Model Activations

Attributing language model outputs to their internal computations is an open problem in interpretability. Existing methods, which use causal interventions, gradients, or learnable masks, either are infeasibly expensive or struggle to identify actual causally-important internal computations. We propose framing attribution as the problem of identifying nested subsets of internal components which minimise a downstream loss. To learn this task, we introduce Matryoshka Attribution (MAttr), a mask learning method that parametrises the mask with a simple differentiable sigmoid top-$k$ operator. We supervise training over all sparsities simultaneously by randomising $k$ over training, resulting in a learned ordering of components by attribution score. MAttr achieves number 1 on the official leaderboard of the Mechanistic Interpretability Benchmark (Mueller et al., 2025); our method identifies sparse and task-transferrable circuits across varying circuit bases. As a practical application, we show that MAttr can be trained with reinforcement learning to identify weight changes responsible for downstream behaviours in LLM finetuning. We train MAttr on refusal judge scores and find that restoring $1\%$ of Llama 3.1 8B Instruct's weights to their base model state is sufficient to remove refusals while maintaining capabilities. We view MAttr as a successful formulation of interpretability into a learnable objective that we can tackle with gradient descent, and encourage future work along these lines.

Linear probes can decode safety-relevant concepts such as truthfulness from language-model activations, but probe accuracy may show only decodability, not that the features the probe weights causally drive model behavior. We demonstrate that this gap cannot be closed from the geometry of probe weights alone: the features geometrically aligned with probe direction need not be the ones the model uses, so causal relevance requires intervention. We introduce a feature-level diagnostic that decomposes a deployed True/False probe into sparse-autoencoder (SAE) features, ranks those features by both probe alignment and by gradient sensitivity of the model's behavior, and ablates the resulting shared, probe-only, and random feature sets under a coherence gate. On the truth probe of Buerger et al. (2024) (TTPD), applied in the instructed truth/deception setting of Long et al. (2025) for Gemma2-9B-Instruct, the two rankings overlap only weakly (about 12%, Spearman rho = 0.10), and ablation dissociates them sharply: features the probe shares with the model flip the output far more (up to 27%) than equally sized probe-only (6%) or random (1%) features at full coherence, while probe-only features instead perturb the probe's own readout. The dissociation holds across five seeds and a held-out split, and an activation-aware selection of features flips behavior nearly three times as often as the probe's geometric top features (17.6% vs. 6.1%). In this setting, therefore, the geometric projection of a probe's weight vector alone does not identify the features the model causally uses; however, combining probe information with feature activation statistics recovers substantially more behaviorally causal features, and coherence-gated SAE intervention is needed to separate them from probe readouts.

Reinforcement learning (RL) is widely utilized in large language model training to improve targeted capabilities, yet how RL reshapes a model remains poorly understood. Prior attempts to explain how RL works largely offer behavioral perspectives, leaving open what RL gives a model at the representation level: can RL create genuinely novel features, and which existing features does it enhance or suppress? Recent developments in mechanistic interpretability suggest sparse autoencoders (SAEs) as a promising lens to decompose internal activations into human-interpretable features; however, they cannot be directly applied to tracking change across training. In this work, we introduce Fixed-SAE Track, a framework that trains one shared SAE per considered layer on activations pooled across the base model and all RL checkpoints, holding every feature direction fixed so that representation shifts are rigorously defined through the activations of interpretable SAE latents, including the detection of emerging novel features. Validated across multiple datasets and RL algorithms, we find that RL-induced drift is small, gradual, concept specific, and concentrated in late layers, mainly enhancing the sampling rates of a small set of ladder tokens, formatting scaffolding such as step breaks and answer delimiters, rather than reshaping problem content. Steering these features into the base model recovers around 80% of RL's performance gain, suggesting that RL primarily elicits capabilities the model already possesses, much as steering does. We further design a synthetic benchmark with features known by construction to test whether RL can instill genuinely novel features. We believe Fixed-SAE Track provides a principled approach to tracking representation shifts and offers representational evidence for understanding how reinforcement learning changes the inner representation of LLMs.

Activation steering provides a lightweight way to control large language models (LLMs) by modifying their hidden activations at inference time. Among these approaches, norm-preserving steering aims to change model behavior without altering the activation norm, reducing the risk of representation collapse and degradation. However, existing norm-preserving methods are limited by predefined steering trajectories and by their reliance on one-step updates, which may fail to capture the complex structure of activation distributions. We propose GeoSteer, an optimization-based method for norm-preserving activation steering. GeoSteer formulates steering as a Riemannian optimization problem and updates activations through a sequence of small geodesic steps on the representation manifold. To avoid fixed steering directions, GeoSteer learns a nonlinear activation-space objective that distinguishes desired from undesired activations, and uses this function to adaptively guide each steering step. This multistep formulation yields smoother, more stable, and more consistent steering behavior while preserving the activation norm. Across TruthfulQA, RealToxicityPrompts, and UltraFeedback benchmarks, GeoSteer consistently improves over state-of-the-art activation steering baselines. These results suggest that norm-preserving steering can be made more effective by replacing predefined one-step edits with adaptive, geometry-aware optimization.

Recent methods in language model interpretability employ techniques such as sparse autoencoders to decompose residual stream contributions into linear, semantically meaningful features. Such methods are commonly interpreted as identifying features that persist in the residual stream and that subsequent layers build upon. We challenge this view by identifying the Transformer Layer Correction Mechanism (TLCM), wherein adjacent transformer layers systematically counteract portions of each other's contributions. TLCM appears in 5 out of 7 major open-source model families and activates across nearly all tokens in diverse texts. We show that TLCM emerges during pretraining, operates most strongly on contextually dependent tokens, and adaptively calibrates its correction strength based on the preceding layer's output. Using the layer Jacobian, we further show that TLCM selectively corrects specific subspaces while reinforcing others, which we interpret through a ``propose-and-reject'' framework in which layers propose candidate features and subsequent layers selectively remove inappropriate ones. This dynamic suggests that the residual stream at any layer contains transient proposals alongside persistent features, helping explain why SAE feature descriptions often have low specificity, why effective model steering requires extreme feature amplification, and why transcoders hold a theoretical advantage over SAEs.

Activation steering promises modular control of language model behavior: a behavior such as politeness corresponds to a direction in a model's activations, and adding that direction while it generates should switch the behavior on and leave everything else alone. It does not. We ask what decides which other behaviors move, and by how much, and find that it is the model rather than the behavior being steered. A steer relaxes the model toward a small set of behaviors it already favors, chiefly refusal, sycophancy, and poeticism, and that set is much the same whatever is steered. Three results across 24 behaviors and ten instruction-tuned models support this, every effect read off the generated text by a language-model judge rather than off a probe. That readout matters: all 24 behaviors are linearly decodable, but only 20 change what the model writes. First, a direction carrying no behavioral content, matched to a real steer only in the size of the vector it adds, moves the same behaviors in the same order as real steers do, while producing none of the behaviors that need a specific direction. Second, most interference runs one way, so it cannot be an overlap between two directions: steering profanity makes the model toxic, while steering toxicity leaves profanity untouched. Third, with a behavior held out entirely, geometry measured on the others explains almost none of the interference it takes part in. The account holds on all ten models, the pull toward defaults strongest below 10B parameters and weakening in each family's largest. Reading a steer as a perturbation whose endpoint the model fixes implies that disentangling behavior directions cannot by itself make steering modular.

Activation steering suppresses undesired behaviors in language models by adding a steering vector to the hidden state during generation. Recent conditional methods such as CAST and DSAS improve the behavior-capability trade-off by deciding when to intervene, but once active, they apply the full dense vector to all hidden dimensions, regardless of whether a neuron carries concept information or already lies in the desired regime. We introduce dimension-level conditioning as a complementary axis of selectivity that also decides which neurons to intervene on. Our method, GAPS (Gated Activation steering via Posterior and Separability), combines two training-free gates: a static separability gate that restricts steering to neurons with statistically reliable concept information (via AUROC), and a dynamic posterior gate that steers a neuron only when its current activation is better explained by the undesired concept under a Gaussian model. The gates add O(D) overhead per token, and they plug into existing conditional methods. On toxicity mitigation (RealToxicityPrompts) and concept removal (OneSeC) with Gemma-3 (4B) and Qwen-3 (1.7B), GAPS consistently matches or improves the Pareto front of its token-level counterparts; under a fixed capability budget, DSAS+GAPS reduces Gemma-3's toxicity rate from 6.52% to 0.48%, versus 3.52% for DSAS alone. Ablations attribute most of the gain to the posterior gate.

Patients with depression present with diverse symptom profiles, yet clinical practice routinely reduces this variation to a single severity score. Large language models (LLMs) can potentially capture various symptoms and their severity from patient speech. However, how depressive symptoms are represented inside LLMs remains poorly understood, limiting clinical trust. To examine whether internal model activations match clinician judgment, we analyzed the residual stream of Gemma-3-27B-PT using mechanistic interpretability techniques. Recording activations across symptom descriptions drawn from validated clinical instruments, we found that symptom groups geometrically separated the most at layer 21 across multiple distance metrics. Using Semantic Projection, we then projected held-out naturalistic text onto Symptom Vectors constructed from these instruments. The resulting per-symptom coefficients preserved clinician-annotated rank ordering across mood, somatic, and suicidality axes. Furthermore, a single depression vector in Layer 21 separates held-out depressive from non-depressive text (AUC = 0.789), which can be used as an emotional valence gate that restricts symptom projection to depressive speech. These results reveal a decorrelated, clinician-aligned symptom signal readable directly from internal activations, offering a mechanistic foundation for interpretable depression-assessment tools.

LLM-based user simulators are increasingly used to evaluate autonomous agents at scale, in place of costly human evaluations. Despite this promise, these simulators exhibit "assistant bias," a tendency to cooperate and pursue task goals. They rarely reproduce the frustration or disengagement that real users exhibit, compromising evaluation validity. Prior work outlines that this bias is baked in during model training, which role-playing prompts fail to override. We analyze this bias from model activations, extracting a user role vector by contrasting how the model represents user versus assistant perspectives on the same dialogue. We observe two findings: (i) the user direction is identifiable in activations, elicits user-like behaviors, and captures characteristics distinct from assistant traits; and (ii) although user-role activation associates with simulation realism and steering strengthens it, it can exaggerate user behaviors and override individual user profiles. Together, our findings provide a representation-level analysis of LLM user simulators, confirming that assistant bias is structurally identifiable and that user behavior can be directionally analyzed.

Large language models (LLMs) trained only on text and code can sometimes generate programs that draw recognizable images. However, it is unclear whether this reflects an internal representation of 2D spatial layout or simply the ability to translate spatial descriptions into code. We introduce Autoregressive Mosaics (AM-Bench), a benchmark that separates these factors: First, a translation task gives a model a fully specified geometry of a picture in words as a prompt and asks for the code that produces it. Second, a layout task requires the model to compose an image from an underspecified prompt. Across eight open-weight text-and-code-only models, all models reliably translate specified geometry into code, but their open-ended layout performance differs substantially, indicating that these differences are not explained by code-generation ability alone. An output-medium ablation further shows that the interface or medium of expression that the model uses matters: replacing procedural code with raw SVG improves layout scores across all models. Finally, probing model activations shows that a coarse layout plan is present before generation, but reflects only the layout implied by the prompt. During generation, models track the evolving geometric state instead of executing an initially fixed plan. Overall, these results show that 2D spatial performance in text-only LLMs depends on both the model and the output medium, and is not explained by code-generation ability alone.

We present the first systematic study of Massive activations (MAs) in layer-interleaved HLA LLMs and uncover two architecture-aligned morphologies: MAs consistently spike immediately before full attention layers, forming pre-attention spikes (PAS), and can persist through intervening linear attention layers, giving rise to inter-spike plateaus (ISP). As full attention becomes denser, successive PAS become increasingly connected through ISP, ultimately recovering the stable MA morphology of full attention LLMs. We establish the recurrence of this organization across five linear attention architectures, six hybridization configurations, five data domains, and representative open-source hybrid models spanning 1.2B to 397B total parameters. Controlled pretraining of GDN-based hybrids at scales up to 1.3B shows that both morphologies emerge early and respond asymmetrically to output gating: full attention output gating strongly attenuates their absolute magnitudes without eliminating their layerwise organization, whereas removing GDN gates yields comparatively modest amplification. Mechanistically, our systematic-outlier analysis supports a shared lifecycle account governed by the timing of MA cancellation. PAS follows a localized write-sink-cancel process, while the extended persistence of ISP is consistent with delayed cancellation. At the full attention limit, this account recovers the stable MA morphology characteristic of full attention LLMs. Our code is available at https://github.com/StartluxLabs/Massive-Activations-HLA.

Lens methods interpret large language models (LLMs) by mapping intermediate activations to the output vocabulary, revealing how next-token predictions develop through the network. Trained lenses remain expensive: affine-translator parameters grow quadratically with model width, while exact, full-vocabulary Kullback--Leibler (KL) training dominates memory. Consequently, prior trained lenses have been applied to models of at most 20B parameters and remain tied to particular component types. We present OmniLens, which applies a single lens family to any model-width activation, whether residual stream, attention, or MLP, and combines two independent scaling techniques. First, low-rank translators make per-lens parameter growth linear in model width and reduce trainable parameters by up to 98.4%. Second, Subset-KL materializes only selected vocabulary logits: its Top-k mode cuts peak training memory by up to 70%, while its importance-sampled variant retains unbiased stochastic gradients for the full KL. These savings enable a dense ensemble of 482 lenses for LLaMA-3.3-70B, providing 6x the coverage of a residual-stream design at the same depth. Model-wide coverage then reveals what single-component lenses cannot: the components where a behavior is most visible need not be those where intervention is most effective, and the most effective interventions lie outside the attention heads examined by prior lens studies. Across three case studies (prompt-injection detection, multi-hop memory injection, and toxicity localization), OmniLens reproduces key published results at substantially lower cost.

Activation-based tools are usually tied to one model's native hidden space, requiring probes, sparse autoencoders, and natural-language interpreters to be rebuilt or rediscovered for each new language model. We present a Universal Activation Bus, a framework that provides a common activation interface across compatible language models. Using a small set of source models, we learn a shared dense space together with one lightweight linear encoder--decoder adapter pair per model. After source training, the interface is frozen; a new model joins by fitting only its adapter pair on unlabeled matched text. The resulting interface allows activation-based tools to be shared across connected models, including common probes and SAE features as well as access to an NLA originally trained for a different model. Across five models, semantically related texts form consistent neighborhoods in the shared space, and an onboarded model reuses these tools effectively without retraining them. We further show that an intermediate activation from one model can be used by another model's frozen upper layers to produce predictions. These results establish a stable, model-wise activation contract for reusable tools across compatible language models.

Activation steering edits the behaviour of a frozen language model by adding a learned vector to its residual stream, and current practice fixes the injection layers globally per task. We argue that the best layers are an instance-level decision, and we make per-instance, multi-layer selection both well understood and deployable. On two open-weight 8B models and six binary persona traits, a per-instance oracle over layer subsets shows that the best layers vary from one input to the next: on most trait-model pairs, no fixed global layer set recovers the per-instance benefit. A greedy rule that ranks layers by single-layer marginal effect recovers nearly all of the oracle's benefit, but both must score candidate layers against the gold answer, so neither can run at deployment; the rule instead becomes the target a prompt-only predictor is trained to reproduce. Our deployable recipe needs no label at inference: a per-instance layer ranker read off the prompt embedding, a classifier that infers the steering direction, and an adaptive gate that scores short steered passes against that inferred direction and steers no more layers than necessary. The recipe recovers most of the oracle's lift (the bulk on the stronger model, a clear majority on the harder one), never drives any trait-model pair below its unsteered alignment baseline on average, and largely avoids the fluency collapse that strong global selection incurs at higher layer counts. A mechanistic account, "direction over magnitude", explains the behavioural flip under a mis-directed global set, the output collapse from steering too many layers, and the ceiling of unsteerable inputs.

Khatri et al. (2026) [DOI: 10.1109/DSN-W70714.2026.00027] show that lightweight MLP probes on final-layer activations of a single 8B model (LLaMA-3.1-8B) detect harmful prompts at F1 competitive with guard models 1000x larger, using one probe per benchmark. We reproduce this pipeline end-to-end and extend it along two axes the original study leaves open. First, we test whether the result generalizes across other model architecture and scale by training identical probes on activations from models like Gemma-4-E4B, Mistral-7B-v0.3, and Qwen2-7B, using the three benchmarks (WildJailbreak, BeaverTails, AEGIS 2.0). Second, we test how much of the reported performance is affected by non-determinism during inference by repeating extraction under five random seeds and measuring the variance of F1 scores. Our results reproduce the original LLaMA model benchmarks within 0.37 percentage points of the original F1 scores (and within 0.2 points on BeaverTails). We find that the original MLP probe architecture extends to other model families with F1 scores within a point of the values reported for LLaMA-3.1-8B. Our experiments varying seed values reveal an interesting observation: final token latent vectors remained the same for all tested architectures irrespective of the seed values used.

Existing measures of how much a text is about a concept read the surface of the text: dictionary word shares, topic proportions, embedding similarities. They score the words a text uses, not the judgment a reader forms about it. Recent work has shown that a gap exists in what Large Language Models (LLMs) know internally versus what they express in their response. This paper asks whether that internal knowledge, read by monitoring the activations of frozen, out-of-the-box LLMs, can stand in for task-specific fine-tuning when measuring concept content, and which extraction method reads it best. We extract such measures via the Recursive Feature Machine (RFM) algorithm and via linear probing, and compare these against an embedding baseline, surface baselines, and the same model's own answer to the question. We demonstrate the approach on financial text, a domain studied extensively and served by established annotated resources, using a human-annotated Environmental, Social and Governance (ESG) dataset. The best linear probe comes within 0.6 percentage points of a fine-tuned domain classifier's accuracy without any task-specific fine-tuning, and outscores the same model's own answer to the question in eleven of twelve comparisons, so the activations carry concept content the response does not report. The simple probe consistently beats the RFM concept vectors, which in turn provide what classification alone does not: a continuous score intended to reflect how strongly a concept is present in a text, whose validation awaits graded labels.

Automatic speaking assessment systems are increasingly deployed in high-stakes settings to mark second language (L2) learners' speaking tests, making it critical to show that their scores depend on speaking proficiency rather than irrelevant speaker attributes such as first language (L1) or age. Transformer-based foundation models have improved the accuracy of these L2 speaking graders, but their black-box representations make fairness and interpretability analysis more difficult. Building on prior work that used Concept Activation Vectors (CAVs) to detect bias towards unwanted attributes (`concepts') in feature-based graders, we extend CAV-based analysis to two neural speaking assessment systems: a text-based BERT grader and a speech-and-text multimodal grader based on Whisper. CAVs represent human-interpretable concepts as directions in a model's activation space, allowing us to distinguish between whether a concept is encoded in a model's internal representations and whether it influences the predicted score, the latter quantified using a gradient-based sensitivity metric. Since CAVs rely on linear separability, which is less likely in complex neural embedding spaces, we also investigate whether sparse autoencoders (SAEs) provide cleaner concept directions by learning CAVs in a sparse latent space and mapping them back to activation space. Our analysis shows that concept recoverability depends strongly on the representation and architecture being probed, rather than on the concept alone. Sensitivity to concepts is also architecture-dependent. SAEs make concepts more linearly recoverable, but attenuate the original activation-space sensitivity, especially in low-dimensional layers. These findings highlight the need to distinguish concept recoverability from concept influence when auditing bias in speaking assessment systems.

Fine-tuning enables a source model to acquire desired capabilities and behaviors in a target domain while retaining much of its general-purpose competence. However, this adaptation process can also degrade alignment properties that were present in the source model. Recent work has shown that large language models can be trained using LoRA-based modules known as introspection adapters (IAs) to describe behavioral changes induced by fine-tuning. However, existing studies primarily consider settings in which the model is fine-tuned on datasets explicitly designed to implant a specific behavior and is then asked to explain the implanted behavior. This differs from practical deployment scenarios, where the central concern is often side-effect misalignment: unintended degradation of alignment caused by fine-tuning on tasks that are not obviously related to safety or alignment. To bridge this gap, we formulate a novel problem setting called \emph{side-effect introspection}, in which the target of introspection is not a behavior explicitly implanted through fine-tuning, but rather alignment shifts that emerge as unintended side effects, and we construct a dataset for this setting. Furthermore, to enhance sensitivity to internal model changes, we propose the Delta-Aware Introspection Adapter (DAIA), a novel mechanism designed to explicitly process both base-model activations and activation differences induced by fine-tuning. Our empirical evaluation shows that introspection learning generalizes to unseen fine-tuned models and safety categories, and that DAIA consistently outperforms existing introspection adapters.

When a language model fails on surface-perturbed input (typos, OCR noise, homophones), "which layer is responsible" has three natural operationalizations: where representations diverge most (sensitivity), where restoring clean activations recovers the prediction (causality), and where a small adapter can repair the damage (compensatory capacity) - and we show these three layer maps dissociate. Across a five-model panel we identify two propagation regimes - spike-and-suppress (Phi-3.5, Gemma-2-9B) and late-accumulation (Llama-3, Mistral, Qwen2.5-7B) - and on the two models meeting an 80% identity-patch gate, sensitivity and causality are anti-correlated (rho = -0.72 to -0.88). Within-family scaling on Qwen2.5 (1.5B to 14B) shows the late-accumulation signature strengthening monotonically with scale, corroborated on a second family. We propose cascade disruption as the mechanism behind the dissociation: adapters placed at causally implicated early layers break intact downstream computation, making diagnostic-flagged sites the worst adapter placements. A fixed-harness layer sweep across four models (3.8-8B) confirms the core prediction on chain-of-thought GSM8K - the flagged sites are the most damaging adapter windows on every adjudicable model - and is sign-consistent but strongly attenuated on a multiple-choice control, consistent with damage that compounds with generation length. The sweep yields practical guidance: a training-free LRD pre-screen and a default-deepest placement rule, though absolute gains over no-adapter baselines remain small. Finally, apparent gains from a representation-stability loss reverse under an adequate generation budget - truncated chain-of-thought had been scored as empty - a methodological warning for any intervention evaluated on chain-of-thought tasks.

While music generative models have recently gained significant attention, how they can be effectively integrated into live music performances still requires further exploration. This paper presents Calliphony, a calligraphy-driven interface for real-time generative music performance. Specifically, we build a low-latency pipeline that captures brush motion with an attachable sensor and maps it to control signals for real-time symbolic music generation. Using a generative model, the system produces multi-track MIDI in performance settings, while brush-derived control signals constrain event timing and activate additional musical layers. The generated melody is then extended with real-time harmony and additional voices, and finally rendered through a DAW for live staging. Calliphony contributes: (1) a performance-oriented prototype that uses calligraphic motion as an external control layer for a real-time symbolic music generation model, controlling note density, pitch constraints, and accompaniment-layer activation; and (2) a cross-modal performance scenario that extends calligraphy beyond a primarily visual practice into an audiovisual, AI-assisted setting.

Post-training can substantially alter language-model behavior, yet aggregate behavior rates do not reveal whether training removes an existing mechanism, creates a new one, or changes how an inherited mechanism is used. We study this question through two mechanistically distinct failures, repetition as a decoding-attractor pathology and sycophancy as a preference-related alignment failure. We introduce behavioral manifold analysis, which isolates behavior-specific geometry by selecting sparse behavior-associated coordinates and lifting them into low-dimensional local charts. We construct these charts in two complementary spaces. ACT captures runtime activation states, while NOC quantifies how strongly the model routes functional information flow through the shared behavior-associated subspace. Across multiple model families, the resulting charts are highly compressed and partially alignable across architectures. Contribution-space charts expose a more architecture-robust shared core, whereas activation-space charts retain stronger family-specific structure. Tracking these charts through controlled post-training reveals a consistent asymmetry. Supervised fine-tuning substantially alters the inherited behavioral geometry, whereas reward optimization changes behavior while largely preserving the underlying chart. This geometric perspective provides a unified framework for understanding the mechanistic distinction between the two objectives. SFT tends to rewrite behavioral geometry, whereas reward optimization primarily reweights it. Code is available at https://github.com/ronglingze/Manifold-Analysis

This work presents Fairness Pruning, a lightweight structural intervention method designed for the management and future mitigation of demographic bias in large language models (LLMs). As a foundational empirical validation of this method, this work focuses on causal bias localization. Using minimally contrastive prompt pairs and inference-time activation capture, the method identifies neurons that react differentially when processing demographic attributes in GLU architectures, evaluating the signal at the down_proj input. Empirical evaluation was conducted on models of up to 3 billion parameters (Llama-3.2 family and Salamandra-2B), combining standardized benchmark evaluation with qualitative text generation experiments. Results demonstrate that zeroing the identified neurons alters how the model responds to associated demographic variables. However, rather than producing flat mitigation, the intervention causes bidirectional bias destabilization: because BiasScore is unsigned, candidate sets mix neurons that push toward and against the stereotype, and the net effect on aggregate bias depends on which sign dominates. The intervention is extremely surgical: zeroing at most 40 neurons in Llama-3.2-1B (less than 0.031% of total MLP width) achieves a mean retention of 99.49% in reasoning and general knowledge capabilities. These findings empirically confirm that demographic bias processing and model capabilities operate on dissociable circuits, establishing the methodological foundations for transitioning from blind zeroing toward directional behavior modulation.

Activation steering controls model behavior by editing internal activations at inference time. We study its input-side dual: optimizing a fluent prompt so that a chosen internal latent is driven toward zero, with no inference-time model access. Our target is an "evaluation-awareness" latent-linearly readable and steerable in recent work-whose control would threaten the validity of safety evaluations if models behave differently when they detect being tested. Adapting Fluent Dreaming / EPO with a negated feature term (GCG-style token optimization plus a self-cross-entropy fluency regularizer, swept over a fluency weight), we suppress the latent under five target constructions-a CAA direction, a subspace norm, an SAE feature, a single MLP neuron, and a behavioral logit-on Llama-3.2-3B and Llama-3.1-8B. The latent is robustly suppressible ($z\approx-7$), and a causally-validated Llama Scope SAE feature can be fully and selectively turned off. But our controls tell a cautionary story about the CAA direction: a placebo random direction is suppressed just as hard and shifts behavior just as far, and when we hold a real eval passage in context and optimize only a prefix, suppressing the eval-direction fails to reduce-and slightly increases-the model's behavioral eval judgment. Activation-readability, in short, is not behavioral controllability. We further find that a single MLP neuron is eval-correlated but not causal at both scales, and that scanning the real Pile yields a natural-text baseline competitive with the optimizer for the internal direction. A positive control validates our erasure detector, bounding an erasure-vs-rotation question earlier left open.

Activation steering modifies a language model by adding a learned direction to its hidden activations, enabling targeted behavioral changes without retraining. While effective, steering often produces unintended side effects on other behaviors, making it difficult to deploy safely. We therefore ask: can these side effects be forecasted before steering is applied? We answer this question by constructing a cross-effect matrix over a taxonomy of 67 behaviors across three open-weight language models. We find that side effects are common, structured, and often asymmetric, revealing interactions that cannot be explained by existing similarity-based heuristics. Despite this complexity, we show that side effects are largely predictable before steering is performed. Their magnitude depends primarily on the target behavior, while their direction can be forecasted from the model's unsteered representations with substantially higher accuracy than simple baselines. Our results demonstrate that activation steering has systematic and forecastable side effects, enabling proactive safety auditing and more informed deployment of steering interventions.

Activation steering controls language models by adding vectors or features to hidden states at inference time, but the upstream source of these steering signals is often treated as a secondary detail. We study this source choice as activation source selection: the combination of source context and activation readout policy used to collect the hidden states from which a steering signal is built. Holding the downstream intervention fixed, we show across three instruction-tuned models and four steering task families that changing only the source activations substantially changes steering success. We further find that effective steering is not explained simply by whether the desired behavior appears in the source text. Instead, strong signals come from execution-boundary states, where the model is about to produce or continue the target behavior. This pre-/post-realization distinction explains why answer-based sources sometimes work: their useful component aligns with execution-boundary directions rather than target appearance alone. Building on this view, we introduce tail subtraction, which removes shared prompt and continuation semantics from boundary states and yields cleaner, more stable steering signals. Overall, our results suggest that steering depends on representations of what the model is about to do, not merely on what has already appeared.

Activation Oracles (AOs) are language models trained to answer natural-language questions about another model's internal activations. They offer a flexible interface for reading hidden information from model states, especially when relevant information is internally represented but absent or incomplete in visible behavior. However, AOs are themselves learned systems: their answers are shaped by training data, objectives, and learned reporting behavior, rather than being neutral readouts of represented information. We study this in a controlled Taboo Word Guessing setting, where subject models are fine-tuned to internally use a hidden concept while avoiding direct disclosure. Contrary to the expectation that an AO trained on such a subject becomes a specialist reader, we find that fine-tuned AOs can become concept-specific anti-readers: they selectively fail to recover the concept persistently present during their own training. This failure is not simply explained by absence of the concept from the subject or oracle representations: the target remains decodable inside the oracle, while LogitLens and layer-ablation analyses indicate that the failure arises in the AO readout pathway. Our results show that behavioral leakage, representation-level decodability, and AO-verbalizability can come apart, raising a reliability concern for learned interpretability interfaces.

Post-training quantization (PTQ) has become a practical solution for deploying deep learning models on resource-constrained edge devices by compressing high-precision floating-point weights into low-precision representations without requiring retraining. Past research has demonstrated that quantization largely preserves classification accuracy; however, whether it also preserves the model's internal reasoning remains an open question. This study presents a systematic evaluation on how static PTQ affects the interpretability / explainability of five widely used CNN architectures: VGG19, ResNet18, EfficientNet-B0, DenseNet161, and MobileNetV2 at INT8 and INT4 precision. We employ a dual interpretability framework that combines Grad-CAM for spatial attention analysis with LIME for input-level feature attribution, and systematically compare full-precision and quantized models on two binary classification datasets. Interpretability is evaluated using three complementary metrics: the Pearson correlation coefficient, structural similarity index, and top-20% IoU to capture distributional and structural variations in model explanations, supplemented by deletion/insertion faithfulness analysis. The results show that classification accuracy is not a reliable indicator of interpretability stability under reduced precision. DenseNet161 maintains strong feature consistency across both precision levels, whereas EfficientNet-B0, despite achieving competitive spatial attention and classification accuracy at INT8 precision, exhibits a substantial degradation in input-level feature attribution. These findings have direct implications for the trustworthy deployment of quantized models in applications with high interpretability requirements, demonstrating that architecture selection is as important as the quantization strategy.

Neural network misclassifications exhibit characteristic spectral instability in internal activations that is invisible at the output layer. This phenomenon is identified and formalized as Spectral Drift -- the frequency-domain distance between consecutive layer activations -- with empirical validation showing that failures exhibit significantly higher drift than correct predictions (1.9% increase, p<0.001). This spectral signature emerges during internal processing but becomes masked in final outputs, explaining why confidence-based detection methods struggle. This work introduces Self-Detecting Neural Networks (SDNN), a framework that monitors spectral dynamics across network depth using Short-Time Fourier Transform, wavelet decomposition, and statistical moments to capture multi-scale spectral features. A lightweight detector network (5% parameter overhead) learns to identify failure-indicative patterns via curriculum learning on progressively challenging distributions: natural misclassifications, distribution shifts, and adversarial perturbations. Experiments on CIFAR-10 demonstrate that SDNN achieves 79.0 +/- 25.3% AUROC across three seeds, substantially outperforming confidence-based baselines including MaxSoftmax (50.5%) and Energy Score (52.9%) by approximately 25-30 percentage points. Ablation studies reveal that wavelet decomposition and statistical features make consistent contributions, while STFT's role remains unclear. This work establishes spectral analysis of internal activations as a promising direction for neural network reliability, revealing diagnostic information inaccessible to output-based approaches.

Activation steering provides a lightweight mechanism for controlling large language models at inference time, but steering vectors can have unintended externalities: utility vectors may weaken safety behavior, while refusal vectors may induce over-refusal on benign prompts. We introduce OPIUM (Optimizing Protected Injections via Utility Manifolds), a training-free method for sanitizing steering vectors through representation matching. Given reference behaviors on two prompt sets, OPIUM optimizes a new steering vector that preserves the downstream representations induced by the desired intervention while matching a safer reference behavior on prompts where the original vector fails. Across steering-externality and over-refusal settings, OPIUM improves the safety--utility tradeoff relative to vanilla steering and directional ablation, suggesting that harmful side effects of activation steering can often be mitigated directly in activation space.

Genomic language models achieve strong performance across regulatory-genomics tasks, yet what these models internally represent remains opaque, and the field lacks a principled procedure for verifying that an apparent concept'' inside a model is real rather than an artifact of sequence composition. We introduce a framework that combines sparse dictionary learning with causal intervention to extract, validate, and causally test interpretable features in genomic foundation models. Training top-$k$ sparse autoencoders on the hidden activations of two architecturally distinct models, Nucleotide Transformer ($6$-mer tokenization) and DNABERT-2 (byte-pair encoding), we recover thousands of monosemantic features that map to transcription-factor (TF) sequence motifs. We show that the naive validation of such features against position weight matrices is severely confounded by GC composition and repetitive elements, producing hundreds of spurious TF features'', and we develop a composition-matched, binding-resolved protocol that removes these confounds. Critically, we move beyond correlation: by ablating individual dictionary directions during the model's forward pass and measuring the induced shift in the model's own predictive distribution, we establish that specific features are \emph{causally} used to represent cell-type-specific TF binding, not merely motif presence. Across three transcription factors (CTCF, GATA1, REST) and both architectures, causally validated binding features emerge reproducibly ($7$--$14$ of $15$ tested features per condition), while two classes of negative control, scrambled binding labels and randomly selected features, yield no detectable signal. The framework is purely computational, uses only public data, and provides a reusable standard for interpretability claims in genomic deep learning.

Interpretability methods for neural network activations span a wide cost spectrum, from cheap, training-free techniques (such as linear probes, PCA, SVD) to more expensive training-based ones (such as SAEs and activation oracles). Training-based methods are typically more powerful, in part because they leverage large activation datasets during training. This raises a natural question - do they actually surface insights that go beyond what is recoverable from the training dataset itself? To address this, we equip an LLM agent with a vector database of activations paired with their textual contexts, along with tools for manipulating activations - projecting out directions in latent space, computing activation differences and averages. The agent iteratively queries the database, forms hypotheses from the retrieved samples, and validates them by constructing linear probes. We call this method HARP, for Hypothesis-driven Agentic Retrieval and Probing. Despite not involving any training, HARP outperforms both activation oracles and SAE-based agents on concept discovery, concept detection, model steering, and secret elicitation. The training-free design also makes HARP substantially cheaper and more flexible: new datasets can be indexed on demand whenever existing ones prove insufficient. More broadly, our results suggest that current training-based methods do not yet extract insights beyond their training data, and motivate benchmarks that explicitly require interpretability methods to demonstrate such insights. We release our code at https://github.com/SriramB-98/HARP

Transcoders have recently emerged as a promising approach for mechanistic interpretability (MI), enabling circuit-level analysis of model behaviour. In this paper, we investigate the use of transcoders to analyse deceptive behaviour in language models, a behaviour that poses a safety and security risk. Using a Qwen3-4B model with pre-trained transcoders, specifically per-layer transcoders (PLTs), we construct attribution graphs that capture feature activations and inter-feature dependencies, allowing circuit-level analysis of deception. Through feature steering and circuit analysis, we identified a dictionary of deception-related features and show that these features exert a stronger influence on deceptive outputs, as they produce predictable shifts between deceptive and non-deceptive responses. These findings suggest that deception emerges from internal model mechanisms and highlight the potential of transcoders for behavioural monitoring and early detection of security vulnerabilities related to malicious behaviours in language models.

Activation steering offers a lightweight alternative to fine-tuning for controlling large language models at inference time. While many existing methods implicitly optimize a log-density-ratio objective between desired and undesired activation distributions, they do so heuristically rather than deriving it from a principled optimization problem. Moreover, these methods produce query-independent steering directions that can degrade performance on both in-distribution and out-of-distribution (OOD) inputs. We introduce \textsc{Cobras} (Conditional Optimal Bridge for Riemannian Activation Steering), which addresses both limitations by casting activation steering as a Schrödinger Bridge on the residual-stream hypersphere. This formulation yields, to our knowledge, the first principled derivation of the log-density-ratio steering objective from a well-posed optimization problem. Solving the bridge via entropic optimal transport and extracting the probability flow ODE recovers the widely used density-ratio gradient as a special case when the Sinkhorn potentials are uniform. Crucially, the Schrödinger potentials are evaluated at the current activation, making the resulting steering direction inherently query-adaptive. Empirically, across four models and three alignment axes (helpfulness, truthfulness, and detoxification), \textsc{Cobras} consistently outperforms prior activation steering baselines while avoiding the OOD degradation commonly observed in existing methods. The code can be found at https://github.com/arshandalili/cobras.

Additive activation steering (injecting a scaled residual-stream direction during generation) is calibrated almost entirely in single-turn chat, yet the models it targets are increasingly deployed as tool-using ReAct agents. We present the first systematic chat-to-agent transfer study of additive steering, coupling behavioral measurement with a representation read-out in a matched-information design: the same items rendered as plain chat or as a ReAct tool-use episode, with matched-norm random-direction controls and the transcript re-encoded every turn to exclude KV-cache contamination. Transfer is real but rescaled, and the right description is a dissociation: the injected direction reaches the late layers at near-full strength in every setting and model tested (install-site agent-over-chat ratios 0.83-1.16 across three families), while the behavioral coupling is reset per model and context. On Qwen2.5-7B a refusal bypass vector amplifies in the agent (T = 1.45, CI [1.20, 1.78], N = 300); across a powered uniform-protocol distribution the coupling spans amplification (Gemma-2-9B T = 2.00) to attenuation (Yi-1.5-9B T = 0.43, CI [0.29, 0.60]), with no universal constant and a single clean attenuator against a universal sign. Directional ablation of the same axis does not amplify (T = 0.93, CI including 1) while additive injection amplifies (T = 1.50), a 20.1-point gain difference (CI [13.4, 26.8]) that identifies an additive-specific mechanism. Two pre-registered instruments converge to localize the rescaling to the ReAct format scaffold, before any tool observation, rather than to the observation boundary where a dilution account would predict it. The safety implication is immediate and unpredictable: agentic deployment amplifies steering-based refusal bypass by up to 2.00x on some models while others attenuate, so a deployment cannot assume a given model is safe under additive steering.

A transformer can be built from operators that are legible by construction -- bounded, named units that read as fuzzy set operations rather than dense activations -- but legibility must be pressed for during training, and the pressure has a failure mode. A crispness penalty meant to sharpen a bounded operator into a decisive detector instead collapses it into a dead constant. An identity, E[v(1-v)] = mu(1-mu) - var, shows why -- the penalty is a variance-minimizer blind to the difference between a live detector and a constant -- and names the fix: a per-channel variance floor, the target legibility metric written as a loss, which recovers both legibility and quality. A learned per-unit fraction then retires the hand-set reserved-GELU partition of prior work: given the choice the model keeps no unit as pure GELU and routes 87% of its load-bearing computation through crisp operators. The result is the most legible transformer we have built -- 78% of its feed-forward operands and 50% of its attention value channels are crisp-and-contextual detectors, and per-head legibility rises from 18% in shallow layers to 78% in deep ones. Read in the correct rotated per-layer frame, these units separate a clean detection (what a unit responds to) from a harder naming (what its output decodes to); and because the objective makes each unit crisp and sparse, edits to them are far more local -- 50-184x in the deep layers where the edit sites concentrate -- and can target explicit conjunctions a single neuron cannot express. Finally, a between-unit decorrelation pressure exposes a legibility dial: it trades a circuit's reuse for independence at no quality cost, turning concepts into single, surgically editable units and a prediction into a short explanation read off a handful of named operations. Quality holds at parity with a conventional baseline throughout.

Large language models process prompts by propagating activations through dozens of layers before generating a response. We ask whether the task-relevant information contained in an instruction prompt can be compressed into a single activation vector and re-injected into the model, replacing the original token sequence? We show this is achievable using a learned weighted sum of activations extracted at an intermediate layer and injected at an early layer of the target LLM. The compressed vector preserves task-relevant information, incurring an accuracy drop of under $2\%$ relative to full prompt processing. Beyond its practical implications, including reducing per-query computation for fixed instruction prompts without reprocessing the original token sequence, our analysis reveals structure in the activation space of LLMs: (i) mid-layer representations transfer meaningfully to early layers, suggesting a degree of cross-layer compatibility in how information is encoded; (ii) a single activation vector encodes a quantifiable and recoverable amount of semantic information; (iii) a weighted sum of activations is a robust representation compressor.

As part of a series on first-principles modeling of cognitive functions, this paper attempts to provide a mathematical formulation of thinking and perception. It formally derives slow thinking or more generally, active perception, and encompasses the design, training and inference of slow thinking large language models. Our starting point is the lifting and projection of probability distributions on the observable and latent spaces, with the objective of representing complex data distributions by simple function families such as neural networks. A theory called "active lifting" is proposed, based on the sampling of latent sequences and an intrinsic drive to reduce uncertainty with maximum rate. It derives a large design space, containing the slow thinking models in a subspace that we call the static theory. These models are positioned on the representation hierarchy and sampler hierarchy induced by the static theory, and can be upgraded by climbing the two hierarchies. Active lifting further derives an inference process with an internal time axis, and a training objective that resembles minimum-length coding as well as the invention of languages. Thus, it characterizes the agency of perception, including the emergence of the slow thinking formats. Technical by-products of this theory include a three-stage pathway for improving slow thinking models, a unified approach to constructing encoders and generative models for all data modalities, a priori formation of human-like visual representations, and a possible solution to policy collapse.

We study how correctness is assembled inside aligned language models, not only whether the final answer is right. Using layer-wise difference-in-differences (DiD) trajectories over polarity-controlled minimal pairs, we identify the wrong-dip: in mid layers (25-90% depth), internal preference transiently commits to the incorrect answer and is rescued only by late-layer correction. We verify this causally with patchscope-style activation transplantation across 17 models, three families, and 64x scale (0.5B-32B). Four findings follow. (1) Alignment amplification of the causal wrong-dip is recipe-specific and emergent: it emerges at 3B in Qwen2.5, remains high, and peaks at 32B (paired t up to 9.7), reverses in Llama-3-8B (t=-2.31), and sits between for Mistral-7B. (2) The dip predicts real compression failures: high-dip items are 3-7x more likely to flip under late-layer low-rank compression, block dropping, or structured pruning, while quantization flips are dip-blind, a double dissociation confirmed by late-layer ablation. (3) The dip is trainable: a LoRA fine-tune with a mid-layer wrong-margin penalty matches output-only SFT accuracy while cutting the causal dip by 67-70% and improving compression robustness; output-only SFT worsens the causal dip by up to 2.8x at perfect surface accuracy. (4) With controlled readouts, the phenomenon survives natural-language I/O: dip stratification of structural-damage failures is significant on naturalistic vignettes, and free-form fragility separates into a dip-auditable late-rescue layer and a dip-blind interface layer. Together, output-level correctness can hide a late-rescue process that governs compression risk, post-training quality, and evaluation distortion.

Interpretability methods aim to reveal the features represented inside large language models (LLMs). Many existing methods begin with labeled examples of a human-defined concept that may reflect human biases, and then identify how that concept is represented within the model, for example in its activation space or through other decomposition methods. We introduce \emph{Mining via Activation Geometry} (MAG), a simple unsupervised framework for extracting reasoning features from model activations by prepending the same natural-language instruction $Q$ to every input $p$, where $Q$ defines the reasoning feature of interest, such as Can this object be found in the desert?'' or Is this prompt malicious?'' We measure how the instruction changes the model's internal representation using $m(Q \mid p) - m(p)$ at a single readout point. We explore eight different MAGs. The extracted reasoning features predict the models' own world understanding and judgment, can be approximated into a single activation direction, we found that some features are more linearly represented and some less, this linear representation, which is vector steering, can change the LLMs' decisions through activation steering by injecting reasoning features. Finally, we use the same method to select the best training datasets for prompt-injection classifier probes: while similarity between ordinary activations is almost unrelated to downstream performance, RFD-based similarity achieves $94.7\%$ Top-1 and $100\%$ Top-2 accuracy.

Large language models (LLMs) are increasingly deployed in domains requiring guardrails to detect unsafe, off-topic, or adversarial prompts. Existing guardrails predominantly rely on fine-tuning to build classifiers, which often suffer from low generalization and high inference latency. We present kNNGuard, a training-free guardrail that utilizes the activation space of an off-the-shelf LLM. Given a small bank of 50 safe and unsafe prompts, kNNGuard extracts hidden activations and performs multi-layer kNN fusing activation-space and embedding-space scores for classification. Across six domains spanning topical and security prompts, kNNGuard achieves competitive or superior F1 compared to fine-tuned state-of-the-art guardrails while running 2.7x faster than the best comparable guardrail, and 10x faster than a fine-tuned safety classifier without gradient updates or fine-tuning. Domain adaptation requires only updating the labeled bank, which can be constructed in under 10 seconds and several orders of magnitude faster than established guardrails. We also analyze the impact of system prompts, layer selection, and integration into production LLM pipelines as a configurable, low-latency guardrail.

Model organisms (MOs) - language models trained to exhibit undesired or unnatural behaviours - are frequently used as testbeds for evaluating white-box interpretability techniques. Current MOs are typically constructed via post-hoc supervised fine-tuning (SFT) on behavioural transcripts or synthetic documents. Prior research has shown that interpretability methods can easily identify hidden behaviours in these MOs. However, recent work suggests that such post-hoc training methods may make interpretability unrealistically easy. We investigate this claim by constructing a suite of 54 $\verb|OLMo2-1B|$- and $\verb|gemma-3-1b-it|$-based MOs trained with seven different techniques, including standard post-hoc SFT, post-hoc DPO, and more realistic integration of MO data into the OLMo post-training DPO phase. We use these MO variants to benchmark activation oracles, activation steering, logit lens, and sparse autoencoders. Our findings show that (i) MO interpretability depends strongly on training objective, target behaviour, model architecture, and training data generation pipeline; (ii) substantial variance remains even after controlling for differences in the strength of target behaviour expression; and (iii) our more realistic $\textit{integrated training}$ often yields less interpretable MOs than standard post-hoc methods. Our results cast substantial doubt on the validity of current MOs as interpretability proxies.

Sleeper agents are the canonical model organism of deception: models trained to behave normally but to emit an unsafe behaviour on a specific trigger. Eliciting that behaviour without knowing the trigger has not been studied systematically. We study fuzzing: injecting Gaussian noise into a model's weights or residual-stream activations and checking whether the perturbed outputs reveal the behaviour. On 6 backdoored models (7B-13B) we compare both forms of fuzzing head-to-head against temperature-sampling baselines. Fuzzing elicits the hidden behaviour more often than temperature sampling on 4 of 6 models (up to ~6x on OpenHermes-13B), and which form wins depends on the task, so both are worth running. Elicitation is uneven across each method's hyperparameter grid: a uniform sweep gives only a few percent on most models, while the best cell is 2-10x higher, so the bottleneck is hyperparameter selection, not the technique. To select hyperparameters without ground-truth access, we use a cheap proxy task (in-context secret elicitation, where a base64-encoded secret is placed in the system prompt for the model to hide) and run Thompson sampling on it to pick candidate cells, which we evaluate on the real backdoor. On the four models that can decode the secret, proxy-selected cells raise activation-fuzzing elicitation ~4x over the uniform-sweep mean (recovering ~70% of the best-cell rate on the best performing model) and weight-fuzzing by 1.3-1.8x. To our knowledge this is the first systematic study of fuzzing on sleeper-agent backdoors and the first to show proxy-task hyperparameter selection transferring to real-task elicitation. We also propose reporting such results as a (uniform-baseline, proxy-selected, oracle) triple, since these are three distinct claims that prior work has often blurred.

We aim to discover diverse, generalizable perturbations of LLM internals that can surface hidden behavioral modes. Such perturbations could help reshape model behavior and systematically evaluate potential risks. We introduce Causal Perturbative Elicitation (CPE), an unsupervised method for discovering interpretable low-rank adapters (LoRAs) that can elicit these latent behaviors. CPE decomposes the computations of a deep transformer slice using a heuristic tensor-decomposition-based algorithm. CPE exhibits remarkable data efficiency, learning a large number of interpretable LoRAs from a single example. Even though CPE is unsupervised, we find that in some cases it can be competitive with supervised elicitation methods via brute-force enumerative search over weight space. For instance, CPE performs similarly to matched-wall-clock-time GRPO on the Countdown task for Qwen3-8B (85% vs 87%), demonstrating that CPE can efficiently elicit complex multi-token behaviors. Since CPE is unsupervised, it can also surface hidden failure modes, such as sandbagging, restoring 85% of locked BigCodeBench performance on a password-locked version of Llama3-70B introduced by Taylor et al. (2025). Additionally, since CPE explores behaviors in weight-space rather than token-space it can potentially ameliorate exploration hacking, a misalignment failure which may arise in sufficiently self-aware AI models (Ngo, 2022). In fact, we find that CPE virtually eliminates alignment-faking (Greenblatt et al., 2024) behavior in a Llama3-70B-based model organism developed by Hughes et al. (2025). Finally, we find that CPE can be used to initialize GPT-OSS-20B in an aligned basin when running GRPO on an environment prone to reward-hacking. By providing a data-efficient method to systematically explore the space of latent model behaviors, CPE yields a powerful tool for aligning AI systems and evaluating their safety.

Do language models know when they are being tested? This question matters for AI safety: a model that recognises an evaluation context could alter its behaviour strategically, making downstream benchmarks harder to interpret. Using 11 models spanning Qwen 2.5, Gemma 2, and Llama 3.2, we find a systematic size-dependent shift in representational depth: in both Qwen 2.5 and Gemma 2, the layer at which evaluation-awareness is most linearly recoverable moves from late layers in smaller models to early layers in larger ones. This suggests that scale changes not only the strength of evaluation-awareness but also where it is most linearly recoverable in the network. This depth shift helps explain why within-family scaling trajectories are non-monotonic or inverse rather than smooth and family-general, showing that a simple universal power-law account is not supported under denser within-family sampling. Finally, white-box probe signals are consistently stronger than black-box behavioural expression, and the relationship between the two varies by family in ways not predicted by probe AUROC alone.

While existing data attribution methods can identify which training examples build specific mechanistic circuits, they cannot explain how training data shapes the high-level behavioral decisions a model learns to make. To bridge this gap, we introduce Symbolic Mechanistic Data Attribution (SMDA), a framework that attributes training pairs to the interpretable symbolic policies governing model behavior. SMDA fits a closed-form Ridge regression over sparse autoencoder (SAE) features to model a target behavior, then analytically decomposes how each supervised fine-tuning example shifts that policy through feature-activation Delta_X and output-probability Delta_Y pathways. We distill a symbolic policy for refusal behavior in Llama-3.2-3B-Instruct and analyze 200 SFT training pairs. Our analysis reveals that (1) the symbolic policy's coefficients expose systematic gaps in the base model's safety behavior for categories like religious stereotyping; (2) per-feature Delta_X/Delta_Y decomposition can mechanistically explain why harmful and harmless pairs exert qualitatively different influences on certain features; and (3) individual training pairs routinely exhibit cross-feature interference, allowing SMDA to identify training pairs whose dominant effect falls on unintended features. These results demonstrate that combining mechanistic interpretability with data attribution yields a diagnostic tool that is both more fine-grained than black-box influence functions and more scalable than manual circuit analysis.

Sparse autoencoders (SAEs) have become a useful tool for extracting interpretable features in language models. However, standard SAE architectures operate on individual token activations, meaning that the number of active features scales linearly with context length, and studying long model transcripts becomes difficult. We introduce turn-averaged SAEs, which represent a single Human or Assistant turn with a fixed number of features by learning to reconstruct the average model activation across the turn. We find that turn-averaged features describe a single turn's high-level characteristics more completely than per-token features when judged by an LLM. We also demonstrate that turn-averaged SAEs greatly simplify common downstream uses of SAEs like attribution graphs. Broadly, turn-averaged SAEs make interpretability techniques practical at long context lengths.

Activation patching is the primary tool in mechanistic interpretability. It attributes causal responsibility for a model behavior to each of its individual components by estimating its natural indirect effect (NIE). Re-deriving the activation patching estimand from causal mediation analysis, we find that the NIE does not solely capture the causal effect through the specific component. It also contains interaction effects (INT) that measure how much the component's causal effect itself depends on the state of other components in the model. A natural response may be to try to eliminate INT by adjusting the estimator or unit of analysis, but each of these potential remedies has predictable failure modes. We demonstrate these failure modes in the GPT-2 IOI circuit; components whose causal importance is conditional on the state of other components are either invisible or artificially inflated, and INT variance explains the previously documented instability of faithfulness scores. We prove that INT scales with the distance between clean and patched component activations, is negligible when the model is locally affine, and decomposes combinatorially into pairwise and higher-order group interactions. Despite its inevitability, INT is not a nuisance to be eliminated, but rather a diagnostic for interpretability studies. Its individual and group-level magnitude and sign signal when causal conclusions are prompt-dependent, and when greedy NIE-based component ranking will miss mechanisms only discoverable through combinatorial search.

Distributed intelligent systems increasingly need to train across data silos without centralizing raw data. Federated learning keeps data local but can suffer under heterogeneous partitions and requires repeated full-model exchange. Split learning reduces communication through cut-layer activations, but standard protocols generally do not recover centralized mini-batch gradient behavior and may expose activations and gradients in plaintext. We present TL++, a two-mode traversal-learning framework that constructs virtual batches across nodes to recover centralized mini-batch gradient behavior under explicit synchronization assumptions. Base mode exchanges cut-layer activations and gradients rather than full models. Secure mode secret-shares each cut-layer activation and gradient between an orchestrator and a non-colluding helper, preventing either server from observing plaintext cut-layer tensors. This protection is limited to a semi-honest two-server setting; labels and loss-related outputs remain visible to the orchestrator. In the lightweight secure path evaluated here, exactness requires a linear or affine server path, while nonlinear operations require nonlinear MPC or approximation. We formalize TL++, analyze communication and computation costs, and evaluate it against federated and split-learning baselines on CIFAR-10 and BioGPT/PubMedQA using full fine-tuning and LoRA. On CIFAR-10, TL++ base cut 1 and exact secure cut 3 achieve accuracies of 91.41% (SD 0.19) and 90.93% (SD 0.17), respectively, exceeding the strongest measured non-TL++ baseline by more than 12 percentage points. TL++ base cut 1 also reduces per-step communication by 13.1-fold relative to full-model synchronization. PubMedQA results similarly favor TL++. Overall, TL++ approaches centralized-training performance while reducing communication and providing activation-level secret sharing.

Jailbreak attacks reveal a persistent weakness in aligned Large Language Models: carefully crafted prompts can elicit policy-violating responses despite safety training. While most defenses operate at the prompt or output level, it remains unclear how harmful intent is encoded within the model's internal representations. We investigate this question by analyzing token-level predictive entropy trajectories across layers of a frozen LLM using the logit lens. We find that static aggregate statistics of prompt-level entropy (e.g., mean, variance) carry little discriminative signal, whereas features capturing how entropy evolves across token positions, such as monotonic rank-based trend scores, are substantially more informative. Importantly, this signal is not uniform across model depth: it is concentrated in intermediate layers and degrades at the final layer, indicating that jailbreak-relevant structure is most pronounced in mid-network representations rather than at the output head. Across multiple models (Llama, Qwen, Gemma) and adversarial benchmarks, these entropy dynamics provide architecture-consistent separation without additional training. Together, our findings show that jailbreak behavior is reflected in structured intermediate uncertainty dynamics, clarifying both which entropy-derived features encode harmful intent and where in the network that signal is most pronounced.

Interpreting and controlling model behaviors through activation steering methods requires many pairs of contrastive samples that clearly exhibit desired or undesired behavior. These data pairs determine the degree to which interpretability frameworks can reliably detect model features responsible for a behavior, and therefore the ability to steer models toward or away from such behavior. In this work, we present an iterative data generation pipeline that isolates cascading linear features responsible for a behavior. Specifically, we show how moving beyond simple binary pairs of samples, and instead isolating samples that show degrees of features that scale linearly with behavior, allows for better disentanglement of features. We focus on detecting and steering away from sycophancy -- the tendency of language models to prioritize user validation. We demonstrate that sycophancy features discovered through cascading samples form linearly separable subspaces, and allow for selection of model activations that more clearly correspond to the desired behavior than baseline approaches. We also evaluate their ability to enable detection, deterministic scoring, and robust steering, and see that they either match or outperform LLM-as-a-judge and system prompting baselines while providing lower computational demand and more interpretability guarantees. Code & Data: https://cascading-feats.github.io/

A central aspiration of mechanistic interpretability is controllability: if we know where a behavior is represented in a model's activations, we should be able to modify it. This rests on a hidden premise -- that the direction which detects a behavior and the direction which controls it are the same, or close. We test this geometrically: what is the angle between the direction that best detects a behavior and the one that best causes it? If detection implies control the cosine is near 1; otherwise it quantifies a detection-intervention gap. On Gemma 2-2B-it, output format (clean JSON vs markdown fencing) collapses both roles onto one axis. Hallucination does not: the model detects fake entities with perfect linear separability (AUC = 1.000 from layer 5), yet that direction sits at cos = 0.12 (about 83 degrees) from the direction producing a refusal -- a small, reproducible alignment, far from the cos = 1 that "detection is control" would require. A detector built from activations, with no chosen tokens, likewise fails to align (cos = -0.06). The gap generalizes: across four models from three families and two scales (1B-9B), cos stays in [0.12, 0.20], identical before and after instruction tuning (0.1197 vs 0.1200), placing its origin in pretraining. A 15-degree rotation toward the refusal direction partially bridges it -- 73% and 60% refusal on two held-out fake-entity categories at 1.8% false positives. We then ask whether this cosine predicts steerability, and it does not: detection is a high-dimensional class, not a single direction, and what separates the steerable case is functional, not readable from a static angle. The cosine is a weight-computable signature of the dissociation between knowing and steering, not a predictor of it.

Vision Transformers have achieved remarkable success in many fields, yet their deployment on edge devices remains challenging due to their substantial computational demands. Post-Training Quantization (PTQ) offers an attractive solution by compressing models using a small calibration set with minimal training overhead. However, most existing PTQ works adopt a static quantization paradigm that is uniformly applied to all instances. Given the substantial diversity of natural images, the activation distributions vary significantly across samples, making these methods inherently suboptimal. In this paper, we propose ScalePredictor, a dynamic quantization framework for accurate and efficient quantization scale learning of ViTs. We first reveal a hidden correlation between the distribution range of shallow-layer activations and the optimal scales of deeper layers. Based on this, we develop a scale learning mechanism that integrates an efficient range extraction approach to capture robust range statistics at the shallow stage, which are then fed into a Taylor-motivated polynomial scale projection module to generate all quantization scales simultaneously. With the efficiency of polynomial approximation, ScalePredictor introduces insignificant computational overhead while avoiding costly just-in-time calibration. Extensive experiments on ImageNet demonstrate that ScalePredictor consistently outperforms prior PTQ methods, achieving a more favorable accuracy-efficiency trade-off. Code and additional results are shown in the supplementary materials.

Transformers achieve strong performance, but their internal computations remain opaque. We view each Transformer layer as a dynamic graph whose nodes are token representations and per-head attention outputs, with Multi-Head Attention (ATT) and MLP as module boundaries. On this graph we use LIG (Layer-wise Integrated Gradients), which applies set-to-set Integrated Gradients (IG) at nonlinear module boundaries. Set-to-set IG applies IG to a map from a set of input token representations to a set of output representations, evaluating token-to-token contributions, which is not standard in prior IG applications. This extends IG from the usual scalar-objective setting to set-to-set maps via an L2 scalarization, and composes within-layer contributions in the spirit of Layer-wise Relevance Propagation (LRP), with IG completeness playing the role of LRP-style conservation at each boundary. We use LIG to analyze (i) the agreement between module-wise composition and layer-whole attribution under an L2 criterion, and (ii) within-layer information flow by tracing separated ATT and MLP contributions. On BERT-base and PTB, configurations that best preserved within-layer consistency used the target token's embedding as the ATT baseline and either the ATT output at a=0 or Zero as the MLP baseline. We therefore present LIG as a diagnostic XAI tool at module-boundary granularity, without model-specific retraining or per-operation interpreter design. Code is available at https://github.com/eightsuzuki/layer-wise-integrated-gradients.

Fine-tuning language models on insecure code induces emergent misalignment with poorly understood internal structure. We investigate whether this misalignment corresponds to a causally actionable activation-space direction shared across architectures. Across four instruction-tuned model families (Qwen2.5-1.5B, Gemma-2-2B, Llama-3.2-1B, Ministral-3-3B) finetuned identically, a difference-in-means direction achieves 99.6% separation of aligned and misaligned activations at each model's final layer. Causal steering by subtracting this direction reduces code spillover by 21-51 points, while a secure-code control confirms content specificity. Cross-architecture transfer via ridge regression maps yields large behavioral suppression (up to 46 points) but fails specificity controls as random and orthogonal directions perform comparably. We identify a two-tier specificity structure: within-model directions are causally specific and actionable; cross-model directions are causally real but non-specific. An asymmetric transfer topology emerges, with Gemma and Qwen acting as geometric donors and Llama as a receiver. These findings define the limits of linear cross-architecture correction and recommend within-model probing for auditing.

Aligned language models gate behaviors such as refusal and language routing through sparse feed forward neurons, yet no theory predicts when a single neuron intervention controls a behavior coherently rather than collapsing the output. We develop a budget normalized control window framework for single neuron steering. A dose along one write direction reduces to one control coordinate: the alignment between the residual stream and the write, driven along a universal saturation curve in units of a coherence budget set by the residual norm divided by the write norm. Coherent control exists when a behavior trigger lies below the collapse ceiling. The same coordinate governs benign mode switches and refusal; the ceiling follows from weights and one generic forward pass, while triggers are measured at rollout. On fifteen held out neurons, the predicted ceiling has mean absolute error 0.14, about 0.07 in bulk layers, and the committed open or closed verdict holds on eleven against a ten of fifteen majority baseline. Closed cases expose three failure modes rather than violations: collapse before trigger, too little depth to propagate, or a normalization that caps how far one neuron can push. The law explains why local gradient attribution anti predicts control: true controllers write off the readout axis and carry a near zero first order gradient. A forward only contrastive screen made precise by the window recovers controllers that attribution misses. On refusal, the hardest case, intervention success is typed, not scalar: coherent bypass and strict actionable reach separate, so a neuron can flip refusal in fluent, on task text with no actionable content, and genuine actionable reach appears only for three of six audited Llama pivots and only at later rollout horizons. Single neuron steering is therefore a budgeted, typed audit of controllability rather than a fixed dose anecdote.

Trained transformers reliably develop massive activations, a small number of hidden dimensions whose magnitude is far above the median and which concentrate on the sequence-start token. Whether these outliers are a removable artifact of the residual stream's overloaded read and write role, or instead a functional necessity, is actively debated. We test the artifact hypothesis directly, with an architectural intervention. Our architecture, Ledger Residuals, splits the residual stream into a mutable scratch stream (Deliberation) that intermediate computation may freely overwrite and a protected, decode-only accumulator (Commitment) that holds the representation the model reads out. If massive activations exist only because one stream is forced to be both scratchpad and answer, then a dedicated answer channel should remove the need for them. We find that it does not. In matched-loss language models at the 160M and 290M scales, the model rebuilds the canonical fixed-dimension, start-token outlier inside the protected channel. The rebuilt feature is smaller in magnitude than in a standard transformer but more sharply concentrated on the start token, and a stronger sparsity penalty makes it more persistent and more concentrated still, rather than removing it. Massive activations therefore look architecturally robust: they re-emerge in whichever representation the model decodes from, which is what we would expect if they are functional rather than incidental. We release our architecture and measurement code.

Large language models are commonly aligned through supervised fine-tuning, yet little is known about how their internal representations evolve during this process. We study alignment dynamics using persistent homology by tracking the topology of activation spaces throughout fine-tuning. Across four transformer language models ranging from 1B to 7B parameters and three alignment objectives corresponding to helpful, harmless, and mixed training data, we find that the majority of topological reorganization occurs during the earliest stages of training. A dense checkpoint analysis reveals a transient peak in topological activity followed by rapid stabilization. We further show that different alignment objectives induce distinguishable topological trajectories, while instruction-tuned and pretrained models exhibit qualitatively different patterns of evolution. Our results suggest that persistent homology provides a complementary perspective on alignment, revealing representation-level changes that are not apparent from behavioral metrics alone.

While Audio Large Language Models (Audio LLMs) excel at multimodal understanding, they suffer from text dominance, a bias where models blindly favor text over acoustic evidence, causing hallucinations. However, the internal mechanisms underlying how these models behave when audio and textual inputs contradict each other remain unexplored. In this work, we present the first mechanistic analysis of this phenomenon by tracing the propagation of internal representations across layers. Our investigation reveals three key findings: (i) text dominance is systematically and empirically across models; (ii) while text and audio rely on functionally distinct pathways, they ultimately converge into a shared semantic space in late layers; and (iii) the text pathway does not erase audio information, but rather actively suppresses intact audio representations. Building on these insights, we leverage back-patching, a training-free intervention that routes late-layer audio activations back into earlier layers. This amplifies the audio representations, enabling them to overcome textual suppression. Our evaluation shows that back-patching consistently reduces text dominance, paving the way for mechanistic multimodal alignment under conflict.

Sparse autoencoders (SAEs) are increasingly used to extract interpretable features from language models (LMs), yet a central question remains: when can an SAE-based explanation be treated as a faithful view of an underlying frozen LM We study this through a post-hoc generalization framework that certifies the LM via a sparse proxy, obtained by replacing a native hidden activation with its pretrained SAE reconstruction. Our framework derives an upper bound on the base model's expected risk using four measurable quantities: proxy risk, SAE reconstruction gap, concept-pool mismatch, and sparse complexity. We interpret this certificate as an operational criterion for explanatory faithfulness. In particular, a non-vacuous bound indicates that the extracted sparse features retain meaningful predictive information, while small reconstruction and mismatch errors indicate that the proxy remains behaviorally close to the original model. Empirically, we show that the bound becomes non-vacuous on GPT-2 Small, Gemma-2B, and Llama-3-8B at practical sample sizes. A detailed layerwise analysis of Llama-3-8B reveals a strong depth dependence, with later layers becoming much easier to certify, associated with both stronger local fidelity and weaker downstream error amplification. Finally, through feature-shuffling ablations, we show that the decomposition distinguishes genuine semantic alignment from mere statistical sparsity, providing a useful diagnostic for when SAE-based explanations become less reliable.

Multimodal Large Language Models (MLLMs) have demonstrated strong performance on vision-language tasks, yet their internal visual representations remain difficult to interpret. Sparse Autoencoders (SAEs) provide a scalable way to decompose dense model activations into sparse, interpretable features. However, existing SAE architectures primarily recover flat feature dictionaries and are less suited for explicit multi-level concept organization. In this paper, we introduce cascaded sparse autoencoders (CSAEs) for learning hierarchical visual concepts in MLLMs. Rather than nesting or stacking SAE sparse activation codes, CSAEs train a second-level SAE directly on the decoder weights of the first-level SAE, treating learned low-level feature directions as inputs for higher-level abstraction. This design enables CSAEs to learn "concepts of concepts" while avoiding drawbacks from the shared-prefix coupling of nesting, Matryoshka-style hierarchies and the bottlenecks of naively stacked SAEs. Experiments across Qwen3-VL, Gemma-3, and LLaVA on multiple visual datasets show that CSAEs improve interpretability in terms of hierarchical concept coherence over state-of-the-art SAE baselines. Results on concept steering further demonstrate that the learned concept groups support effective group-level interventions in MLLM outputs.