization and prediction.

Jiuding Sun, Sidharth Baskaran, Zhengxuan Wu et al.

Steering language models (LMs) by modifying internal activations is a popular approach for controlling text generation. Unsupervised dictionary learning methods, e.g., sparse autoencoders, can be scaled to produce many steering vectors, but lack guarantees on the individual efficacy of each vector and control over the coverage of relevant steering tasks. In contrast, supervised methods for constructing steering vectors are targeted and effective, but require more data collection and training for each additional steering vector produced. In this work, we introduce HyperSteer, a family of hypernetwork-based architectures which are trained end-to-end to generate steering vectors conditioned on the natural language steering prompts and the internals of the steered LM. In our evaluations, we show that scaling HyperSteer with thousands of steering prompts exceeds the performance of state-of-the-art activation steering methods, even on steering prompts never seen during training. Moreover, HyperSteer performs on par with steering-via-prompting.

Or Shafran, Atticus Geiger, Mor Geva · deepmind

A central goal for mechanistic interpretability has been to identify the right units of analysis in large language models (LLMs) that causally explain their outputs. While early work focused on individual neurons, evidence that neurons often encode multiple concepts has motivated a shift toward analyzing directions in activation space. A key question is how to find directions that capture interpretable features in an unsupervised manner. Current methods rely on dictionary learning with sparse autoencoders (SAEs), commonly trained over residual stream activations to learn directions from scratch. However, SAEs often struggle in causal evaluations and lack intrinsic interpretability, as their learning is not explicitly tied to the computations of the model. Here, we tackle these limitations by directly decomposing MLP activations with semi-nonnegative matrix factorization (SNMF), such that the learned features are (a) sparse linear combinations of co-activated neurons, and (b) mapped to their activating inputs, making them directly interpretable. Experiments on Llama 3.1, Gemma 2 and GPT-2 show that SNMF derived features outperform SAEs and a strong supervised baseline (difference-in-means) on causal steering, while aligning with human-interpretable concepts. Further analysis reveals that specific neuron combinations are reused across semantically-related features, exposing a hierarchical structure in the MLP's activation space. Together, these results position SNMF as a simple and effective tool for identifying interpretable features and dissecting concept representations in LLMs.

Theodora-Mara Pîslar, Sara Magliacane, Atticus Geiger

Mechanistic interpretability aims to reverse engineer neural networks by uncovering which high-level algorithms they implement. Causal abstraction provides a precise notion of when a network implements an algorithm, i.e., a causal model of the network contains low-level features that realize the high-level variables in a causal model of the algorithm. A typical problem in practical settings is that the algorithm is not an entirely faithful abstraction of the network, meaning it only partially captures the true reasoning process of a model. We propose a solution where we combine different simple high-level models to produce a more faithful representation of the network. Through learning this combination, we can model neural networks as being in different computational states depending on the input provided, which we show is more accurate to GPT 2-small fine-tuned on two toy tasks. We observe a trade-off between the strength of an interpretability hypothesis, which we define in terms of the number of inputs explained by the high-level models, and its faithfulness, which we define as the interchange intervention accuracy. Our method allows us to modulate between the two, providing the most accurate combination of models that describe the behavior of a neural network given a faithfulness level.

Yoav Gur-Arieh, Mor Geva, Atticus Geiger · deepmind

A key component of in-context reasoning is the ability of language models (LMs) to bind entities for later retrieval. For example, an LM might represent "Ann loves pie" by binding "Ann" to "pie", allowing it to later retrieve "Ann" when asked "Who loves pie?" Prior research on short lists of bound entities found strong evidence that LMs implement such retrieval via a positional mechanism, where "Ann" is retrieved based on its position in context. In this work, we find that this mechanism generalizes poorly to more complex settings; as the number of bound entities in context increases, the positional mechanism becomes noisy and unreliable in middle positions. To compensate for this, we find that LMs supplement the positional mechanism with a lexical mechanism (retrieving "Ann" using its bound counterpart "pie") and a reflexive mechanism (retrieving "Ann" through a direct pointer). Through extensive experiments on nine models and ten binding tasks, we uncover a consistent pattern in how LMs mix these mechanisms to drive model behavior. We leverage these insights to develop a causal model combining all three mechanisms that estimates next token distributions with 95% agreement. Finally, we show that our model generalizes to substantially longer inputs of open-ended text interleaved with entity groups, further demonstrating the robustness of our findings in more natural settings. Overall, our study establishes a more complete picture of how LMs bind and retrieve entities in-context.

Amir Zur, Elisa Kreiss, Karel D'Oosterlinck et al.

Although CLIPScore is a powerful generic metric that captures the similarity between a text and an image, it fails to distinguish between a caption that is meant to complement the information in an image and a description that is meant to replace an image entirely, e.g., for accessibility. We address this shortcoming by updating the CLIP model with the Concadia dataset to assign higher scores to descriptions than captions using parameter efficient fine-tuning and a loss objective derived from work on causal interpretability. This model correlates with the judgements of blind and low-vision people while preserving transfer capabilities and has interpretable structure that sheds light on the caption--description distinction.

Jingyuan Selena She, Christopher Potts, Samuel R. Bowman et al.

A number of recent benchmarks seek to assess how well models handle natural language negation. However, these benchmarks lack the controlled example paradigms that would allow us to infer whether a model had learned how negation morphemes semantically scope. To fill these analytical gaps, we present the Scoped Negation NLI (ScoNe-NLI) benchmark, which contains contrast sets of six examples with up to two negations where either zero, one, or both negative morphemes affect the NLI label. We use ScoNe-NLI to assess fine-tuning and in-context learning strategies. We find that RoBERTa and DeBERTa models solve ScoNe-NLI after many shot fine-tuning. For in-context learning, we test InstructGPT models and find that most prompt strategies are not successful, including those using step-by-step reasoning. To better understand this result, we extend ScoNe with ScoNe-NLG, a sentence completion test set that embeds negation reasoning in short narratives. Here, InstructGPT is successful, which reveals the model can correctly reason about negation, but struggles to do so on prompt-adapted NLI examples outside of its core pretraining regime.

Zhengxuan Wu, Atticus Geiger, Josh Rozner et al.

Distillation efforts have led to language models that are more compact and efficient without serious drops in performance. The standard approach to distillation trains a student model against two objectives: a task-specific objective (e.g., language modeling) and an imitation objective that encourages the hidden states of the student model to be similar to those of the larger teacher model. In this paper, we show that it is beneficial to augment distillation with a third objective that encourages the student to imitate the causal computation process of the teacher through interchange intervention training(IIT). IIT pushes the student model to become a causal abstraction of the teacher model - a simpler model with the same causal structure. IIT is fully differentiable, easily implemented, and combines flexibly with other objectives. Compared with standard distillation of BERT, distillation via IIT results in lower perplexity on Wikipedia (masked language modeling) and marked improvements on the GLUE benchmark (natural language understanding), SQuAD (question answering), and CoNLL-2003 (named entity recognition).

Atticus Geiger, Zhengxuan Wu, Hanson Lu et al.

In many areas, we have well-founded insights about causal structure that would be useful to bring into our trained models while still allowing them to learn in a data-driven fashion. To achieve this, we present the new method of interchange intervention training (IIT). In IIT, we (1) align variables in a causal model (e.g., a deterministic program or Bayesian network) with representations in a neural model and (2) train the neural model to match the counterfactual behavior of the causal model on a base input when aligned representations in both models are set to be the value they would be for a source input. IIT is fully differentiable, flexibly combines with other objectives, and guarantees that the target causal model is a causal abstraction of the neural model when its loss is zero. We evaluate IIT on a structural vision task (MNIST-PVR), a navigational language task (ReaSCAN), and a natural language inference task (MQNLI). We compare IIT against multi-task training objectives and data augmentation. In all our experiments, IIT achieves the best results and produces neural models that are more interpretable in the sense that they more successfully realize the target causal model.

Atticus Geiger, Hanson Lu, Thomas Icard et al.

Structural analysis methods (e.g., probing and feature attribution) are increasingly important tools for neural network analysis. We propose a new structural analysis method grounded in a formal theory of causal abstraction that provides rich characterizations of model-internal representations and their roles in input/output behavior. In this method, neural representations are aligned with variables in interpretable causal models, and then interchange interventions are used to experimentally verify that the neural representations have the causal properties of their aligned variables. We apply this method in a case study to analyze neural models trained on Multiply Quantified Natural Language Inference (MQNLI) corpus, a highly complex NLI dataset that was constructed with a tree-structured natural logic causal model. We discover that a BERT-based model with state-of-the-art performance successfully realizes parts of the natural logic model's causal structure, whereas a simpler baseline model fails to show any such structure, demonstrating that BERT representations encode the compositional structure of MQNLI.

Atticus Geiger, Alexandra Carstensen, Michael C. Frank et al.

The notion of equality (identity) is simple and ubiquitous, making it a key case study for broader questions about the representations supporting abstract relational reasoning. Previous work suggested that neural networks were not suitable models of human relational reasoning because they could not represent mathematically identity, the most basic form of equality. We revisit this question. In our experiments, we assess out-of-sample generalization of equality using both arbitrary representations and representations that have been pretrained on separate tasks to imbue them with structure. We find neural networks are able to learn (1) basic equality (mathematical identity), (2) sequential equality problems (learning ABA-patterned sequences) with only positive training instances, and (3) a complex, hierarchical equality problem with only basic equality training instances ("zero-shot'" generalization). In the two latter cases, our models perform tasks proposed in previous work to demarcate human-unique symbolic abilities. These results suggest that essential aspects of symbolic reasoning can emerge from data-driven, non-symbolic learning processes.

Atticus Geiger, Kyle Richardson, Christopher Potts

We address whether neural models for Natural Language Inference (NLI) can learn the compositional interactions between lexical entailment and negation, using four methods: the behavioral evaluation methods of (1) challenge test sets and (2) systematic generalization tasks, and the structural evaluation methods of (3) probes and (4) interventions. To facilitate this holistic evaluation, we present Monotonicity NLI (MoNLI), a new naturalistic dataset focused on lexical entailment and negation. In our behavioral evaluations, we find that models trained on general-purpose NLI datasets fail systematically on MoNLI examples containing negation, but that MoNLI fine-tuning addresses this failure. In our structural evaluations, we look for evidence that our top-performing BERT-based model has learned to implement the monotonicity algorithm behind MoNLI. Probes yield evidence consistent with this conclusion, and our intervention experiments bolster this, showing that the causal dynamics of the model mirror the causal dynamics of this algorithm on subsets of MoNLI. This suggests that the BERT model at least partially embeds a theory of lexical entailment and negation at an algorithmic level.

Atticus Geiger, Ignacio Cases, Lauri Karttunen et al.

Deep learning models for semantics are generally evaluated using naturalistic corpora. Adversarial methods, in which models are evaluated on new examples with known semantic properties, have begun to reveal that good performance at these naturalistic tasks can hide serious shortcomings. However, we should insist that these evaluations be fair -that the models are given data sufficient to support the requisite kinds of generalization. In this paper, we define and motivate a formal notion of fairness in this sense. We then apply these ideas to natural language inference by constructing very challenging but provably fair artificial datasets and showing that standard neural models fail to generalize in the required ways; only task-specific models that jointly compose the premise and hypothesis are able to achieve high performance, and even these models do not solve the task perfectly.

Atticus Geiger, Ignacio Cases, Lauri Karttunen et al.

Standard evaluations of deep learning models for semantics using naturalistic corpora are limited in what they can tell us about the fidelity of the learned representations, because the corpora rarely come with good measures of semantic complexity. To overcome this limitation, we present a method for generating data sets of multiply-quantified natural language inference (NLI) examples in which semantic complexity can be precisely characterized, and we use this method to show that a variety of common architectures for NLI inevitably fail to encode crucial information; only a model with forced lexical alignments avoids this damaging information loss.