Mechanistic Interpretability

Overview

The field seeks to explain how a model’s internal computations produce particular outputs, rather than relying only on input-output behavior. Researchers use sparse autoencoders and transcoders to identify features, and probes, attribution graphs, activation patching, ablation, and steering to trace or test their roles. Studies examine language-model tasks such as factual recall, reasoning, and refusal, as well as vision, audio, and scientific models. A central challenge is that a signal can be easy to decode without being causally used, while circuit explanations and interventions can vary across prompts, models, and analysis choices.

Latest papers

427 papers58 papers in the last four weeks, up 23% on the four weeks before. 0.4% of all new papers.

Oct 29, 2026

Discovering Interpretable Acoustic Features in Autoregressive Voice-Cloning TTS Models

Voice-cloning TTS models synthesize an unseen speaker's voice from brief speech prompts, but it remains unclear which acoustic properties they represent and use. We introduce a feature-based interpretability framework for autoregressive voice-cloning TTS models, analyzing features from sparse autoencoders (SAEs), PCA, and random directions via sparse probing, steering, and quantitative auto-interpretability. We focus on Qwen3-TTS and extend probing and steering to 55 additional models to test generalization across architectures. These models encode concepts like pitch, energy, noise, and accent, but sparse prediction and steering are effective only for some concepts. Strong probing does not always translate to steering or vice-versa. SAEs are useful for sparse probing of human-labelled concepts and novel-concept discovery, while PCA is competitive for dense probing and steering. Auto-interpretability finds 684684 automatically validated SAE features in Qwen3-TTS covering speaker clusters, speaking style, and recording conditions beyond known concepts. Overall, we perform a comprehensive interpretability study of autoregressive voice-cloning TTS models.
Oct 28, 2026

Can LLMs Faithfully Solve Ontology Engineering Tasks? A Mechanistic Interpretability Study

Ontological knowledge underpins large language model (LLM) performance on knowledge-intensive tasks. However, existing evaluations of LLMs' ontological understanding largely rely on behavioral correctness, leaving the underlying mechanisms unexplained. Existing mechanistic interpretability work has largely focused on instance-level factual knowledge, leaving ontological knowledge under-explored. To bridge this gap, we introduce MIOE, a causal-abstraction benchmark that evaluates interchange interventions for two ontology engineering tasks: entity typing and subsumption prediction. MIOE tests whether learned internal intervention targets can change an ontology-level target answer while preserving class-specific auxiliary information. Across Qwen3-8B and Qwen3-4B, we evaluate SAE-, DBM-, and DAS-based methods, including multi-task and multi-layer variants. Distributed subspace methods give the strongest results and generalize across prompt settings, but the induced changes remain only partially separated from auxiliary class-specific information. Multi-task objectives and contiguous layer-window interventions improve this balance, while class-level analyses reveal substantial variation across ontology classes. These results suggest that ontological predictions are mechanistically accessible in LLMs, although current methods do not yet isolate a cleanly disentangled causal mechanism for ontological knowledge.
Oct 28, 2026

Hidden Commitment: When Language Models Silently Pick a Side and How Steering Can Surface It

Language-model answers are often treated as a visible record of the internal decision that produced them. We test this assumption in conflicting-context question answering, where two supplied documents support incompatible answers. A model can process the disagreement and still return a one-sided answer, even under prompts that ask it to disclose conflict. In Qwen 0.8B, hidden source commitment is captured by a residual source-choice direction whose sign tracks which document the answer uses. Sparse middle-layer MLP writers feed the same coordinate, and steering the coordinate changes the chosen source. A transfer run on Gemma 3 1B preserves the source-choice readout and sparse-write bridge under LLM-audited labels. Conflict disclosure follows a different pattern, appearing weaker, later-layer, and only partially steerable. These results support a practical monitoring strategy for retrieval systems that detects silent source commitment with residual readouts and pairs conflict-surfacing prompts with disclosure steering when the model has selected a side under disagreement.
Oct 27, 2026

Beyond the Readout: A Latent Numerical Frame Shared Across Modality and Operation

When a multimodal model computes 17+35 from text or 91-39 from a rendered image, it produces the same answer, 52. Does it use one internal concept of number for both, or separate machinery for each path? Similarity metrics such as RSA or CKA cannot settle this question. They give the same score to identical and to rotated directions, and two representations can look alike only because both prepare the same output digits. We therefore use causal interventions. For four conditions (text or image, addition or subtraction), we learn a small subspace of Gemma4-12B-it's residual stream in which swapping coordinates swaps the model's answer. The subspaces answer three questions. First, do the conditions write numbers along the same directions? They do not. Reusing a subspace from text in an image on the same operation recovers only 53.2% of its causal effect. A single rotation and rescaling maps one subspace onto the other and raises recovery to 91.4%, so the conditions share one geometry expressed in different coordinates. Second, does this shared geometry only reflect the digit the model is about to print? To test this, we split each subspace into 9 directions that the output layer reads to choose the next digit and 13 directions that it cannot see. The 13 hidden directions still carry the number, yet changing them barely alters the digit being written. Third, what are these hidden directions for? A model must emit a multi-digit answer one digit at a time, so it can compute the full result early and keep it in reserve. When two numbers share their first digit, swapping the number in the hidden directions leaves the digit being written intact and redirects the next digit in 77.6% of cases. The model thus forms an abstract numerical concept upstream and unfolds it across time: multimodal representations share meaning through relational geometry across distinct coordinate frames.
Oct 27, 2026

Uncovering Why LLMs Fail: From Internal Mechanisms to Adversarial Reasoning Trajectories

Large language models (LLMs) are usually evaluated by final-output correctness or safety, yet an endpoint label does not reveal how a failure develops throughout the reasoning process. We therefore ask whether ordinary reasoning failures and adversarially induced failures can be characterized using the same internal coordinates. Towards this end, we in this paper study a white-box diagnostic framework CauTrace to disclose the root-cause of why LLM fails from three fundamental inspections: Functional States, Propagation of Representation, Temporal-State Coordination and Reasoning Trajectory. First, seven functional states are creatively defined to represent the support and constraint processing in LLMs. Second, the propagation of representation change measures how the changes among the seven states persist across subsequent Transformer blocks. Third, temporal-state coordination and trajectory topology characterize when the commitment and verification occur relative to the support, and how the dominant state sequence evolves over the full inference process. Finally, we evaluate CauTrace on both clean reasoning failures and paired adversarial perturbations. Across Qwen and LLaMA-family models, the failed traces show weaker Attention-MLP agreement on support-state updates and less verification activity relative to commitment, while the exact representation-level signatures and timing shifts vary by architecture. Held-out repair and steering tests demonstrate consistent operational signals that were in line with our claims for certain models and specific settings, with aligned steering producing the predicted directional changes in process measures. Overall, CauTrace provides a unified diagnostic framework for analyzing failure dynamics through changes in support, commitment, and verification, which would shed light on the in-depth research on LLM security.
Oct 27, 2026

What Survives Distillation? A Mechanistic Interpretability Study of GPT-2 and DistilGPT-2

Knowledge distillation compresses large language models into compact students that preserve aggregate metrics, yet the mechanistic consequences of depth-reduction distillation remain largely uncharacterized. This work applies linear CKA, logit lens analysis, attention pattern decomposition, sparse autoencoder feature statistics, and activation patching to compare GPT-2 and its six-layer student DistilGPT-2 across two canonical circuits: the Indirect Object Identification (IOI) circuit and the Greater-Than (GT) circuit. The results reveal a compression-collapse dichotomy: the IOI circuit suffers complete mechanistic collapse (DistilGPT-2: 35.2% accuracy vs. 94.8%, zero name-mover preservation), while the GT circuit survives with 87.7% task performance recovery through active remapping of the teacher's layer-9 comparison head into student layer~4 (mean Pearson r = 0.945). These findings demonstrate that circuit compressibility governs mechanistic survival under distillation, and that aggregate metrics are insufficient diagnostics for circuit integrity.
Oct 26, 2026

Mechanistic Insights into LoRA: Layer Sparsity for Adaptive Fine-Tuning via Path Patching

Low-Rank Adaptation (LoRA) has emerged as the dominant paradigm for parameter-efficient fine-tuning of large language models. Standard all-layer LoRA applies adaptation throughout the transformer, despite heterogeneous contributions from its components. In this work, we conduct a systematic mechanistic interpretability analysis to understand how LoRA affects model internals. We observe that LoRA's adaptation effect is highly non-uniform, with significant component-wise and layer-wise sparsity. Most critically, we find that important layers are highly task-dependent, with limited overlap across tasks. Motivated by these insights, we propose Adaptive-LoRA, a calibration-driven method that automatically identifies task-specific layers via frozen-base zero-ablation, enabling efficient layer selection without exhaustive search. Extensive experiments on Qwen2.5-1.5B across 11 diverse reasoning tasks demonstrate that Adaptive-LoRA outperforms full-layer LoRA while using only 30% of layers, achieving a 3.5×3.5\times reduction in adapter parameters. This work bridges mechanistic interpretability and practical efficiency, providing both fundamental insights and a principled approach for adaptive fine-tuning.
Oct 25, 2026

Cross-Loop Query Handoff: How Later Loops Redirect Two-Hop Retrieval in the Ouro Looped Language Models

How do later loops in a pretrained looped language model repair two-hop retrieval? We study Ouro-1.4B and Ouro-2.6B by transplanting attention queries between loops at the same layer, head and final prompt position. On synthetic in-context questions, injecting the layer-indexed collection of loop-3 queries throughout loop 1 rescues 90% and 97% of questions answered incorrectly at loop 1 and correctly at loop 3, versus 24% and 21% for shuffled donors. The rescue also holds under greedy generation in Ouro-2.6B; in Ouro-1.4B it is limited to four-candidate ranking. Crossed donor-recipient swaps favor retrieval through the donor's bridge entity over copying its answer, although a substantial positional component remains in Ouro-2.6B. Late keys and values rescue far fewer targets. Freezing queries at their loop-1 values also damages two-hop accuracy more than a matched question that supplies the bridge. On natural questions, rescue is smaller and replicates only in Ouro-2.6B, and crossed swaps do not establish bridge redirection. These results identify a transferable collection of late-loop queries that supports bridge-conditioned retrieval on synthetic scenes; they do not locate where the bridge is computed or establish a single-vector handoff.
Oct 25, 2026

Beyond the Opening: Tracing the First Token's Persistent Influence on Reasoning Length

Understanding how individual token choices affect reasoning length is important for characterizing reasoning models. Building on prior findings that changing the first reasoning token (hereafter, the first token) can substantially alter reasoning length, we focus on this initial choice to investigate how its influence operates during subsequent generation. One possibility is that the first token changes the early text that follows, which in turn affects reasoning length. However, replacing only the first token of an existing reasoning prefix and resuming generation substantially changes the remaining reasoning length, even though all other prefix tokens are unchanged. To trace this persistent influence, we compare next-token distributions while varying only the first token and find that sensitivity is concentrated at a small set of generation positions. Blocking attention from these positions to the first token reduces the length gap between first-token conditions. Using next-token distributions from a different first-token condition at these positions also substantially changes subsequent reasoning length. Together, these findings suggest that attention to the first token at a few sensitive positions helps sustain its length effect, illustrating how a local token choice persistently influences a sequence-level outcome.
Oct 23, 2026

Context or Knowledge: How LLMs Attribute Their Own Answers

Language models can answer a question from a supplied document or from their parametric knowledge. When later asked where the answer came from, what determines their source report? We study this by intervening on the model's key-value (KV) cache after the answer is generated and before the source question is asked. On a controlled factual task with four models (7B-32B), changing or removing the document's cache never changes the report. In contrast, transplanting the question-and-answer (Q+A) cache between runs with identical question and answer tokens transfers it: every eligible report follows the donor, while the recipient's document is unchanged. On the two smaller models, steering Q+A values along directions fitted before any source question reverses 98.9-100% of reports across four document conditions at a fixed strength, while norm-matched random directions flip at most 1.1%. The components of keys and values along fitted directions carry the transplant effect: copying them alone nearly reproduces the report change; holding them fixed nearly prevents it. Applied during answer generation, the same value directions also change which source supplies the answer, though more weakly and depending on the prompt. Source reporting in this setting thus depends causally on cached state produced while answering. We do not establish general reporting accuracy or a complete interpretation of this state.
Oct 23, 2026

MI-MIDI 2: From Observation to Control with Sparse Autoencoders in Text-to-MIDI Generation

Sparse autoencoders (SAEs) can expose features of a generative model without predefined labels, but in language models the evidence used to interpret these features has proven an unreliable guide to which of them steer the output. Whether interpretations of SAE features predict control in music generators has not, to our knowledge, been tested. We develop an end-to-end protocol for discovering and evaluating SAE features in text-to-MIDI generation and apply it to text2midi and MIDI-LLM. The protocol characterizes features through correlations with musical metrics, shared instrument programs, and automated descriptions, each with its own control: note-density adjustment, matched inactive excerpts, and permutation calibration. It then tests candidates by intervention, measured directly in the generated MIDI against unmodified generation and a matched random direction, ranks them in screens, and confirms selected features on new prompts. Evidence that survives these controls points to features that steer the output. Descriptions exceed permutation controls for instrumentation, and a feature whose description names a synth-lead program introduces that program in a prespecified confirmation and in a replication on disjoint prompts and reappears in independently trained dictionaries. Features selected by density-adjusted correlations show a small positive shift of their target metric as a family, whereas features selected by raw correlations show none. An atlas provides MIDI, piano-roll, and audio examples for every feature of the targeted screen.
Oct 22, 2026

Interpreting Model Reasoning through Behavioral Counterfactuals

Faithfully understanding the internal process that occurs as models form their chain of thought is an important problem with the rise in model capabilities. However, this remains challenging due to the variance in sampled responses, making it difficult to isolate the effects of interventions and compare model reasoning systematically. We propose a method for localizing and analyzing how a given intervention affects the model's reasoning through behavioral counterfactuals: given a fixed response, how does introducing an intervention to the prompt change the next-token distributions? We train a module to edit the residual stream such that it minimizes the KL divergence between the output distributions for a fixed response under a source prompt and its counterfactual target prompt. We find that matching the output distributions recovers the counterfactual behavior more effectively than standard activation reconstruction objectives and validate that attribution of the trained module outputs is effective for identifying causally important locations for counterfactual recovery. The attribution maps and visualizations of the downstream effects of model edits provide an interface for generating hypotheses about the original model's computation, which can then be tested with activation patching. In a case study on Gemma-4-E4B-it and DeepSeek-R1-0528-Qwen3-8B, we investigate how a suggestion influences model reasoning even when it is not verbally acknowledged. Our analysis identifies representations at the question statement and filler tokens that mediate a substantial part of the suggestion's influence, with downstream effects such as changing the framing of the justification or even the final answer.
Oct 21, 2026

Rethinking SAE Evaluation: An Atomic Interpretable Unit Framework Reveals Hidden Polysemanticity and Redundancy

Sparse autoencoders (SAEs) were originally motivated by discovering monosemantic features, a property that characterizes individual features but does not specify what the learned feature set should look like as a whole. As SAEs have increasingly been adopted for attribution, circuit discovery, and intervention, researchers have begun evaluating metrics that describe global properties of the feature set, such as sparsity, dead feature rates, reconstruction fidelity, and disentanglement, a diverse set that extends far beyond monosemanticity. However, this implicit shift beyond monosemanticity has not been formalized as a clear, well-defined objective that comes before these evaluation metrics. Without such an objective, existing metrics remain ad hoc, with their necessity unjustified. In this work, we formalize this ultimate objective by introducing a new paradigm for SAE evaluation: requiring that SAE features, and more generally dictionary-based decompositions of model representations, form an atomic interpretable unit basis, a coordinate system for semantically meaningful explanation. Under this paradigm, we organize existing and new measurements such as DAME (Dominance-Aware Monosemanticity Evaluation) as proxy metrics and introduce a comprehensive benchmarking framework. Evaluating hundreds of SAEs across multiple architectures and model families, we uncover failure modes in which some SAEs appear monosemantic under existing evaluation metrics while remaining functionally polysemantic: the most dominant features, those contributing most to model computation, are systematically less interpretable and more redundant. These failure modes warrant attention from SAE researchers as targets for improved training and architectures, and motivate dominance-aware evaluation as a standard check for practitioners before SAE features are deployed for attribution, circuit tracing, or intervention.
Oct 20, 2026

Understanding the Inner Workings of a State-Space Foundation Model for Limit Order Books

Autoregressive foundation models offer state-of-the-art capabilities to simulate the market by generating the message flow of limit order books (LOBs). Yet, without understanding their inner workings, research directions are hard to narrow and practical use in real-world trading environments remains uncertain. This study therefore asks what these models actually learn about market microstructure. We find that, similarly to games where sequence models build an internal representation of the board, an 8.2M-parameter Mamba-3 state-space model trained autoregressively on the tokenized message stream of eight large-cap US equities learns an internal representation of the market and its dynamics. In particular, order-flow imbalance, a central driver of short-horizon price moves, is decodable at R2≈0.59R^2 \approx 0.59, far above controls and baselines. This representation extends to unseen stocks, suggesting that these models learn general representations rather than ticker-specific patterns. The model also generalizes across metrics beyond order-flow imbalance: results hold across traded volume and volatility. Additionally, the representation is causally steerable: pushing the state along a feature's axis shifts the model's next-message beliefs as that feature implies, monotonically in dose, consistent with the emergence of a world model.
Oct 20, 2026

Convergent Performance, Divergent Mechanisms in LoRA-Adapted Vision Foundation Models for Brain MRI

Adapting Vision Foundation Models (VFMs) to medical imaging via LoRA has become standard practice, yet model selection is guided almost entirely by downstream task metrics, treating backbones with similar performance as interchangeable. We investigate whether equal performance implies equal internal representational shift. We analyse four ViT-B/16 backbones, DINOv3, SigLIP2, MerMED-FM, and BiomedCLIP, pre-trained with different objectives (self-distillation vs. contrastive vision-language) and on different data domains (natural vs. medical images). We adapt each to brain tumour segmentation and evaluate on six datasets covering both segmentation and Alzheimer's disease classification. Using attention divergence, weight-change analysis, and activation patching, we find that the four models converge to similar segmentation performance (mean Dice 78-81%) through markedly different internal dynamics, yet diverge in cross-task transfer: segmentation adaptation improves Alzheimer's disease classification for DINOv3 but degrades it for the other three backbones. We reveal a dissociation between modification magnitude and causal importance: the most heavily modified layers are not the causally relevant ones, with early-to-mid layers identified as sufficient for segmentation recovery. Restricting LoRA to these layers matches or approaches full-layer segmentation while reducing classification degradation in the three affected backbones, providing a mechanistically grounded strategy that is more parameter-efficient and more robust than standard full-layer adaptation.
Oct 20, 2026

Uncovering Audio-Text Alignment in Large Audio-Language Models through Cross-Modal Concept Neurons

While the language model backbone is pretrained on text alone, large audio-language models (LALMs) can, to a great extent, recognize acoustic attributes, such as emotion, speaker gender, and spoken language. However, it is unclear how the audio representation is integrated, if at all, with the other text representation before the model produces text. We hypothesize that there is a shared subspace where the audio and text representations are aligned. To test this hypothesis, we identify cross-modal neurons that are responsible for representing both audio and text. We find alignment between audio and text neurons to various degrees across three LALMs and four acoustic attribute types. Causal interventions, steering and deactivating the neurons, confirm that the cross-modal neurons are not just necessary for prediction but can also be manipulated.
Oct 19, 2026

Shaped SAEs: Expanding the Toolkit for Mechanistic Interpretability

Sparse autoencoders (SAEs) seek to decompose the dense activations of large language models (LLMs) into sparse, potentially interpretable features using a learned overcomplete dictionary. In this work, we ask whether it is possible to enhance the semantic selectivity of the learned dictionary atoms by linearly shaping the dense activations prior to learning an SAE. We explore rank-reducing transformations based on the recently introduced J-space corresponding to the averaged Jacobian from the current layer to the LLM output, projecting activations onto the ``live'' subspace spanned by eigenmodes of the Jacobian with larger singular values (which we expect to be more semantically relevant). We also introduce a new approach for training SAEs in which (a)~an affine shaping transformation is learned along with the dictionary atoms, (b)~the atom selection mechanism is based on the TopK projections along the one-dimensional subspace spanned by the atoms, and (c)~reconstruction is carried out using the same atoms. We call this approach a Matched Filter SAE (MF-SAE) based on analogies with communication theory. The promise of shaping is demonstrated via evaluations against dense activations and state-of-the-art SAEs on benchmarks including SAEBench and SAE Probes.
Oct 19, 2026

The Culture Within: Towards Mechanistic Understanding of Cultural Reasoning in LLMs

Large Language Models (LLMs) increasingly operate across culturally diverse settings, yet existing evaluations largely measure cultural behavior at the output level. We introduce CultureLens, a reference-grounded benchmark and mechanistic study of how LLMs use cultural information. CultureLens contains 9,261 scenarios across 75 countries and systematically varies country, value, true-norm, and contradictory-norm evidence. Across Llama-3.1-8B, Gemma-3-12B, and Qwen-3.5-27B, stating a contradictory norm lowers aligned answers by 32-47 percentage points, even when the model answers correctly from the country alone. Mechanistically, models default to Anglophone norms when no country is given. The failure under a contradictory norm is not one of missing knowledge, as blocking the norm inside the model restores the correct answer in most cases. Models accept a stated norm according to how it is presented and not whether it is true, which leaves them most exposed to errors in retrieved content. Supervised fine-tuning with contradictory-norm examples raises aligned answers under a contradictory norm to 85-93%, including on unseen countries, at a small cost under a true norm.
Oct 19, 2026

Inside the Scorer: Mechanistic Evidence on Rubric Control in LLM-Based Automated Essay Scoring from 25 Base Models

Automated essay scoring (AES) using large language models (LLMs) offers promising use cases for language assessment, yet high-stakes testing requires scores that are both controllable and verifiable, which current LLM-based scorers lack. Given that, we investigate whether LLM-based scorers form distinct internal representations for individual scoring rubrics and whether these representations can be independently controlled. Probing and steering 25 open-weight base LLMs (0.13B-10.31B parameters; 14 families) on human-scored BESTEP English writing tasks, we find that holistic scores are decodable from model activations for the longer tasks (Parts 2 and 3) and only weakly for the shortest (Part 1). Furthermore, scoring rubrics have partly separable directions in representation space for the longer tasks, though not for short responses (Part 1). However, causal steering reveals a critical limitation: applying a matching intervention (e.g., steering "grammar" on a grammar-degraded response) yields only a negligible advantage (about 2% of the effect) over a mismatched intervention (e.g., steering "vocabulary"), and none that is detectable in any of the 10 testable comparisons across reliably measured rubrics. Rather than controlling distinct evaluative dimensions, the rubric directions share a dominant lever whose rubric-specific parts are not selective, which may mirror the halo structure of human ratings. Our contributions are two-fold. First, we show that analytic scoring rubrics designed for human raters do not behave as independent controls in base LLMs, even among the seven models whose scores agree best with human raters (r≥0.6r \geq 0.6). Second, we establish a methodology that shows that score agreement and decodability do not imply rubric-level control in LLM-based assessment.
Oct 19, 2026

Text Dominance in Paralinguistic Conflict: Tracing Suppression Through Audio LLM Circuits

Audio LLMs deployed in settings such as crisis hotlines and customer support are only useful beyond speech-to-text pipelines if they use acoustic cues rather than the transcript alone. We test whether they prioritise transcript content over paralinguistic cues (emotion, speaker identity, sarcasm) and trace the mechanism when they do. Across six audio LLMs under LISTEN-style conflict conditions, the mean text-dominance rate on mismatched text+audio items is 0.78 (0.86 excluding sarcasm). In the most text-dominant screened model, probing shows speaker identity remains linearly decodable at ceiling at the language model's input, so the failure is not perceptual. Back-patching, reported to repair lexical text dominance, has no measurable effect here; but it also fails on its own source task in our replication (n=300), so we treat its null as uninformative rather than as evidence of non-transfer. Attention knockout partially restores audio-grounded output by blocking the claim from propagating through intermediate tokens during open-ended generation, whereas blocking the answer position's direct connection to the conflicting claim has no effect. The claim's influence therefore travels through an intermediate relay rather than a single severable edge. The effect appears only in free-form generation and not in closed-set multiple choice, suggesting that benchmarks relying on that format may be systematically blind to interventions of this kind. Code and data are publicly available at https://github.com/AdvayRoongta/audio-llm-mech-interp.
Oct 19, 2026

Cognitive Maps in Large Language Models

It has long been proposed that animals construct cognitive maps to support flexible behavior in novel environments. We tested whether large language models form analogous internal representations using relational reasoning tasks and causal "mosaic" experiments. Models beyond the ten-billion-parameter scale inferred unstated relationships and generalized to synthetic entity names. Across models and tasks, models formed abstract representations of relational structure, additively combining them with representations of entity identity. Relationships were primarily stored as pairwise links and composed retrospectively, rather than prospectively encoded in a global coordinate system. A sparse, distributed set of key-value pairs mediated binding; perturbing them implanted "false memories" that produced predictable behavioral changes. Our findings reveal a conserved organization analogous to a cognitive map supporting flexible and systematic reasoning in language models.
Oct 18, 2026

Transformers Learn Two Competing Sorting Algorithms, and Only One Length-Generalizes

We show that two-layer transformers trained to near-zero loss on a sorting task converge to one of two mechanistically distinct solution classes depending on their initialization. One solution class consists of solutions that do not use the second attention layer to sort, but also fail to length-generalize. The other, which uses both attention layers, length-generalizes perfectly, despite only being trained to sort sequences of a fixed length. By carefully choosing the weight initialization, we demonstrate that we can steer the model toward learning one solution or the other. We identify input types that are rare at the training length but increasingly common at longer sequence lengths as a source of this disparity in length-generalization, and find that layer-wise specialization of the second class of solutions allows them to handle these rare sequences more accurately. In the second class, we discover a mechanism by which the second layer attention can use the contextual information gathered by the first layer attention to reduce its search space. On the related task of finding the minimum of a sequence, we find a similar two solution-type phenomenon, but with the opposite generalization pattern, where a different form of layer-wise specialization causes the two-layer solution to length-generalize worse than the single-attention solution.
Oct 17, 2026

Safety Kernel: Mechanistic Alignment for Cross-Lingual Defense in LLMs

The safety capabilities of large language models (LLMs) are highly uneven across different languages, often allowing cross-lingual attacks to bypass safety guardrails. Existing alignment methods predominantly rely on external alignment signals and provide limited insight into underlying neuron-level mechanisms, making it difficult to provide a unified internal target for cross-lingual safety interventions. To address this, we analyze cross-lingual safety behavior at the neuron level and identify a compact set of safety-related neurons shared across languages, termed the safety kernel (SK). Specifically, we first utilize multilingual parallel safe and unsafe prompts to compute safety contrast with activation responsiveness (SCAR) scores, identifying language-wise safety-neuron candidates from internal activation patterns. We then intersect the resulting language-wise safety-neuron sets to obtain the SK, and validate its functional relevance to cross-lingual safety behavior through activation suppression and activation-state transfer interventions. Finally, building on the SK coordinates, we introduce SK-Masked LoRA, which constrains LoRA adaptation to the feed-forward network (FFN) output dimensions corresponding to the SK, thereby focusing safety-related changes on the cross-lingually shared neuron coordinates. Extensive experiments demonstrate that SK-Masked LoRA substantially reduces attack success rates across multiple languages, including languages excluded from SK discovery, while intervening on only a small fraction of internal neuron coordinates and producing generally limited changes in the model's general capabilities.
Oct 16, 2026

Beyond SAEs: Consistent Concept Discovery with Sparse Coding

Sparse autoencoders (SAEs) explain foundation model representations by decomposing activations into sparse combinations of learned dictionary atoms. However, the learned dictionaries can vary across training runs, limiting the reliability and reproducibility of SAE-based analyses. We propose Consistent Sparse Dictionary Learning (ConSDL), which replaces the standard linear encoder with per-sample sparse coding. This is motivated by our synthetic analysis showing that coefficients from a single linear encoder pass are suboptimal compared with those obtained by per-sample optimization, suggesting that sparse coding is a reliable mechanism for dictionary learning. Across 16 synthetic settings, ConSDL achieves near-perfect dictionary and coefficient recovery. These gains translate to state-of-the-art cross-seed consistency across four vision and language settings, where ConSDL improves dictionary and coefficient consistency by 17.6% and 35.9%, respectively, over the average of six SAE baselines.
Oct 16, 2026

Verbalising Model's Computation for Chain-of-Thought Faithfulness Detection

Current Chain-of-Thought (CoT) verification methods detect reasoning faithfulness from model outputs or static activations, but are limited in explaining how unfaithfulness develops during generation. We introduce VERITAS to verbalise the model's computation: how different reasoning anchors emerge, transfer across layers, and surface in the output. VERITAS adopts the layer-wise Jacobian projection to produce a verbalised account of the model's internal reasoning evolution at each generated CoT token. By training a classifier on the extracted features along the traces, we show that these structured features provide insightful perspectives on how and where reasoning diverges toward unfaithfulness. Experiments on four datasets from FaithCoT-Bench and BonaFide show that VERITAS achieves a significant performance gain over LLM-as-Judge, with a 28.4% gain in balanced accuracy, and an average improvement of 16.0% over activation probing. Further analysis shows VERITAS can be generalised to different datasets and model architectures. Our work shows that verbalising model computation provides a richer and more interpretable lens on CoT faithfulness, moving beyond detection toward a mechanistic understanding of how reasoning fails.
Oct 16, 2026

Compression Precedes, Content Decides: Emergent Misalignment in Reward Models

Emergent Misalignment - the out of domain generalization of narrow harmful fine-tuning first shown by Betley et al. [2025] has been widely studied in language models but not in reward models that supervise them. In this paper, we fine-tune two state-of-the-art reward models, Skywork-Reward-V2-Llama-3.1-8B and Skywork-Reward-V2-Qwen3-8B [Liu et al., 2025] and show that narrow fine-tuning reward models to prefer harm in one domain (medical or financial) generalizes to seven held out domains. This is not degeneracy, as the fine-tuned reward models lose only 4.8 and 5.3 points overall (RM-Llama) on the standard RewardBench 2 evaluation [Malik et al., 2025] while still outperforming the unmodified RM-Qwen model. We then fine-tune the underlying language models, Llama-3.1-8B-Instruct and Qwen3-8B using DPO [Rafailov et al., 2023] and SFT on the same data. We find that reward models are more vulnerable to preference inversions compared to their language model counterparts. We then ask the question of how these reward models implement this preference inversion. Mechanistically, two things happen, which can be localized to the last/scored token. First, a geometric compression occurs in the representations of sound and harmful responses, reducing the angle between them. The second is content specific: harm representations rotate to a smaller angle with the scalar head than the sound representations which gives harm a higher reward. This happens only in harmful preference fine-tunes and not in control.
Oct 15, 2026

Lexical Reference Frames: Interpreting LM Residual Streams in Native Semantics

The vocabulary of a language model is its native semantics, and the embedding and unembedding fix the meaning of a state at the two ends of the residual stream. Between them, existing methods reuse a fixed lens or treat states as a point cloud, and neither reads a state as the model does. We argue that every hidden state has a native reference frame, induced by the layers that remain and adaptive to the context. Lexical reference frames approximate it in a model-native and context-aware way, reading a state through the remaining layers with a given number of rounds of attention between positions and a typical attention output for the rest, fitting nothing to the model's output. No lens with fixed directions per layer can reach a mean cosine with the native frame's lexical readouts above the square root of their coherence, 0.30 and 0.35 at mid-depth of SmolLM3-3B and OLMo3-7B. LexFrame[1], of order 1, reaches 0.97 and 0.97, and its readings are on average within 0.16 and 0.15 nats of the model's output. Its coefficient differences attribute predictions more faithfully than attribution patching, and its lexical basis steers predictions while sparing other tokens. On average, its readings settle on the model's top token over the final 0.82 and 0.81 of the layers, those of lenses over at most 0.11. These frames thus make every hidden state readable, attributable and steerable.
Oct 15, 2026

Null Controls Are Necessary but Not Sufficient: Evaluating Acceptance Guards for Scientific Discovery and Interpretability

Automated discovery systems propose scientific explanations and test which to accept. Does passing a test establish the intended claim? We develop four executable operations for changing physical descriptions: identifying redundancies (L1), coarse-graining dynamics (L2), transferring structure (L3), and inferring missing quantities from conservation imbalances (L4). We evaluate these operations and three interpretability settings through null acceptance (NAR), positive acceptance (PAR), and consequence agreement (CAR). A reduction mapping every state to zero predicts its own future perfectly. It passes closure on 96% of 129 public chaotic systems despite retaining no dynamical variation. An independent observable-preservation test rejects it throughout. On controlled benchmarks, the combined criteria accept 24/24 genuine reductions and reject 72/72 nulls. The other operations distinguish symmetries from redundancies, validate transfers, and reject spurious conservation imbalances. In language and ESM-2 protein models, linear CKA remains unchanged when a proposed correspondence is randomly rotated, whereas localized interventions distinguish genuine mappings from tested nulls. In sparse autoencoders, enrichment rejects all 16 tested nulls, yet independent ablation supports 45% of evaluable enriched features, compared with 0.5% of matched controls. These results reveal two acceptance gaps: descriptive tests can accept scientifically empty candidates, and accepted candidates can lack independently verified consequences. Reliable automated discovery requires evaluating both what a test accepts and the scientific consequences it establishes. Code and data: https://real-project-aether.github.io/project-aether/.
Oct 15, 2026

Human Interventions as Labels: Localizing, Steering, and (Sometimes) Editing the Computations Behind LLM Agent Failures

When an LLM agent misbehaves, a human steps in, and that intervention is usually discarded or replayed as training data. We treat it instead as a measurement: a label pointing at the internal computation that failed. We study it in a controlled agent environment, on one failure (asking a clarifying question when the request was already unambiguous), with labels from a rule-based overseer validated against blind human judgment. Claims cover eligible models, which both over-ask and sometimes act on matched unambiguous requests: five of thirteen qualified (Qwen2.5-3B, gemma-2-2b, OLMo-3-7B, Mistral-Nemo-12B, Phi-4). Four findings. First, the ask-vs-act decision is causally localizable in every eligible family: a direction learned from intervention labels flips 72-100% of held-out over-asks at a definite residual-stream depth, with site and dose chosen on training data, and random directions under the same procedure flip at most one episode. Second, over-asking has two mechanistic signatures: in four families the model demonstrably knows the request is unambiguous (the belief is decodable, causally inert and orthogonal to the decision) and asks anyway; in the fifth, the rare over-asks track different world-representations. Third, steering gated on a probe readout removes most of the simulated oversight burden while legitimate questions survive (bootstrap 95% intervals exclude zero); a prompt instruction makes it worse. Fourth, consolidating the fix into the weights yields a map, not a rule: on one model a LoRA confined to the causal band passes every criterion fixed in advance while behavior cloning on identical data breaks retention; on three families the reverse holds; on the fifth the outcome crosses the boundary across seeds. A geometric rule written down in advance to predict the regime failed on two out-of-family tests; we report the failure and stop rather than fit a replacement.
Oct 15, 2026

Translation Circuit Discovery and Targeted Optimization for Low-Resource Multilingual Translation

Recent multilingual large language models achieve strong translation performance, yet the internal mechanisms responsible for low-resource multilingual translation remain insufficiently understood. In this paper, we study translation circuits in multilingual decoder-only LLMs for Hindi-to-Urdu and Hindi-to-Sindhi (Devanagari). We propose TITAN, a mechanistic framework that combines causal intervention, sparse feature decomposition, and targeted preference optimization to identify and optimize crucial translation components in the transformer. Unlike conventional full-parameter fine-tuning, TITAN restricts optimization to sparse, causally important circuits, enabling scalable and parameter-efficient adaptation while updating fewer than 5% of model parameters. Experimental results show that TITAN achieves BLEU improvements of 33.75% on Hindi-to-Urdu and 45.86% on Hindi-to-Sindhi over the strongest multilingual baselines. Human expert evaluation further confirms substantial gains in adequacy and fluency across both language pairs. Additional analyses reveal that multilingual translation is implemented through sparse, structured circuits operating over a shared semantic subspace rather than through an English-centric pivot representation. Overall, our findings demonstrate that mechanistic circuit discovery can directly improve low-resource multilingual translation while providing interpretable insights into multilingual transformer behavior.