LLM Interpretability

Large language models appear to develop internal representations of emotion -- "emotion circuits," "emotion neurons," and structured emotional manifolds have been reported across multiple model families. But every study making these claims uses stimuli signalled by explicit emotion keywords, leaving a fundamental question unanswered: do these circuits detect genuine emotional meaning, or do they detect the word "devastated"? We present the first clinical validity test of emotion circuit claims using mechanistic interpretability methods grounded in clinical psychology -- clinical vignettes that evoke emotions through situational and behavioural cues alone, emotion keywords removed. Across six models (Llama-3.2-1B, Llama-3-8B, Gemma-2-9B; base and instruct variants), we apply four convergent mechanistic interpretability methods -- linear probing, causal activation patching, knockout experiments, and representational geometry -- and discover two dissociable emotion processing mechanisms. Affect reception -- detecting emotionally significant content -- operates with near-perfect accuracy (AUROC 1.000), consistent with early-layer saturation, and replicates across all six models. Emotion categorization -- mapping affect to specific emotion labels -- is partially keyword-dependent, dropping 1-7% without keywords and improving with scale. Causal activation patching confirms keyword-rich and keyword-free stimuli share representational space, transferring affective salience rather than emotion-category identity. These findings falsify the keyword-spotting hypothesis, establish a novel mechanistic dissociation, and introduce clinical stimulus methodology as a rigorous standard for testing emotion processing claims in large language models -- with direct implications for AI safety evaluation and alignment. All stimuli, code, and data are released for replication.

Trained Transformers have been shown to compute abstract features that appear redundant for predicting the immediate next token. We identify which components of the gradient signal from the next-token prediction objective give rise to this phenomenon, and we propose a method to estimate the influence of those components on the emergence of specific features. After validating our approach on toy tasks, we use it to interpret the origins of the world model in OthelloGPT and syntactic features in a small language model. Finally, we apply our framework to a pretrained LLM, showing that features with extremely high or low influence on future tokens tend to be related to formal reasoning domains such as code. Overall, our work takes a step toward understanding hidden features of Transformers through the lens of their development during training.

A language model can give the wrong answer even when the correct answer is decodable from its intermediate states. To study this gap between decodability and selection, we distinguish read from write at the first answer token. Read asks whether the gold token can be decoded from intermediate residual states under same-relation decoy controls. Write asks whether the final readout ranks that token first among content tokens. Under three different readers, with a randomized-label control, a substantial fraction of failures remain readable while another content token is selected. We explain this through the selection margin at the final readout, the difference between the answer logit and the logit of its strongest alternative, which is answer support minus alternative support, and can also be split into a context-averaged baseline linked to token frequency and an item-specific term. Setting the answer support to the level typical of successful generations is sufficient to recover first-token selection for the majority of failures in most of the models we study; the original alternative remains ahead in most remaining failures under this edit, and this outcome follows directly from the readout geometry. Removing the frequency direction alone shifts selection but rarely recovers the answer. Prompt variants of the same fact that succeed supply support that transfers to failing variants through the residual stream and through late MLP outputs, with less consistent effects through late attention. First-token recovery leaves most full answers wrong, which limits the recovery achieved by these edits and separates three things that are easily conflated, decodability, recoverability, and generation.

Despite the strong performance of large audio language models (LALMs) in various tasks, exactly how and where they integrate acoustic features with textual context remains unclear. We adapt causal tracing to investigate the internal information flow of LALMs during audio comprehension. By conducting layer-wise and token-wise analyses across DeSTA, Qwen, and Voxtral, we evaluate the causal effects of individual hidden states. Layer-wise analysis identifies different fusion strategies, from progressive integration in DeSTA to abrupt late-stage fusion in Qwen. Token-wise analysis shows that the final sequence token acts as an informational bottleneck where the network decisively retrieves relevant information from the audio. We also observe an attention-like query mechanism at intermediate token positions that triggers the model to pull task-relevant audio context. These findings provide a clear characterization of when and where multi-modal integration occurs within LALMs.

Large language models are often viewed as high-dimensional nonlinear systems and treated as black boxes. Here, we show that transformer depth dynamics admit accurate low-order linear surrogates within context. Across tasks including toxicity, irony, hate speech and sentiment, a 32-dimensional linear surrogate reproduces the layerwise sensitivity profile of GPT-2-large with near-perfect agreement, capturing how the final output shifts under additive injections at each layer. We then uncover a surprising scaling principle: for a fixed-order linear surrogate, agreement with the full model improves monotonically with model size across the GPT-2 family. This linear surrogate also enables principled multi-layer interventions that require less energy than standard heuristic schedules when applied to the full model. Together, our results reveal that as language models scale, low-order linear depth dynamics emerge within contexts, offering a systems-theoretic foundation for analyzing and controlling them.

Large language models (LLMs) trained with canonical tokenization exhibit surprising robustness to non-canonical inputs such as character-level tokenization, yet the mechanisms underlying this robustness remain unclear. We study this phenomenon through mechanistic interpretability and identify a core process we term word recovery. We first introduce a decoding-based method to detect word recovery, showing that hidden states reconstruct canonical word-level token identities from character-level inputs. We then provide causal evidence by removing the corresponding subspace from hidden states, which consistently degrades downstream task performance. Finally, we conduct a fine-grained attention analysis and show that in-group attention among characters belonging to the same canonical token is critical for word recovery: masking such attention in early layers substantially reduces both recovery scores and task performance. Together, our findings provide a mechanistic explanation for tokenization robustness and identify word recovery as a key mechanism enabling LLMs to process character-level inputs.

Large language models (LLMs) tend to externalize their reasoning in their chain of thought, making the chain of thought a good target for monitoring. This is partially an inherent feature of the Transformer architecture: sufficiently long serial cognition must pass through the chain of thought (Korbak et al., 2025). We formalize this argument through the notion of opaque serial depth, given by the length of the longest computation that can be done without the use of interpretable intermediate steps like chain of thought. Given this formalization, we compute numeric upper bounds on the opaque serial depth of Gemma 3 models, as well as asymptotic results for additional architectures beyond standard LLMs. We also open-source an automated method that can calculate upper bounds on the opaque serial depth of arbitrary neural networks, and use it to demonstrate that Mixture-of-Experts models likely have lower depth than dense models. Overall, our results suggest that opaque serial depth is a useful tool for understanding the potential for models to do significant reasoning that is not externalized.

Large language models are routinely deployed on text that varies widely in emotional tone, yet their reasoning behavior is typically evaluated without accounting for emotion as a source of representational variation. Prior work has largely treated emotion as a prediction target, for example in sentiment analysis or emotion classification. In contrast, we study emotion as a latent factor that shapes how models attend to and reason over text. We analyze how emotional tone systematically alters attention geometry in transformer models, showing that metrics such as locality, center-of-mass distance, and entropy vary across emotions and correlate with downstream question-answering performance. To facilitate controlled study of these effects, we introduce Affect-Uniform ReAding QA (AURA-QA), a question-answering dataset with emotionally balanced, human-authored context passages. Finally, an emotional regularization framework is proposed that constrains emotion-conditioned representational drift during training. Experiments across multiple QA benchmarks demonstrate that this approach improves reading comprehension in both emotionally-varying and non-emotionally varying datasets, yielding consistent gains under distribution shift and in-domain improvements on several benchmarks.

Do the features learned by Sparse Autoencoders (SAEs) represent abstract meaning, or are they tied to how text is written? We investigate this question using Serbian digraphia as a controlled testbed: Serbian is written interchangeably in Latin and Cyrillic scripts with a near-perfect character mapping between them, enabling us to vary orthography while holding meaning exactly constant. Crucially, these scripts are tokenized completely differently, sharing no tokens whatsoever. Analyzing SAE feature activations across the Gemma model family (270M-27B parameters), we find that identical sentences in different Serbian scripts activate highly overlapping features, far exceeding random baselines. Strikingly, changing script causes less representational divergence than paraphrasing within the same script, suggesting SAE features prioritize meaning over orthographic form. Cross-script cross-paraphrase comparisons provide evidence against memorization, as these combinations rarely co-occur in training data yet still exhibit substantial feature overlap. This script invariance strengthens with model scale. Taken together, our findings suggest that SAE features can capture semantics at a level of abstraction above surface tokenization, and we propose Serbian digraphia as a general evaluation paradigm for probing the abstractness of learned representations.

Transformer models are now a cornerstone in natural language processing. Yet, explaining their decisions remains a challenge. It was shown recently that the same model trained on the same data with a different randomness can lead to very different explanations. In this paper, we investigate how the (syntactic) context, the classes to be learned and the tasks influence this explanations' sensitivity to randomness. We show that they all have statistically significant impact: smallest for the (syntactic) context, medium for the classes and largest for the tasks.

Sparse autoencoders (SAEs) trained on large language model activations output thousands of features that enable mapping to human-interpretable concepts. The current practice for analyzing these features primarily relies on inspecting top-activating examples, manually browsing individual features, or performing semantic search on interested concepts, which makes exploratory discovery of concepts difficult at scale. In this paper, we present Concept Explorer, a scalable interactive system for post-hoc exploration of SAE features that organizes concept explanations using hierarchical neighborhood embeddings. Our approach constructs a multi-resolution manifold over SAE feature embeddings and enables progressive navigation from coarse concept clusters to fine-grained neighborhoods, supporting discovery, comparison, and relationship analysis among concepts. We demonstrate the utility of Concept Explorer on SAE features extracted from SmolLM2, where it reveals coherent high-level structure, meaningful subclusters, and distinctive rare concepts that are hard to identify with existing workflows.

Explicit Chain-of-Thought improves the reasoning performance of large language models but often incurs high inference cost due to verbose token-level traces. While recent approaches reduce this overhead via concise prompting or step pruning, they largely truncate what the model says rather than internalize what the model thinks. Latent reasoning offers a promising alternative by performing computation in the hidden space, yet prior methods face two critical challenges. Many existing approaches rely on rigid point-to-point alignment, forcing a latent token to approximate the final representation of a reasoning step, which can be insufficient to capture the dense, variable-length semantics of an entire reasoning segment. Furthermore, these methods often suffer from a lack of interpretability: latent states are commonly produced by unconstrained optimization or embedding mixing, yielding vectors that are difficult to decode or audit under the pretrained language head. We propose SPOT, a flexible framework that compresses explicit CoT into compact latent pause tokens without enforcing a fixed response template. At the core of SPOT is Span-level Semantic Alignment, a Sinkhorn optimal-transport objective that softly matches each pause token to the semantics of an entire reasoning segment, overcoming the rigidity of step-end alignment. To further improve interpretability, SPOT introduces a Frozen-Head Decoding Constraint that keeps latent states directly decodable as token distributions under the frozen pretrained LM head, enabling readable keyword interpretations of latent thoughts. Experiments on reasoning benchmarks demonstrate that SPOT improves accuracy by 2.3 points on average while reducing generated tokens by 37.5% and provides faithful semantic interpretations of the latent reasoning process.

Trust plays a pivotal role in enabling effective cooperation, reducing uncertainty, and guiding decision-making in both human interactions and multi-agent systems. Although it is significant, there is limited understanding of how large language models (LLMs) internally conceptualize and reason about trust. This work presents a white-box analysis of trust representation in EleutherAI/gpt-j-6B, using contrastive prompting to generate embedding vectors within the activation space of the LLM for diadic trust and related interpersonal relationship attributes. We first identified trust-related concepts from five established human trust models. We then determined a threshold for significant conceptual alignment by computing pairwise cosine similarities across 60 general emotional concepts. Then we measured the cosine similarities between the LLM's internal representation of trust and the derived trust-related concepts. Our results show that the internal trust representation of EleutherAI/gpt-j-6B aligns most closely with the Castelfranchi socio-cognitive model, followed by the Marsh Model. These findings indicate that LLMs encode socio-cognitive constructs in their activation space in ways that support meaningful comparative analyses, inform theories of social cognition, and support the design of human-AI collaborative systems.

Large language models (LLMs) have demonstrated remarkable performance in text generation and knowledge-intensive question answering. Nevertheless, they are prone to producing hallucinated content, which severely undermines their reliability in high-stakes application domains. Existing hallucination attribution approaches, based on either external knowledge retrieval or internal model mechanisms, primarily focus on semantic similarity matching or representation-level discrimination. As a result, they have difficulty establishing structured correspondences at the span level between hallucination types, underlying error generation mechanisms, and external factual evidence, thereby limiting the interpretability of hallucinated fragments and the traceability of supporting or opposing evidence. To address these limitations, we propose HART, a fine-grained hallucination attribution and evidence retrieval framework for large language models. HART formalizes hallucination tracing as a structured modeling task comprising four stages: span localization, mechanism attribution, evidence retrieval, and causal tracing. Based upon this formulation, we develop the first structured dataset tailored for hallucination tracing, in which hallucination types, error mechanisms, and sets of counterfactual evidence are jointly annotated to enable causal-level interpretability evaluation. Experimental results on the proposed dataset demonstrate that HART substantially outperforms strong retrieval baselines, including BM25 and DPR, validating the effectiveness and generalization capability of the proposed tracing paradigm for hallucination analysis and evidence alignment.

A growing literature shows that variables can be linearly decoded from the activations of large language models (LLMs). These range from properties of the world, such as the locations of cities and the lifetimes of historical figures, to emotions and pain. Such findings are often taken as evidence that language models go beyond surface text statistics and form internal models of the world. We show that static word embeddings (fixed, context-insensitive representations learned from corpus statistics) of the same or matched stimuli support much of the same decoding. Across four published cases (place, time, pain and emotion), static vectors predict coordinates and year of death (R^2 = 0.42-0.59), separate pain from matched control sentences (held-out AUC 0.85-0.88), and classify twelve emotions in stories written to avoid naming them (AUC 0.84-0.88). Because static embeddings assign each word a single, context-independent vector, these results are a lower bound on what word associations alone can support. The LLMs retain clear advantages on representational tests, and causal and behavioral findings remain outside the scope of the baseline. On the original authors' entities, where we reproduce their Llama-2 results, the transformer's advantage lies mostly in placing historical figures in the right century and places in the right country, coarse sorting that richer word associations would be expected to improve; within those groups every representation orders items poorly. Static vectors for disambiguated Wikipedia entities, which carry the associations of a particular place or person rather than of the words in its name, close most of the remaining gap, matching Pythia-2.8B on coordinates and Llama-2-7B on year of death. These results indicate that decodability alone cannot distinguish a representation of a property from information already available in fixed distributional associations.

Sparse autoencoders (SAEs) are widely used to extract human-interpretable features from neural network activations, but their learned features can vary substantially across random seeds and training choices. To improve stability, we studied weight regularization by adding L1 or L2 penalties on encoder and decoder weights, and evaluate how regularization interacts with common SAE training defaults. On MNIST, we observe that L2 weight regularization produces a core of highly aligned features and, when combined with tied initialization and unit-norm decoder constraints, it dramatically increases cross-seed feature consistency. For TopK SAEs trained on language model activations (Pythia-70M-deduped), adding a small L2 weight penalty increased the fraction of features shared across three random seeds and roughly doubles steering success rates, while leaving the mean of automated interpretability scores essentially unchanged. Finally, in the regularized setting, activation steering success becomes better predicted by auto-interpretability scores, suggesting that regularization can align text-based feature explanations with functional controllability.

In-context learning enables large language models to perform novel tasks through few-shot demonstrations. However, demonstrations per se can naturally contain noise and conflicting examples, making this capability vulnerable. To understand how models process such conflicts, we study demonstration-dependent tasks requiring models to infer underlying patterns, a process we characterize as rule inference. We find that models suffer substantial performance degradation from a single demonstration with corrupted rule. This systematic misleading behavior motivates our investigation of how models process conflicting evidence internally. Using linear probes and logit lens analysis, we discover that under corruption models encode both correct and incorrect rules in intermediate layers but develop prediction confidence only in late layers, revealing a two-phase computational structure. We then identify attention heads for each phase underlying the reasoning failures: Vulnerability Heads in early-to-middle layers exhibit positional attention bias with high sensitivity to corruption, while Susceptible Heads in late layers significantly reduce support for correct predictions when exposed to the corrupted evidence. Targeted ablation validates our findings, with masking a small number of identified heads improving performance by over 10%.

Large language models (LLMs) have shown strong empirical performance across pharmacology and drug discovery tasks, yet the internal mechanisms by which they encode pharmacological knowledge remain poorly understood. In this work, we investigate how drug-group semantics are represented and retrieved within Llama-based biomedical language models using causal and probing-based interpretability methods. We apply activation patching to localize where drug-group information is stored across model layers and token positions, and complement this analysis with linear probes trained on token-level and sum-pooled activations. Our results demonstrate that early layers play a key role in encoding drug-group knowledge, with the strongest causal effects arising from intermediate tokens within the drug-group span rather than the final drug-group token. Linear probing further reveals that pharmacological semantics are distributed across tokens and are already present in the embedding space, with token-level probes performing near chance while sum-pooled representations achieve maximal accuracy. Together, these findings suggest that drug-group semantics in LLMs are not localized to single tokens but instead arise from distributed representations. This study provides the first systematic mechanistic analysis of pharmacological knowledge in LLMs, offering insights into how biomedical semantics are encoded in large language models.

Large Language Models (LLMs) have achieved strong complex reasoning capabilities through Chain-of-Thought (CoT) reasoning. However, their reasoning patterns remain too complicated to analyze. While Sparse Autoencoders (SAEs) have emerged as a powerful tool for interpretability, existing approaches predominantly operate at the token level, creating a granularity mismatch when capturing more critical step-level information, such as reasoning direction and semantic transitions. In this work, we propose step-level sparse autoencoder (SSAE), which serves as an analytical tool to disentangle different aspects of LLMs' reasoning steps into sparse features. Specifically, by precisely controlling the sparsity of a step feature conditioned on its context, we form an information bottleneck in step reconstruction, which splits incremental information from background information and disentangles it into several sparsely activated dimensions. Experiments on multiple base models and reasoning tasks show the effectiveness of the extracted features. By linear probing, we can easily predict surface-level information, such as generation length and first token distribution, as well as more complicated properties, such as the correctness and logicality of the step. These observations indicate that LLMs should already at least partly know about these properties during generation, which provides the foundation for the self-verification ability of LLMs. Our code is available at https://github.com/Miaow-Lab/SSAE.

In recent years, pre-trained large language models have achieved remarkable success across diverse tasks. Besides the pivotal role of self-supervised pre-training, their effectiveness in downstream applications also depends critically on the post-training process, which adapts models to task-specific data and objectives. However, this process inevitably introduces model shifts that can influence performance in different domains, and how such shifts transfer remains poorly understood. To open up the black box, we propose the SAE-based Transferability Score (STS), a new metric that leverages sparse autoencoders (SAEs) to forecast post-training transferability. Taking supervised fine-tuning as an example, STS identifies shifted dimensions in SAE representations and calculates their correlations with downstream domains, enabling reliable estimation of transferability before fine-tuning. Extensive experiments across multiple models and domains show that STS accurately predicts the transferability of supervised fine-tuning, achieving Pearson correlation coefficients above 0.7 with actual performance changes. Beyond this, we take an initial step toward extending STS to reinforcement learning. We believe that STS can serve as an {\color{black} interpretable} tool for guiding post-training strategies in LLMs. Code is available at https://github.com/PKU-ML/STS.

Both humans and Large Language Models (LLMs) store a vast repository of semantic memories. In humans, efficient and strategic access to this memory store is a critical foundation for a variety of cognitive functions. Such access has long been a focus of psychology and the computational mechanisms behind it are now well characterized. Much of this understanding has been gleaned from a widely-used neuropsychological and cognitive science assessment called the Semantic Fluency Task (SFT), which requires the generation of as many semantically constrained concepts as possible. Our goal is to apply mechanistic interpretability techniques to bring greater rigor to the study of semantic memory foraging in LLMs. To this end, we present preliminary results examining SFT as a case study. A central focus is on convergent and divergent patterns of generative memory search, which in humans play complementary strategic roles in efficient memory foraging. We show that these same behavioral signatures, critical to human performance on the SFT, also emerge as identifiable patterns in LLMs across distinct layers. Potentially, this analysis provides new insights into how LLMs may be adapted into closer cognitive alignment with humans, or alternatively, guided toward productive cognitive disalignment to enhance complementary strengths in human-AI interaction.

As chain of thought (CoT) has become central to scaling reasoning capabilities in large language models (LLMs), it has also emerged as a promising tool for interpretability, suggesting the opportunity to understand model decisions through verbalized reasoning. However, the utility of CoT toward interpretability depends upon its faithfulness---whether the model's stated reasoning reflects the underlying decision process. We provide mechanistic evidence that instruction-tuned models often determine their answer before generating CoT. Training linear probes on residual stream activations at the last token before CoT, we can predict the model's final answer with >0.9 AUC on most tasks. We find that these directions are not only predictive, but also causal: steering activations along the probe direction often flips model answers, with flip rates substantially exceeding norm-matched orthogonal baselines across most model-dataset pairs. When steering induces incorrect answers, we observe two distinct failure modes: confabulation (fabricating false premises) and non-entailment (stating correct premises but drawing unsupported conclusions). While post-hoc reasoning may be instrumentally useful when the model has a correct pre-CoT belief, these failure modes suggest it can result in undesirable behaviors when reasoning from a false belief.

Existing explainability methods for Large Language Models (LLMs) typically treat hidden states as static points in activation space, assuming that correct and incorrect inferences can be separated using representations from an individual layer. However, these activations are saturated with polysemantic features, leading to linear probes learning surface-level lexical patterns rather than underlying reasoning structures. We introduce Truth as a Trajectory (TaT), which models the transformer inference as an unfolded trajectory of iterative refinements, shifting analysis from static activations to layer-wise geometric displacement. By analyzing displacement of representations across layers, TaT uncovers geometric invariants that distinguish valid reasoning from spurious behavior. We evaluate TaT across dense and Mixture-of-Experts (MoE) architectures on benchmarks spanning commonsense reasoning, question answering, and toxicity detection. Without access to the activations themselves and using only changes in activations across layers, we show that TaT effectively mitigates reliance on static lexical confounds, outperforming conventional probing, and establishes trajectory analysis as a complementary perspective on LLM explainability.

We develop a discrete gauge-theoretic framework for superposition in large language models (LLMs) that replaces the single-global-dictionary premise with a sheaf-theoretic atlas of local semantic charts. Contexts are clustered into a stratified context complex; each chart carries a local feature space and a local information-geometric metric (Fisher/Gauss-Newton) identifying predictively consequential feature interactions. This yields a Fisher-weighted interference energy and three measurable obstructions to global interpretability: (O1) local jamming (active load exceeds Fisher bandwidth), (O2) proxy shearing (mismatch between geometric transport and a fixed correspondence proxy), and (O3) nontrivial holonomy (path-dependent transport around loops). We prove and instantiate four results on a frozen open LLM (Llama-3.2-3B Instruct) using WikiText-103, a C4-derived English web-text subset, and the-stack-smol. (A) After constructive gauge fixing on a spanning tree, each chord residual equals the holonomy of its fundamental cycle, making holonomy computable and gauge-invariant. (B) Shearing lower-bounds a data-dependent transfer mismatch energy, turning $D_{\mathrm{shear}}$ into an unavoidable failure bound. (C) We obtain non-vacuous certified jamming/interference bounds with high coverage and zero violations across seeds and hyperparameters. (D) Bootstrap and sample-size experiments show stable estimation of $D_{\mathrm{shear}}$ and $D_{\mathrm{hol}}$, with improved concentration on well-conditioned subsystems.

Large language models (LLMs) are increasingly used for high-stakes decision-making, yet existing approaches struggle to reconcile scalability, interpretability, and reproducibility. Black-box models obscure their reasoning, while recent LLM-based rule systems rely on per-sample evaluation, causing costs to scale with dataset size and introducing stochastic, hallucination-prone outputs. We propose reframing LLMs as code generators rather than per-instance evaluators. A single LLM call generates executable, human-readable decision logic that runs deterministically over structured data, eliminating per-sample LLM queries while enabling reproducible and auditable predictions. We combine code generation with automated statistical validation using precision lift, binomial significance testing, and coverage filtering, and apply cluster-based gap analysis to iteratively refine decision logic without human annotation. We instantiate this framework in venture capital founder screening, a rare-event prediction task with strong interpretability requirements. On VCBench, a benchmark of 4,500 founders with a 9% base success rate, our approach achieves 37.5% precision and an F0.5 score of 25.0%, outperforming GPT-4o (at 30.0% precision and an F0.5 score of 25.7%) while maintaining full interpretability. Each prediction traces to executable rules over human-readable attributes, demonstrating verifiable and interpretable LLM-based decision-making in practice.

We present gradiend, an open-source Python package that operationalizes the GRADIEND method for learning feature directions from factual-counterfactual MLM and CLM gradients in language models. The package provides a unified workflow for feature-related data creation, training, evaluation, visualization, persistent model rewriting via controlled weight updates, and multi-feature comparison. We demonstrate gradiend through an English pronoun running example, a semantic sentiment use case that evaluates lexical generalization to held-out target words, and a large-scale feature comparison.

We present GLUScope, an open-source resource for analyzing neurons in Transformer-based language models (LMs), intended for interpretability researchers. We focus on the now-standard GLU-based activation functions such as SwiGLU, used in virtually all recent open-weights LMs. This introduces a new challenge: understanding positive activations is not enough. Instead, both the gate and the in activation of a neuron can be positive or negative, leading to four different possible sign combinations (activation quadrants) that in some cases have quite different functionalities (e.g. promoting vs. suppressing a specific set of tokens). Accordingly, for any neuron, our tool shows text examples for each of the four sign combinations, and indicates how often each combination occurs. Our resource can lead to novel insights: in particular, thanks to analyzing all four activation quadrants, it can surface interpretable behavior that other tools would miss. We also release the code used to create the resource, thus enabling researchers to build on it by adding more models, text corpora, or neuron pages.

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.

Large language model (LLM) hallucinations, meaning fluent but factually incorrect generations, fall into two types: faithfulness violations, where the model misuses provided context, and factuality violations, where answers reflect errors in internal knowledge. Proper mitigation depends on knowing which source drives each answer. We study contributive attribution, i.e. the classification of the dominant knowledge source behind each output, and show that a simple linear probe trained on hidden representations can reliably identify it. We introduce AttriWiki, a self-supervised pipeline that automatically generates labelled training data by prompting models to recall withheld entities from memory or read them from context without relying on knowledge conflicts. Probes trained on AttriWiki achieve up to 0.96 Macro-$F_1$ on Llama-3.1-8B, Mistral-7B, and Qwen-7B, transfer to SQuAD and WebQuestions with 0.94-0.99 Macro-$F_1$, and generalise zero-shot to Tighidet et al. (2024)'s benchmark, outperforming their probe on conflicting settings without retraining. Furthermore, attribution mismatches raise error rates by up to 70%, though correct attribution does not guarantee correct answers, pointing to the need for broader detection frameworks.

State-space models (SSMs) have emerged as an efficient strategy for building powerful language models, avoiding the quadratic complexity of computing attention in transformers. Despite their promise, the interpretability and steerability of modern SSMs remain relatively underexplored. We take a major step in this direction by identifying activation subspace bottlenecks in the Mamba family of SSM models using tools from mechanistic interpretability. We then introduce a test-time steering intervention that simply multiplies the activations of the identified bottlenecks by a scalar. Across 7 SSMs and 6 diverse benchmarks, this intervention improves performance by an average of 8.27%, without requiring any task-specific tuning. Finally, we validate that the identified bottlenecks are indeed hindering performance by modifying them to yield an architecture we call Stable-Mamba, which achieves long-context performance gains when retrained from scratch.