Yonatan Belinkov

Kevin Meng, Arnab Sen Sharma, Alex Andonian et al. · mit

Recent work has shown exciting promise in updating large language models with new memories, so as to replace obsolete information or add specialized knowledge. However, this line of work is predominantly limited to updating single associations. We develop MEMIT, a method for directly updating a language model with many memories, demonstrating experimentally that it can scale up to thousands of associations for GPT-J (6B) and GPT-NeoX (20B), exceeding prior work by orders of magnitude. Our code and data are at https://memit.baulab.info.

Rohit Gandikota, Hadas Orgad, Yonatan Belinkov et al.

Text-to-image models suffer from various safety issues that may limit their suitability for deployment. Previous methods have separately addressed individual issues of bias, copyright, and offensive content in text-to-image models. However, in the real world, all of these issues appear simultaneously in the same model. We present a method that tackles all issues with a single approach. Our method, Unified Concept Editing (UCE), edits the model without training using a closed-form solution, and scales seamlessly to concurrent edits on text-conditional diffusion models. We demonstrate scalable simultaneous debiasing, style erasure, and content moderation by editing text-to-image projections, and we present extensive experiments demonstrating improved efficacy and scalability over prior work. Our code is available at https://unified.baulab.info

Evan Hernandez, Arnab Sen Sharma, Tal Haklay et al. · microsoft-research, mit

Much of the knowledge encoded in transformer language models (LMs) may be expressed in terms of relations: relations between words and their synonyms, entities and their attributes, etc. We show that, for a subset of relations, this computation is well-approximated by a single linear transformation on the subject representation. Linear relation representations may be obtained by constructing a first-order approximation to the LM from a single prompt, and they exist for a variety of factual, commonsense, and linguistic relations. However, we also identify many cases in which LM predictions capture relational knowledge accurately, but this knowledge is not linearly encoded in their representations. Our results thus reveal a simple, interpretable, but heterogeneously deployed knowledge representation strategy in transformer LMs.

Dor Muhlgay, Ori Ram, Inbal Magar et al.

Before deploying a language model (LM) within a given domain, it is important to measure its tendency to generate factually incorrect information in that domain. Existing methods for factuality evaluation of LLM generation focus on facts sampled from the LM itself, and thus do not control the set of evaluated facts and might under-represent domain specific or rare facts. We propose FACTOR: Factual Assessment via Corpus TransfORmation, a scalable approach for evaluating LM factuality. FACTOR automatically transforms a factual corpus of interest into a benchmark evaluating an LM's propensity to generate true facts from the corpus vs. similar but incorrect statements. We use our framework to create three benchmarks: Wiki-FACTOR, News-FACTOR and Expert-FACTOR. We show that: (i) our benchmark scores increase with model size and improve when the LM is augmented with retrieval; (ii) benchmark score and perplexity do not always agree on model ranking; (iii) when perplexity and benchmark score disagree, the latter better reflects factuality in open-ended generation, as measured by human annotators. We make our data and code publicly available in https://github.com/AI21Labs/factor.

Nir Ratner, Yoav Levine, Yonatan Belinkov et al.

When applied to processing long text, Large Language Models (LLMs) are limited by their context window. Existing efforts to address this limitation involve training specialized architectures, and cannot be easily applied to off-the-shelf LLMs. We present Parallel Context Windows (PCW), a method that alleviates the context window restriction for any off-the-shelf LLM without further training. The key to the approach is to carve a long context into chunks (``windows''), restrict the attention mechanism to apply only within each window, and re-use the positional embeddings across the windows. Our main results test the PCW approach on in-context learning with models that range in size between 750 million and 178 billion parameters, and show substantial improvements for tasks with diverse input and output spaces. We show additional benefits in other settings where long context windows may be beneficial: multi-hop questions and retrieval-augmented question answering with multiple retrieved documents. Our results highlight Parallel Context Windows as a promising method for applying off-the-shelf LLMs in a range of settings that require long text sequences. We make our code publicly available at https://github.com/ai21labs/parallel-context-windows.

Jamba Team, Barak Lenz, Alan Arazi et al.

We present Jamba-1.5, new instruction-tuned large language models based on our Jamba architecture. Jamba is a hybrid Transformer-Mamba mixture of experts architecture, providing high throughput and low memory usage across context lengths, while retaining the same or better quality as Transformer models. We release two model sizes: Jamba-1.5-Large, with 94B active parameters, and Jamba-1.5-Mini, with 12B active parameters. Both models are fine-tuned for a variety of conversational and instruction-following capabilties, and have an effective context length of 256K tokens, the largest amongst open-weight models. To support cost-effective inference, we introduce ExpertsInt8, a novel quantization technique that allows fitting Jamba-1.5-Large on a machine with 8 80GB GPUs when processing 256K-token contexts without loss of quality. When evaluated on a battery of academic and chatbot benchmarks, Jamba-1.5 models achieve excellent results while providing high throughput and outperforming other open-weight models on long-context benchmarks. The model weights for both sizes are publicly available under the Jamba Open Model License and we release ExpertsInt8 as open source.

Yanai Elazar, Nora Kassner, Shauli Ravfogel et al. · cmu

Large amounts of training data are one of the major reasons for the high performance of state-of-the-art NLP models. But what exactly in the training data causes a model to make a certain prediction? We seek to answer this question by providing a language for describing how training data influences predictions, through a causal framework. Importantly, our framework bypasses the need to retrain expensive models and allows us to estimate causal effects based on observational data alone. Addressing the problem of extracting factual knowledge from pretrained language models (PLMs), we focus on simple data statistics such as co-occurrence counts and show that these statistics do influence the predictions of PLMs, suggesting that such models rely on shallow heuristics. Our causal framework and our results demonstrate the importance of studying datasets and the benefits of causality for understanding NLP models.

Ori Ram, Liat Bezalel, Adi Zicher et al. · deepmind

Dual encoders are now the dominant architecture for dense retrieval. Yet, we have little understanding of how they represent text, and why this leads to good performance. In this work, we shed light on this question via distributions over the vocabulary. We propose to interpret the vector representations produced by dual encoders by projecting them into the model's vocabulary space. We show that the resulting projections contain rich semantic information, and draw connection between them and sparse retrieval. We find that this view can offer an explanation for some of the failure cases of dense retrievers. For example, we observe that the inability of models to handle tail entities is correlated with a tendency of the token distributions to forget some of the tokens of those entities. We leverage this insight and propose a simple way to enrich query and passage representations with lexical information at inference time, and show that this significantly improves performance compared to the original model in zero-shot settings, and specifically on the BEIR benchmark.

Michael Hanna, Yonatan Belinkov, Sandro Pezzelle

Animacy - whether an entity is alive and sentient - is fundamental to cognitive processing, impacting areas such as memory, vision, and language. However, animacy is not always expressed directly in language: in English it often manifests indirectly, in the form of selectional constraints on verbs and adjectives. This poses a potential issue for transformer language models (LMs): they often train only on text, and thus lack access to extralinguistic information from which humans learn about animacy. We ask: how does this impact LMs' animacy processing - do they still behave as humans do? We answer this question using open-source LMs. Like previous studies, we find that LMs behave much like humans when presented with entities whose animacy is typical. However, we also show that even when presented with stories about atypically animate entities, such as a peanut in love, LMs adapt: they treat these entities as animate, though they do not adapt as well as humans. Even when the context indicating atypical animacy is very short, LMs pick up on subtle clues and change their behavior. We conclude that despite the limited signal through which LMs can learn about animacy, they are indeed sensitive to the relevant lexical semantic nuances available in English.

Ehud Karpas, Omri Abend, Yonatan Belinkov et al.

Huge language models (LMs) have ushered in a new era for AI, serving as a gateway to natural-language-based knowledge tasks. Although an essential element of modern AI, LMs are also inherently limited in a number of ways. We discuss these limitations and how they can be avoided by adopting a systems approach. Conceptualizing the challenge as one that involves knowledge and reasoning in addition to linguistic processing, we define a flexible architecture with multiple neural models, complemented by discrete knowledge and reasoning modules. We describe this neuro-symbolic architecture, dubbed the Modular Reasoning, Knowledge and Language (MRKL, pronounced "miracle") system, some of the technical challenges in implementing it, and Jurassic-X, AI21 Labs' MRKL system implementation.

Hadas Orgad, Bahjat Kawar, Yonatan Belinkov

Text-to-image diffusion models often make implicit assumptions about the world when generating images. While some assumptions are useful (e.g., the sky is blue), they can also be outdated, incorrect, or reflective of social biases present in the training data. Thus, there is a need to control these assumptions without requiring explicit user input or costly re-training. In this work, we aim to edit a given implicit assumption in a pre-trained diffusion model. Our Text-to-Image Model Editing method, TIME for short, receives a pair of inputs: a "source" under-specified prompt for which the model makes an implicit assumption (e.g., "a pack of roses"), and a "destination" prompt that describes the same setting, but with a specified desired attribute (e.g., "a pack of blue roses"). TIME then updates the model's cross-attention layers, as these layers assign visual meaning to textual tokens. We edit the projection matrices in these layers such that the source prompt is projected close to the destination prompt. Our method is highly efficient, as it modifies a mere 2.2% of the model's parameters in under one second. To evaluate model editing approaches, we introduce TIMED (TIME Dataset), containing 147 source and destination prompt pairs from various domains. Our experiments (using Stable Diffusion) show that TIME is successful in model editing, generalizes well for related prompts unseen during editing, and imposes minimal effect on unrelated generations.

Itay Itzhak, Gabriel Stanovsky, Nir Rosenfeld et al.

Recent studies show that instruction tuning (IT) and reinforcement learning from human feedback (RLHF) improve the abilities of large language models (LMs) dramatically. While these tuning methods can help align models with human objectives and generate high-quality text, not much is known about their potential adverse effects. In this work, we investigate the effect of IT and RLHF on decision making and reasoning in LMs, focusing on three cognitive biases - the decoy effect, the certainty effect, and the belief bias - all of which are known to influence human decision-making and reasoning. Our findings highlight the presence of these biases in various models from the GPT-3, Mistral, and T5 families. Notably, we find a stronger presence of biases in models that have undergone instruction tuning, such as Flan-T5, Mistral-Instruct, GPT3.5, and GPT4. Our work constitutes a step toward comprehending cognitive biases in instruction-tuned LMs, which is crucial for the development of more reliable and unbiased language models.

Hadas Orgad, Yonatan Belinkov

Considerable efforts to measure and mitigate gender bias in recent years have led to the introduction of an abundance of tasks, datasets, and metrics used in this vein. In this position paper, we assess the current paradigm of gender bias evaluation and identify several flaws in it. First, we highlight the importance of extrinsic bias metrics that measure how a model's performance on some task is affected by gender, as opposed to intrinsic evaluations of model representations, which are less strongly connected to specific harms to people interacting with systems. We find that only a few extrinsic metrics are measured in most studies, although more can be measured. Second, we find that datasets and metrics are often coupled, and discuss how their coupling hinders the ability to obtain reliable conclusions, and how one may decouple them. We then investigate how the choice of the dataset and its composition, as well as the choice of the metric, affect bias measurement, finding significant variations across each of them. Finally, we propose several guidelines for more reliable gender bias evaluation.

Dana Arad, Hadas Orgad, Yonatan Belinkov

Our world is marked by unprecedented technological, global, and socio-political transformations, posing a significant challenge to text-to-image generative models. These models encode factual associations within their parameters that can quickly become outdated, diminishing their utility for end-users. To that end, we introduce ReFACT, a novel approach for editing factual associations in text-to-image models without relaying on explicit input from end-users or costly re-training. ReFACT updates the weights of a specific layer in the text encoder, modifying only a tiny portion of the model's parameters and leaving the rest of the model unaffected. We empirically evaluate ReFACT on an existing benchmark, alongside a newly curated dataset. Compared to other methods, ReFACT achieves superior performance in both generalization to related concepts and preservation of unrelated concepts. Furthermore, ReFACT maintains image generation quality, making it a practical tool for updating and correcting factual information in text-to-image models.

Hadas Orgad, Yonatan Belinkov

Models trained on real-world data tend to imitate and amplify social biases. Common methods to mitigate biases require prior information on the types of biases that should be mitigated (e.g., gender or racial bias) and the social groups associated with each data sample. In this work, we introduce BLIND, a method for bias removal with no prior knowledge of the demographics in the dataset. While training a model on a downstream task, BLIND detects biased samples using an auxiliary model that predicts the main model's success, and down-weights those samples during the training process. Experiments with racial and gender biases in sentiment classification and occupation classification tasks demonstrate that BLIND mitigates social biases without relying on a costly demographic annotation process. Our method is competitive with other methods that require demographic information and sometimes even surpasses them.

Hadas Orgad, Seraphina Goldfarb-Tarrant, Yonatan Belinkov

Common studies of gender bias in NLP focus either on extrinsic bias measured by model performance on a downstream task or on intrinsic bias found in models' internal representations. However, the relationship between extrinsic and intrinsic bias is relatively unknown. In this work, we illuminate this relationship by measuring both quantities together: we debias a model during downstream fine-tuning, which reduces extrinsic bias, and measure the effect on intrinsic bias, which is operationalized as bias extractability with information-theoretic probing. Through experiments on two tasks and multiple bias metrics, we show that our intrinsic bias metric is a better indicator of debiasing than (a contextual adaptation of) the standard WEAT metric, and can also expose cases of superficial debiasing. Our framework provides a comprehensive perspective on bias in NLP models, which can be applied to deploy NLP systems in a more informed manner. Our code and model checkpoints are publicly available.

Rachit Bansal, Danish Pruthi, Yonatan Belinkov

Neural networks are known to exploit spurious artifacts (or shortcuts) that co-occur with a target label, exhibiting heuristic memorization. On the other hand, networks have been shown to memorize training examples, resulting in example-level memorization. These kinds of memorization impede generalization of networks beyond their training distributions. Detecting such memorization could be challenging, often requiring researchers to curate tailored test sets. In this work, we hypothesize -- and subsequently show -- that the diversity in the activation patterns of different neurons is reflective of model generalization and memorization. We quantify the diversity in the neural activations through information-theoretic measures and find support for our hypothesis on experiments spanning several natural language and vision tasks. Importantly, we discover that information organization points to the two forms of memorization, even for neural activations computed on unlabelled in-distribution examples. Lastly, we demonstrate the utility of our findings for the problem of model selection. The associated code and other resources for this work are available at https://rachitbansal.github.io/information-measures.

Boaz Carmeli, Ron Meir, Yonatan Belinkov

The field of emergent communication aims to understand the characteristics of communication as it emerges from artificial agents solving tasks that require information exchange. Communication with discrete messages is considered a desired characteristic, for both scientific and applied reasons. However, training a multi-agent system with discrete communication is not straightforward, requiring either reinforcement learning algorithms or relaxing the discreteness requirement via a continuous approximation such as the Gumbel-softmax. Both these solutions result in poor performance compared to fully continuous communication. In this work, we propose an alternative approach to achieve discrete communication -- quantization of communicated messages. Using message quantization allows us to train the model end-to-end, achieving superior performance in multiple setups. Moreover, quantization is a natural framework that runs the gamut from continuous to discrete communication. Thus, it sets the ground for a broader view of multi-agent communication in the deep learning era.

Kerem Zaman, Yonatan Belinkov

Most evaluations of attribution methods focus on the English language. In this work, we present a multilingual approach for evaluating attribution methods for the Natural Language Inference (NLI) task in terms of faithfulness and plausibility. First, we introduce a novel cross-lingual strategy to measure faithfulness based on word alignments, which eliminates the drawbacks of erasure-based evaluations.We then perform a comprehensive evaluation of attribution methods, considering different output mechanisms and aggregation methods. Finally, we augment the XNLI dataset with highlight-based explanations, providing a multilingual NLI dataset with highlights, to support future exNLP studies. Our results show that attribution methods performing best for plausibility and faithfulness are different.

Sarah Wiegreffe, Oyvind Tafjord, Yonatan Belinkov et al.

Multiple-choice question answering (MCQA) is a key competence of performant transformer language models that is tested by mainstream benchmarks. However, recent evidence shows that models can have quite a range of performance, particularly when the task format is diversified slightly (such as by shuffling answer choice order). In this work we ask: how do successful models perform formatted MCQA? We employ vocabulary projection and activation patching methods to localize key hidden states that encode relevant information for predicting the correct answer. We find that the prediction of a specific answer symbol is causally attributed to a few middle layers, and specifically their multi-head self-attention mechanisms. We show that subsequent layers increase the probability of the predicted answer symbol in vocabulary space, and that this probability increase is associated with a sparse set of attention heads with unique roles. We additionally uncover differences in how different models adjust to alternative symbols. Finally, we demonstrate that a synthetic task can disentangle sources of model error to pinpoint when a model has learned formatted MCQA, and show that logit differences between answer choice tokens continue to grow over the course of training.

Omer Antverg, Eyal Ben-David, Yonatan Belinkov

Large pre-trained models are usually fine-tuned on downstream task data, and tested on unseen data. When the train and test data come from different domains, the model is likely to struggle, as it is not adapted to the test domain. We propose a new approach for domain adaptation (DA), using neuron-level interventions: We modify the representation of each test example in specific neurons, resulting in a counterfactual example from the source domain, which the model is more familiar with. The modified example is then fed back into the model. While most other DA methods are applied during training time, ours is applied during inference only, making it more efficient and applicable. Our experiments show that our method improves performance on unseen domains.

Adir Rahamim, Yonatan Belinkov

Recent work has compared neural network representations via similarity-based analyses to improve model interpretation. The quality of a similarity measure is typically evaluated by its success in assigning a high score to representations that are expected to be matched. However, existing similarity measures perform mediocrely on standard benchmarks. In this work, we develop a new similarity measure, dubbed ContraSim, based on contrastive learning. In contrast to common closed-form similarity measures, ContraSim learns a parameterized measure by using both similar and dissimilar examples. We perform an extensive experimental evaluation of our method, with both language and vision models, on the standard layer prediction benchmark and two new benchmarks that we introduce: the multilingual benchmark and the image-caption benchmark. In all cases, ContraSim achieves much higher accuracy than previous similarity measures, even when presented with challenging examples. Finally, ContraSim is more suitable for the analysis of neural networks, revealing new insights not captured by previous measures.

Omri Koshorek, Niv Granot, Aviv Alloni et al.

Retrieval-Augmented Generation (RAG) has become the dominant approach for answering questions over large corpora. However, current datasets and methods are highly focused on cases where only a small part of the corpus (usually a few paragraphs) is relevant per query, and fail to capture the rich world of aggregative queries. These require gathering information from a large set of documents and reasoning over them. To address this gap, we propose S-RAG, an approach specifically designed for such queries. At ingestion time, S-RAG constructs a structured representation of the corpus; at inference time, it translates natural-language queries into formal queries over said representation. To validate our approach and promote further research in this area, we introduce two new datasets of aggregative queries: HOTELS and WORLD CUP. Experiments with S-RAG on the newly introduced datasets, as well as on a public benchmark, demonstrate that it substantially outperforms both common RAG systems and long-context LLMs.

Adir Rahamim, Naomi Saphra, Sara Kangaslahti et al.

Parameter efficient finetuning methods like low-rank adaptation (LoRA) aim to reduce the computational costs of finetuning pretrained Language Models (LMs). Enabled by these low-rank settings, we propose an even more efficient optimization strategy: Fast Forward, a simple and effective approach to accelerate large segments of training. In a Fast Forward stage, we repeat the most recent optimizer step until the loss stops improving on a tiny validation set. By alternating between regular optimization steps and Fast Forward stages, Fast Forward provides up to an 87\% reduction in FLOPs and up to an 81\% reduction in train time over standard SGD with Adam. We validate Fast Forward by finetuning various models on different tasks and demonstrate that it speeds up training without compromising model performance. Additionally, we analyze when and how to apply Fast Forward.

Gal Kesten-Pomeranz, Yaniv Nikankin, Anja Reusch et al.

Protein sequences are abundant in repeating segments, both as exact copies and as approximate segments with mutations. These repeats are important for protein structure and function, motivating decades of algorithmic work on repeat identification. Recent work has shown that protein language models (PLMs) identify repeats, by examining their behavior in masked-token prediction. To elucidate their internal mechanisms, we investigate how PLMs detect both exact and approximate repeats. We find that the mechanism for approximate repeats functionally subsumes that of exact repeats. We then characterize this mechanism, revealing two main stages: PLMs first build feature representations using both general positional attention heads and biologically specialized components, such as neurons that encode amino-acid similarity. Then, induction heads attend to aligned tokens across repeated segments, promoting the correct answer. Our results reveal how PLMs solve this biological task by combining language-based pattern matching with specialized biological knowledge, thereby establishing a basis for studying more complex evolutionary processes in PLMs.

Aaron Mueller, Jannik Brinkmann, Millicent Li et al.

Interpretability provides a toolset for understanding how and why neural networks behave in certain ways. However, there is little unity in the field: most studies employ ad-hoc evaluations and do not share theoretical foundations, making it difficult to measure progress and compare the pros and cons of different techniques. Furthermore, while mechanistic understanding is frequently discussed, the basic causal units underlying these mechanisms are often not explicitly defined. In this article, we propose a perspective on interpretability research grounded in causal mediation analysis. Specifically, we describe the history and current state of interpretability taxonomized according to the types of causal units (mediators) employed, as well as methods used to search over mediators. We discuss the pros and cons of each mediator, providing insights as to when particular kinds of mediators and search methods are most appropriate. We argue that this framing yields a more cohesive narrative of the field and helps researchers select appropriate methods based on their research objective. Our analysis yields actionable recommendations for future work, including the discovery of new mediators and the development of standardized evaluations tailored to these goals.

Adi Simhi, Jonathan Herzig, Idan Szpektor et al.

Large language models (LLMs) are prone to hallucinations, which sparked a widespread effort to detect and prevent them. Recent work attempts to mitigate hallucinations by intervening in the model's generation, typically computing representative vectors of hallucinations vs. grounded generations, for steering the model's hidden states away from a hallucinatory state. However, common studies employ different setups and do not properly separate different possible causes of hallucinations, making interventions misguided. In this work, we introduce a method for categorizing examples based on the model's prior knowledge, named WACK. We construct WACK benchmarks that support interventions in two settings: open-book and closed-book question answering. Using the benchmarks, we perform an extensive investigation of the effect of different choices for intervention, such as the intervened components, and how often and how strongly to intervene. We find that intervention success varies depending on the component, with the attention blocks performing well and the residual stream proving detrimental to language modeling capabilities. We also show that interventions can benefit from representative vectors collected before, rather than after, a hallucination occurs. Finally, we introduce a new dynamic intervention, which intervenes only if needed, and thus is more robust than standard static interventions. The code is available at https://github.com/technion-cs-nlp/hallucination-mitigation .

Adi Simhi, Jonathan Herzig, Idan Szpektor et al.

Large language models (LLMs) are susceptible to hallucinations -- factually incorrect outputs -- leading to a large body of work on detecting and mitigating such cases. We argue that it is important to distinguish between two types of hallucinations: ones where the model does not hold the correct answer in its parameters, which we term HK-, and ones where the model answers incorrectly despite having the required knowledge, termed HK+. We first find that HK+ hallucinations are prevalent and occur across models and datasets. Then, we demonstrate that distinguishing between these two cases is beneficial for mitigating hallucinations. Importantly, we show that different models hallucinate on different examples, which motivates constructing model-specific hallucination datasets for training detectors. Overall, our findings draw attention to classifying types of hallucinations and provide means to handle them more effectively. The code is available at https://github.com/technion-cs-nlp/hallucination-mitigation .

Opher Lieber, Barak Lenz, Hofit Bata et al.

We present Jamba, a new base large language model based on a novel hybrid Transformer-Mamba mixture-of-experts (MoE) architecture. Specifically, Jamba interleaves blocks of Transformer and Mamba layers, enjoying the benefits of both model families. MoE is added in some of these layers to increase model capacity while keeping active parameter usage manageable. This flexible architecture allows resource- and objective-specific configurations. In the particular configuration we have implemented, we end up with a powerful model that fits in a single 80GB GPU. Built at large scale, Jamba provides high throughput and small memory footprint compared to vanilla Transformers, and at the same time state-of-the-art performance on standard language model benchmarks and long-context evaluations. Remarkably, the model presents strong results for up to 256K tokens context length. We study various architectural decisions, such as how to combine Transformer and Mamba layers, and how to mix experts, and show that some of them are crucial in large scale modeling. We also describe several interesting properties of these architectures which the training and evaluation of Jamba have revealed, and plan to release checkpoints from various ablation runs, to encourage further exploration of this novel architecture. We make the weights of our implementation of Jamba publicly available under a permissive license.

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.

William Rudman, Michal Golovanevsky, Dana Arad et al.

Large vision-language models (VLMs) are highly capable, yet often hallucinate by favoring textual prompts over visual evidence. We study this failure mode in a controlled object-counting setting, where the prompt overstates the number of objects in the image (e.g., asking a model to describe four waterlilies when only three are present). At low object counts, models often correct the overestimation, but as the number of objects increases, they increasingly conform to the prompt regardless of the discrepancy. Through mechanistic analysis of three VLMs, we identify a small set of attention heads whose ablation substantially reduces prompt-induced hallucinations (PIH) by at least 40% without additional training. Across models, PIH-heads mediate prompt copying in model-specific ways. We characterize these differences and show that PIH ablation increases correction toward visual evidence. Our findings offer insights into the internal mechanisms driving prompt-induced hallucinations, revealing model-specific differences in how these behaviors are implemented.

Samuel Marks, Can Rager, Eric J. Michaud et al.

We introduce methods for discovering and applying sparse feature circuits. These are causally implicated subnetworks of human-interpretable features for explaining language model behaviors. Circuits identified in prior work consist of polysemantic and difficult-to-interpret units like attention heads or neurons, rendering them unsuitable for many downstream applications. In contrast, sparse feature circuits enable detailed understanding of unanticipated mechanisms. Because they are based on fine-grained units, sparse feature circuits are useful for downstream tasks: We introduce SHIFT, where we improve the generalization of a classifier by ablating features that a human judges to be task-irrelevant. Finally, we demonstrate an entirely unsupervised and scalable interpretability pipeline by discovering thousands of sparse feature circuits for automatically discovered model behaviors.

Adi Simhi, Itay Itzhak, Fazl Barez et al.

Prior work on large language model (LLM) hallucinations has associated them with model uncertainty or inaccurate knowledge. In this work, we define and investigate a distinct type of hallucination, where a model can consistently answer a question correctly, but a seemingly trivial perturbation, which can happen in real-world settings, causes it to produce a hallucinated response with high certainty. This phenomenon, which we dub CHOKE (Certain Hallucinations Overriding Known Evidence), is particularly concerning in high-stakes domains such as medicine or law, where model certainty is often used as a proxy for reliability. We show that CHOKE examples are consistent across prompts, occur in different models and datasets, and are fundamentally distinct from other hallucinations. This difference leads existing mitigation methods to perform worse on CHOKE examples than on general hallucinations. Finally, we introduce a probing-based mitigation that outperforms existing methods on CHOKE hallucinations. These findings reveal an overlooked aspect of hallucinations, emphasizing the need to understand their origins and improve mitigation strategies to enhance LLM safety. The code is available at https://github.com/technion-cs-nlp/Trust_me_Im_wrong .

Adi Simhi, Jonathan Herzig, Martin Tutek et al.

As large language models (LLMs) evolve from conversational assistants into autonomous agents, evaluating the safety of their actions becomes critical. Prior safety benchmarks have primarily focused on preventing generation of harmful content, such as toxic text. However, they overlook the challenge of agents taking harmful actions when the most effective path to an operational goal conflicts with human safety. To address this gap, we introduce ManagerBench, a benchmark that evaluates LLM decision-making in realistic, human-validated managerial scenarios. Each scenario forces a choice between a pragmatic but harmful action that achieves an operational goal, and a safe action that leads to worse operational performance. A parallel control set, where potential harm is directed only at inanimate objects, measures a model's pragmatism and identifies its tendency to be overly safe. Our findings indicate that the frontier LLMs perform poorly when navigating this safety-pragmatism trade-off. Many consistently choose harmful options to advance their operational goals, while others avoid harm only to become overly safe and ineffective. Critically, we find this misalignment does not stem from an inability to perceive harm, as models' harm assessments align with human judgments, but from flawed prioritization. ManagerBench is a challenging benchmark for a core component of agentic behavior: making safe choices when operational goals and alignment values incentivize conflicting actions. Benchmark & code available at https://github.com/technion-cs-nlp/ManagerBench.

Rom Himelstein, Amit LeVi, Yonatan Belinkov et al.

Padding tokens are widely used in large language models (LLMs) to equalize sequence lengths during batched inference. While they should be fully masked, implementation errors can cause them to influence computation, and the extent of this influence is not well understood. We systematically study this effect across three open-source model families (Llama, Gemma, Qwen), inserting controlled amounts of padding and evaluating outcomes along four axes: activations, generation quality, bias, and safety. Even small amounts of padding shift hidden representations, degrade quality in smaller models, alter bias in unpredictable ways, and weaken safety guardrails. These findings demonstrate that padding is not a harmless detail but a robustness risk that must be carefully handled in deployment.

Zachary Bamberger, Ofek Glick, Chaim Baskin et al.

Language Models (LMs) struggle with linguistic understanding at the discourse level, even though discourse patterns such as coherence, cohesion, and narrative flow are prevalent in their pre-training data. To improve the discourse capabilities of LMs already at the pre-training stage, we introduce DEPTH, an encoder-decoder model that learns latent representations for sentences using a discourse-oriented pre-training objective. DEPTH combines hierarchical sentence representations with two objectives: (1) Sentence Un-Shuffling, and (2) Span-Corruption. Our approach trains the model to represent both sub-word-level and sentence-level dependencies over a pre-training corpora. When trained either from scratch or continuing from a pre-trained T5 checkpoint, DEPTH learns semantic and discourse-level representations faster than T5, outperforming it in span-corruption loss despite the additional sentence-un-shuffling objective. Evaluations on the GLUE, DiscoEval, and NI benchmarks demonstrate DEPTH's ability to quickly learn diverse downstream tasks, which require syntactic, semantic, and discourse capabilities. Our approach extends the discourse capabilities of T5, while minimally impacting other natural language understanding (NLU) capabilities in the resulting LM. We share our codebase for reproducibility: https://github.com/zbambergerNLP/depth.git.

Rabeeh Karimi Mahabadi, Yonatan Belinkov, James Henderson

While large-scale pretrained language models have obtained impressive results when fine-tuned on a wide variety of tasks, they still often suffer from overfitting in low-resource scenarios. Since such models are general-purpose feature extractors, many of these features are inevitably irrelevant for a given target task. We propose to use Variational Information Bottleneck (VIB) to suppress irrelevant features when fine-tuning on low-resource target tasks, and show that our method successfully reduces overfitting. Moreover, we show that our VIB model finds sentence representations that are more robust to biases in natural language inference datasets, and thereby obtains better generalization to out-of-domain datasets. Evaluation on seven low-resource datasets in different tasks shows that our method significantly improves transfer learning in low-resource scenarios, surpassing prior work. Moreover, it improves generalization on 13 out of 15 out-of-domain natural language inference benchmarks. Our code is publicly available in https://github.com/rabeehk/vibert.

Rabeeh Karimi Mahabadi, Yonatan Belinkov, James Henderson

Several recent studies have shown that strong natural language understanding (NLU) models are prone to relying on unwanted dataset biases without learning the underlying task, resulting in models that fail to generalize to out-of-domain datasets and are likely to perform poorly in real-world scenarios. We propose two learning strategies to train neural models, which are more robust to such biases and transfer better to out-of-domain datasets. The biases are specified in terms of one or more bias-only models, which learn to leverage the dataset biases. During training, the bias-only models' predictions are used to adjust the loss of the base model to reduce its reliance on biases by down-weighting the biased examples and focusing the training on the hard examples. We experiment on large-scale natural language inference and fact verification benchmarks, evaluating on out-of-domain datasets that are specifically designed to assess the robustness of models against known biases in the training data. Results show that our debiasing methods greatly improve robustness in all settings and better transfer to other textual entailment datasets. Our code and data are publicly available in \url{https://github.com/rabeehk/robust-nli}.

Nikhil Prakash, Tamar Rott Shaham, Tal Haklay et al.

Fine-tuning on generalized tasks such as instruction following, code generation, and mathematics has been shown to enhance language models' performance on a range of tasks. Nevertheless, explanations of how such fine-tuning influences the internal computations in these models remain elusive. We study how fine-tuning affects the internal mechanisms implemented in language models. As a case study, we explore the property of entity tracking, a crucial facet of language comprehension, where models fine-tuned on mathematics have substantial performance gains. We identify the mechanism that enables entity tracking and show that (i) in both the original model and its fine-tuned versions primarily the same circuit implements entity tracking. In fact, the entity tracking circuit of the original model on the fine-tuned versions performs better than the full original model. (ii) The circuits of all the models implement roughly the same functionality: Entity tracking is performed by tracking the position of the correct entity in both the original model and its fine-tuned versions. (iii) Performance boost in the fine-tuned models is primarily attributed to its improved ability to handle the augmented positional information. To uncover these findings, we employ: Patch Patching, DCM, which automatically detects model components responsible for specific semantics, and CMAP, a new approach for patching activations across models to reveal improved mechanisms. Our findings suggest that fine-tuning enhances, rather than fundamentally alters, the mechanistic operation of the model.

Michael Hanna, Sandro Pezzelle, Yonatan Belinkov

Many recent language model (LM) interpretability studies have adopted the circuits framework, which aims to find the minimal computational subgraph, or circuit, that explains LM behavior on a given task. Most studies determine which edges belong in a LM's circuit by performing causal interventions on each edge independently, but this scales poorly with model size. Edge attribution patching (EAP), gradient-based approximation to interventions, has emerged as a scalable but imperfect solution to this problem. In this paper, we introduce a new method - EAP with integrated gradients (EAP-IG) - that aims to better maintain a core property of circuits: faithfulness. A circuit is faithful if all model edges outside the circuit can be ablated without changing the model's performance on the task; faithfulness is what justifies studying circuits, rather than the full model. Our experiments demonstrate that circuits found using EAP are less faithful than those found using EAP-IG, even though both have high node overlap with circuits found previously using causal interventions. We conclude more generally that when using circuits to compare the mechanisms models use to solve tasks, faithfulness, not overlap, is what should be measured.

Yaniv Nikankin, Anja Reusch, Aaron Mueller et al.

Do large language models (LLMs) solve reasoning tasks by learning robust generalizable algorithms, or do they memorize training data? To investigate this question, we use arithmetic reasoning as a representative task. Using causal analysis, we identify a subset of the model (a circuit) that explains most of the model's behavior for basic arithmetic logic and examine its functionality. By zooming in on the level of individual circuit neurons, we discover a sparse set of important neurons that implement simple heuristics. Each heuristic identifies a numerical input pattern and outputs corresponding answers. We hypothesize that the combination of these heuristic neurons is the mechanism used to produce correct arithmetic answers. To test this, we categorize each neuron into several heuristic types-such as neurons that activate when an operand falls within a certain range-and find that the unordered combination of these heuristic types is the mechanism that explains most of the model's accuracy on arithmetic prompts. Finally, we demonstrate that this mechanism appears as the main source of arithmetic accuracy early in training. Overall, our experimental results across several LLMs show that LLMs perform arithmetic using neither robust algorithms nor memorization; rather, they rely on a "bag of heuristics".

Shahar Katz, Yonatan Belinkov, Mor Geva et al. · deepmind

Understanding how Transformer-based Language Models (LMs) learn and recall information is a key goal of the deep learning community. Recent interpretability methods project weights and hidden states obtained from the forward pass to the models' vocabularies, helping to uncover how information flows within LMs. In this work, we extend this methodology to LMs' backward pass and gradients. We first prove that a gradient matrix can be cast as a low-rank linear combination of its forward and backward passes' inputs. We then develop methods to project these gradients into vocabulary items and explore the mechanics of how new information is stored in the LMs' neurons.

Michael Toker, Hadas Orgad, Mor Ventura et al.

Text-to-image diffusion models (T2I) use a latent representation of a text prompt to guide the image generation process. However, the process by which the encoder produces the text representation is unknown. We propose the Diffusion Lens, a method for analyzing the text encoder of T2I models by generating images from its intermediate representations. Using the Diffusion Lens, we perform an extensive analysis of two recent T2I models. Exploring compound prompts, we find that complex scenes describing multiple objects are composed progressively and more slowly compared to simple scenes; Exploring knowledge retrieval, we find that representation of uncommon concepts requires further computation compared to common concepts, and that knowledge retrieval is gradual across layers. Overall, our findings provide valuable insights into the text encoder component in T2I pipelines.

Zeping Yu, Yonatan Belinkov, Sophia Ananiadou

We investigate how large language models perform latent multi-hop reasoning in prompts like "Wolfgang Amadeus Mozart's mother's spouse is". To analyze this process, we introduce logit flow, an interpretability method that traces how logits propagate across layers and positions toward the final prediction. Using logit flow, we identify four distinct stages in single-hop knowledge prediction: (A) entity subject enrichment, (B) entity attribute extraction, (C) relation subject enrichment, and (D) relation attribute extraction. Extending this analysis to multi-hop reasoning, we find that failures often stem from the relation attribute extraction stage, where conflicting logits reduce prediction accuracy. To address this, we propose back attention, a novel mechanism that enables lower layers to leverage higher-layer hidden states from different positions during attention computation. With back attention, a 1-layer transformer achieves the performance of a 2-layer transformer. Applied to four LLMs, back attention improves accuracy on five reasoning datasets, demonstrating its effectiveness in enhancing latent multi-hop reasoning ability.

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.

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.

Michal Shur-Ofry, Bar Horowitz-Amsalem, Adir Rahamim et al.

How diverse are the outputs of large language models when diversity is desired? We examine the diversity of responses of various models to questions with multiple possible answers, comparing them with human responses. Our findings suggest that models' outputs are highly concentrated, reflecting a narrow, mainstream 'worldview', in comparison to humans, whose responses exhibit a much longer-tail. We examine three ways to increase models' output diversity: 1) increasing generation randomness via temperature sampling; 2) prompting models to answer from diverse perspectives; 3) aggregating outputs from several models. A combination of these measures significantly increases models' output diversity, reaching that of humans. We discuss implications of these findings for AI policy that wishes to preserve cultural diversity, an essential building block of a democratic social fabric.

Alon Mor, Yonatan Belinkov, Benny Kimelfeld

Local explanation methods highlight the input tokens that have a considerable impact on the outcome of classifying the document at hand. For example, the Anchor algorithm applies a statistical analysis of the sensitivity of the classifier to changes in the token. Aggregating local explanations over a dataset provides a global explanation of the model. Such aggregation aims to detect words with the most impact, giving valuable insights about the model, like what it has learned in training and which adversarial examples expose its weaknesses. However, standard aggregation methods bear a high computational cost: a naïve implementation applies a costly algorithm to each token of each document, and hence, it is infeasible for a simple user running in the scope of a short analysis session. % We devise techniques for accelerating the global aggregation of the Anchor algorithm. Specifically, our goal is to compute a set of top-$k$ words with the highest global impact according to different aggregation functions. Some of our techniques are lossless and some are lossy. We show that for a very mild loss of quality, we are able to accelerate the computation by up to 30$\times$, reducing the computation from hours to minutes. We also devise and study a probabilistic model that accounts for noise in the Anchor algorithm and diminishes the bias toward words that are frequent yet low in impact.

Yaniv Nikankin, Dana Arad, Yossi Gandelsman et al.

Vision-Language models (VLMs) show impressive abilities to answer questions on visual inputs (e.g., counting objects in an image), yet demonstrate higher accuracies when performing an analogous task on text (e.g., counting words in a text). We investigate this accuracy gap by identifying and comparing the \textit{circuits} - the task-specific computational sub-graphs - in different modalities. We show that while circuits are largely disjoint between modalities, they implement relatively similar functionalities: the differences lie primarily in processing modality-specific data positions (an image or a text sequence). Zooming in on the image data representations, we observe they become aligned with the higher-performing analogous textual representations only towards later layers, too late in processing to effectively influence subsequent positions. To overcome this, we patch the representations of visual data tokens from later layers back into earlier layers. In experiments with multiple tasks and models, this simple intervention closes a third of the performance gap between the modalities, on average. Our analysis sheds light on the multi-modal performance gap in VLMs and suggests a training-free approach for reducing it.

Martin Tutek, Fateme Hashemi Chaleshtori, Ana Marasović et al.

When prompted to think step-by-step, language models (LMs) produce a chain of thought (CoT), a sequence of reasoning steps that the model supposedly used to produce its prediction. Despite much work on CoT prompting, it is unclear if reasoning verbalized in a CoT is faithful to the models' parametric beliefs. We introduce a framework for measuring parametric faithfulness of generated reasoning, and propose Faithfulness by Unlearning Reasoning steps (FUR), an instance of this framework. FUR erases information contained in reasoning steps from model parameters, and measures faithfulness as the resulting effect on the model's prediction. Our experiments with four LMs and five multi-hop multi-choice question answering (MCQA) datasets show that FUR is frequently able to precisely change the underlying models' prediction for a given instance by unlearning key steps, indicating when a CoT is parametrically faithful. Further analysis shows that CoTs generated by models post-unlearning support different answers, hinting at a deeper effect of unlearning.