Hyunsoo Yun, Eun Hak Lee, Jiaru Zhang +2cs.LG cs.AI
Conventional discrete choice and machine learning models are estimated primarily from observational data and typically treat explanatory covariates as parallel inputs, providing no internal mechanism for determining how related attributes should adjust when one is deliberately changed. This paper proposes Neural-Bayesian Structure Learning (Neural-BSL), a framework coupling differentiable structure learning with random-utility-based discrete choice estimation in a single differentiable procedure. To prevent mutually exclusive choice outcome from distorting the recovered attribute structure, the observed choice is maintained outside the graph as an alternative-specific utility comparison, while the attribute structure and random-utility parameters are learned jointly. The learned structure enters the choice model through structure-weighted attribute interactions and provides the structural basis for propagating interventions through downstream attributes. An intervention is evaluated by updating the intervened attribute, propagating its model-implied downstream changes in topological order, and then recomputing utilities and choice probabilities. This yields both predicted mode-share responses and the associated changes in downstream traveler or trip attributes. We evaluate Neural-BSL using stated-preference data from Seoul and the revealed-preference data from London. Neural-BSL achieves predictive performance comparable to conventional benchmarks while recovering behaviorally coherent dependency structures. Across policy scenarios, propagating interventions through the learned structure changes the predicted redistribution across modes while exposing the downstream traveler and trip adjustments underlying those responses.
When a context asserts two values for one fact, a model commits to a cue -- recency, repetition, position -- but natural data rarely makes these disagree, so behavior cannot reveal which. We train 26M-parameter transformers on a synthetic language where recency and rarity are exactly coextensive, and separate them with a minimal causal edit that inverts one cue while holding the truth, token count and answer position fixed. All 75 runs reach accuracy >= 0.999, including where the trivial heuristic fails, so no held-in evaluation distinguishes them. Under intervention the per-cell readout does not replicate: 13 of 25 cells differ by more than 0.3 in sign fraction across three seeds, the largest by 0.879 against a standard error of 0.025. The construction predicts this -- coextensive rules leave the objective indifferent between them -- and the variance is ordered by how much of the optimization each comparison releases. What replicates is timing: escape from a positional shortcut with a closed-form ceiling, monotone in redundancy. Probed before that escape, attribution reverses sign in 32 of 75 runs at unchanged accuracy, and gating on circuit formation is necessary but not sufficient. The corpus fixes when a mechanism appears, not which one -- a criterion for when mechanistic attribution to data is available at all, and our construction makes the unavailable case exact.
Vision-language models are known to encode spatial information in their hidden states, yet often fail to use it when answering. However, it remains unclear when and where this encoded information reaches the answer. We address this with direction patching, a class-conditioned causal intervention applied across layers, token positions, and prompt formats. Using spatial-ID directions constructed following prior encoding evidence, we find that causal influence on answer logits emerges only at mid-to-deep depths. Text chain-of-thought suppresses immediate object-word argmax-level transport in most models, while visually grounded prompts keep it open. Positive target-logit gain can remain below the argmax threshold, and transport can re-emerge at the final prefix token or at the answer step in deeper layers. Across the ten VLMs we study, these local effects form descriptive transport patterns. Complementary experiments characterize how these patterns shift across datasets, attributes, and encoding amplitudes. Together, these results reframe the encoding-grounding gap as a problem of conditional transport in VLMs.
Foundation models for astronomy are trained on survey pixels together with the catalogue products derived from those pixels. Those catalogues are incomplete at a measurable rate, and a model trained on both inherits that incompleteness as a systematic. We audit AION-1, a 39-modality transformer trained on more than 200 million objects, using causal interventions on its inputs. Holding the image tokens byte-identical and editing only the survey segmentation map changes every quantity the model reports -- flux, size, ellipticity, redshift -- by 110-4400 times a matched placebo. The mechanism is detection gating, presence at the field centre (r = 0.47), not the light the mask encloses (r = 0.30); across 322 real blends the model ignores how the pipeline partitioned the light (R = -0.006). Nor is the preference specific to that channel: contradicted catalogue photometry leaves the model nine times worse than supplying no metadata at all. The Legacy Survey pipeline leaves 3.68% of targets with no segment covering their position. Propagating that rate, with a miss represented by the fields the pipeline actually returns, shifts tomographic mean redshifts by a median 0.71 times the LSST DESC requirement over 40 assignments and exceeds it in 12; observed positional errors take the worst bin to 8.3 times. Drawing the misses by their measured magnitude dependence rather than uniformly does not change it. Spectroscopy removes the effect, withholding the detection channel removes it at no measurable cost, and the effect grows with model scale. Two further limits lie in the tokeniser: its image codec resolves 28 effective states on source patches against 934 for the spectrum codec, and the redshift readout is quantisation-limited. Sparse dictionaries are unreliable causal handles: across 15, recovery spans 26-75% and moves up to 18 points on the seed alone.
Behavioral accuracy, linear decodability, and successful activation interventions do not by themselves show that a model carries an operation-level structure from one symbolic domain to another. We ask a narrower question in finite isomorphic state spaces: if the hidden-state difference between two operations is estimated separately for each source input, does adding that difference to a mapped recipient input move the model toward the corresponding recipient answer? The design compares this input-specific intervention with wrong-operation, norm-matched random, and no-op controls, and separates candidate construction from an independently isolated confirmation split. On a frozen Qwen2.5-7B-Instruct model at layers 20--21, one route--domain--operation candidate from a family pre-specified and frozen before confirmation access, transparent | integer_mod16--letters16 | successor->predecessor, passed both PyVene splits; its confirmation intersection--union p-value was 0.000198 and its 36-family Holm-adjusted p-value was 0.006943. A subsequent NNsight 0.7.0 experiment, pre-specified and frozen before its confirmation access, tested only this selected prompt route, without candidate or layer reselection. It reproduced all 12 confirmation effect estimates, confidence intervals, and exact sign-flip p-values numerically; its 36-family Holm-adjusted p-value was 0.007141. The result is therefore limited to one prompt route and one candidate, replicated across two intervention implementations on one model revision and one layer interval. It does not establish cross-model generalization, full-family backend independence, domain-general transfer, or algebraic invariance.
Large language models (LLMs) often shift their outputs in response to implicit demographic cues even when users never state a demographic identity. Previous work has documented this behavior, but the connection between these behavioral changes and the model's internal activations remains unclear. Using matched cued and neutral conversations across five LLMs, we establish that a localized internal activation signal tracks changes in recommendations, with correlations up to r=0.87. When multiple cues appear together, their internal signals largely combine, but the changes in output do not simply add up. We further show that removing the internal signal associated with one cue can suppress its influence, often more effectively than asking the model to ignore demographics via prompting, while largely preserving general benchmark performance. However, the ability to selectively remove one dimension's influence while leaving co-present dimensions intact remains highly model- and attribute-specific. These results connect implicit personalization behavior to an internal signal that can be analyzed and causally controlled.
Vision-Language-Action (VLA) models have driven significant progress in robotic manipulation, yet they fundamentally struggle with the vision-override phenomenon. Driven by the severe modality imbalance between dense visual streams and sparse linguistic instructions, VLAs frequently fall prey to causal confusion. Instead of treating language as the primary causal driver, the policy entirely bypasses the original instruction by overfitting to spurious visual confounders, such as prominent objects or familiar layouts. To systematically alleviate this bias, we formalize the process of action generation as a Dual-path Deconfounding Graph (DDG) and propose CofactVLA, a novel causal intervention framework. By dynamically constructing a language-masked counterfactual branch within a single forward pass, CofactVLA isolates and neutralizes visual confounders through two synergistic mechanisms. First, Action-Level Orthogonal Projection Guidance (OPG) geometrically projects the factual velocity field away from the counterfactual visual bias during continuous flow matching, extracting the pure semantic intent. Second, Feature-Level Counterfactual Covariance Reduction (CCR) mathematically deconfounds latent representations by penalizing the positive eigenspace of the covariance difference, explicitly suppressing dominant visual shortcuts while preserving the causal language intent. Extensive experiments demonstrate that CofactVLA establishes a new state-of-the-art across diverse simulation benchmarks. Beyond simulation, real-world robot experiments demonstrate the causal efficacy of our method in bridging the generalization gap, yielding a 52.3\% absolute success rate gain under out-of-distribution scenarios.
Markus J. Buehlercs.AI cond-mat.mes-hall cond-mat.mtrl-sci cs.CL
Large language models can answer scientific questions, yet a correct output does not reveal whether the model represents or uses the governing physics. Here we show that materials science mechanism information in the open-weight google/gemma-4-E4B-it model has three experimentally separable forms: concepts are readable in individual hidden states, constitutive orientation is carried by controlled transformations between states, and selected internal representations causally control engineering answers. We combine matched direct and Jacobian vocabulary readouts, option-free state geometry, a 60-law counterfactual benchmark and causal interventions. In 50 held-out materials descriptions, three independently fitted Jacobian lenses reproduced concept ranks, and target-free word sets from both readouts enabled blinded identification of 9 of 10 mechanism families. A separate 72-prompt benchmark produced mechanism-specific hidden-state neighborhoods, but an exact graph audit showed that this apparent physical organization was equally explained by numerical comparison. We therefore compared otherwise identical prompts in which only the direction of the physical input was reversed, asking whether the resulting hidden-state movement followed the supplied constitutive law. These state transformations ordered direct, physically neutral and inverse laws across 60 frozen relations and correctly oriented 39 of 40 directional laws, whereas lexical controls were near chance. Bidirectional interventions shifted answer probabilities toward or away from the physically appropriate outcome across all 12 matched cases, while counterfactual state patches transferred opposing decision signals across mechanisms and answer formats. Physical relationships were therefore more visible in controlled state changes than in absolute states alone.
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.
Brian Song, Michael A. Lepori, Ellie Pavlickcs.AI cs.CV
Goal-directed visual processing is a hallmark of human visual intelligence, resulting in representations that support downstream tasks such as categorization or search. Though vision-language models (VLMs) are often faced with these same tasks, their ability to recode visual representations when presented with goal-directed language remains poorly characterized. Indeed, prior work largely treats visual representations in VLMs as static repositories of visual information that are manipulated by language representations. In the present work, we provide evidence for two concrete instances of language-induced recoding of visual representations. First, we identify an abstract reference representation that denotes which objects are goal-relevant under a natural language prompt. We extract contrastive steering vectors corresponding to this reference representation and demonstrate that they are causally implicated in model predictions. These reference representations are abstract in that they generalize to different objects, different task contexts, and even from synthetic to naturalistic images. Second, we demonstrate language-induced attribute modulation: later layers selectively amplify goal-relevant attributes in visual representations of objects. We demonstrate this phenomenon across a range of different prompts. Finally, we provide a causal intervention that demonstrates that attribute modulation mediates a VLM's response distribution. Together, our results support a more dynamic account of cross-modality processing in VLMs -- rather than vision tokens serving as static repositories of information, they are modulated to support queries articulated in language.
Compositional visual question answering requires Vision-Language Models (VLMs) to execute multiple reasoning operations like object selection, spatial relation resolution, and attribute verification. Despite strong aggregate performance, the mechanistic basis of VLM failures on this task remains underexplored. To address this gap, we analyze vision-operation misalignment in VLMs by examining how failures relate to specific reasoning operations and the internal computational pathways through which they arise and propagate. We introduce an Operation-centric mechanistic framework that decomposes VLM failures by both the reasoning operation where they originate and the internal computational pathway through which they propagate. Our analysis reveals four dominant failure modes: grounding failure, reasoning failure, attribute extraction failure, and language-prior dominance, each characterized by a distinct relationship between visual grounding strength and answer correctness. Through three complementary causal interventions applied across all transformer layers, we find that object-selection failures are associated primarily with feedforward computation, multi-step relational failures with late-layer direct attention, and attribute-extraction failures with answer-position feedforward computation. Validation on VSR further shows that single-step spatial failures are concentrated at object-position encoding, distinguishing them from multi-step relational composition. These findings reveal distinct computational bottlenecks across operation types and provide a principled basis for targeted diagnosis of VLM failures in multimedia reasoning.
Large language models (LLMs) exhibit remarkable capabilities but remain highly vulnerable to adversarial prompts and jailbreak attacks. Existing approaches primarily analyze these failures through input-output behaviors or attribution methods, offering limited insight into how adversarial perturbations alter the model's internal reasoning. Consequently, the mechanisms underlying unsafe or incorrect behaviors remain poorly understood. We introduce a mechanistic framework for diagnosing LLM vulnerabilities using paired internal computation graphs, which represent prompt-specific inference as structured causal interactions among latent features. By constructing and aligning computation graphs for clean and attacked prompts, we reveal that adversarial attacks induce systematic transformations of internal reasoning, including suppression of safety-relevant components, emergence of attack-specific features, and rerouting of computation paths. Building on this representation, we propose a unified framework that (i) decomposes computation into invariant, suppressed, and emergent structures, (ii) identifies recurring vulnerability motifs associated with failure modes, and (iii) performs causal interventions on nodes, paths, and subgraphs to directly evaluate their contributions to attack success. This enables a transition from descriptive attribution to causal diagnosis of model failures. Experiments across multiple open-source LLMs and diverse adversarial and jailbreak benchmarks demonstrate that structural deviations in internal computation graphs strongly correlate with unsafe behaviors. Furthermore, targeted interventions on identified vulnerability motifs improve model robustness, establishing internal computation graphs as a principled foundation for understanding, diagnosing, and mitigating LLM vulnerabilities.
A frozen vision-language-action model (VLA) receives recent observations at every decision step, yet prior work has focused on adding memory rather than asking how existing history is represented and used. We study this temporal axis using layer-resolved linear probing and causal interchange interventions across three VLAs from two architecture families. We find a three-part dissociation. First, past-frame content remains linearly decodable throughout the network. Second, information unique to history beyond the current frame is nearly absent, indicating that stored history is largely a redundant copy of the present. Third, history is causally deployed only when the current frame is heavily degraded, while the action readout progressively loses dependence on history through the network. Although all models encode history similarly, their deployment strategies differ: under the same occlusion, one architecture increasingly relies on history as a fallback, whereas the other relies on it less. We further introduce a training-free temporal deployment audit that distinguishes these regimes. In the fallback regime, re-injecting history neither repairs occlusion nor disambiguates actions, confirming the redundancy of the stored representation. In the other regime, the same intervention reliably steers the predicted action toward the donor history. These results show that steerability depends on how history is deployed rather than whether it is encoded. VLAs do not forget the past; they largely fail to represent it as information distinct from the present. Our findings suggest that future memory augmentation should inject information unique to the past rather than simply more history.
Large Vision-Language Models (LVLMs) exhibit sophisticated reasoning but remain susceptible to object hallucination. Deviating from the prevailing attention intensity assumption, we reveal a deeper dynamic structural misalignment: hallucination is triggered at decision-critical steps where specific attention heads, acting as risky mediators, decouple from visual evidence to lock onto language priors. This establishes a pathological shortcut that bypasses visual grounding. To dismantle this, we propose Fox (Faithfulness and Observational-flow via eXpression-rectification), a training-free inference-time framework. Fox diagnoses structural misalignment using a visual attention entropy probe to localize risky mediators unsupervisedly. We then execute a targeted causal intervention via numerical logit saturation to physically sever the shortcut path. Finally, a conflict-gated cooperative decoding strategy reconciles interventional faithfulness with observational fluency. Extensive experiments demonstrate that Fox achieves SOTA performance, outperforming SID by 29.1% while preserving linguistic richness. Code is available at https://github.com/Cc2021start/Fox.
Arjun Pillai, Christian Hoang, Anjelo Jann Larozacs.CL cs.AI
Why do multilingual language models sometimes generate in the wrong language, and why is this so hard to fix? We introduce Language Identity Head Ablation (LIHA), a causal intervention that zeros each attention head individually and measures the resulting language switch rate across a parallel dataset of 2,700 prompt-language pairs spanning seven languages. Applied to GPT-2, LIHA identifies a small set of first-token broadcaster heads - led by L6H1 (switch rate 0.32, 3.23 $σ$ above the population mean) - that attend persistently to the first prompt token, propagating its language signal throughout generation. Compensatory redistribution when heads are ablated is statistically significant (p < $10^{-5}$) and follows a directional, hierarchical pattern: compensation always recruits heads in layers above the ablated head, suggesting a feedforward cascade rather than global diffusion. To probe how training regime shapes these circuits, we apply LIHA to a controlled pair - Qwen2.5-1.5B-Base and Qwen2.5-1.5B-Instruct - identical in architecture and size, differing only in training. The base model is nearly flat (max SR=0.016, 200/336 heads at SR=0.0); the instruct model concentrates causal influence sharply at layer 0, led by L0H5 (SR=0.224, 8.93 $σ$ above mean), with all other layers near zero. This controlled comparison provides direct causal evidence that instruction tuning reorganizes language identity circuits toward early-layer localization. Extended experiments with Chinese and Russian confirm that first-token broadcasting is script-specific in GPT-2, with non-Latin languages handled at layer 0 - the same locus as the instruction-tuned model. Code and data will be released upon publication.
Human speech is richly expressive, with prosody carrying linguistic and emotional information beyond the lexical content. A capable large audio-language model (audio-LLM) should therefore support expressive speech understanding, not only transcribing what was said but also interpreting how it was said. Yet behavioral evaluations alone cannot reveal why a model fails on prosodic input. An error may reflect loss of acoustic information, incorrect internal interpretation, or failure to use a representation that is already available inside the model. We introduce a stage-specific probe ladder for localizing these failure modes in audio-LLMs. Across four understanding-only audio-LLMs, prosodic information is usually preserved in the audio path and decodable in late LLM states. Yet it is only partially expressed in the model's final response. We test the causal status of this latent representation with targeted hidden-state interventions. Every intervention shifts the answer distribution in the predicted direction, and in most model--task cells a single edit at the relevant layer is sufficient to drive the model toward the suppressed prosodic decision, though this recovery is directional rather than a selective restoration of the correct class. Feature-level analysis further suggests that this recoverable signal can be expressed through a small subspace. Some of the highest-attribution features in this analysis align with acoustic cues known to carry prosodic information. Within the matched-content contrasts we test, these results locate the recurring bottleneck not in perceiving prosody but in using it. Models that hear and correctly represent a prosodic cue can still fail to express it in their answers.
When an LLM agent fails -- issues a refund it should not have, calls the wrong tool, leaks data -- existing tooling answers what happened (observability) or whether it passed (evaluation), but not which step caused the failure. The obvious heuristics are wrong: the step that executes the harmful action is usually not the step that decided on it, and LLM-judge attribution is correlational and unreliable (state-of-the-art step-level accuracy on the Who&When benchmark is about 14%). We present Causal Agent Replay (CAR), which answers the question by intervention: it models an agent run as a structural causal model, applies a do-operation to a step, and re-executes the trajectory forward under the same stochastic policy, measuring the shift in the outcome distribution. We define an intervention algebra over agent steps, a single-step contrastive estimator whose point-of-commitment rule resolves a confound specific to stochastic run-forward, and a budget-bounded Monte-Carlo Shapley estimator that splits credit across interacting steps. Every effect is reported with confidence intervals. We validate against synthetic structural causal models with planted ground truth: the contrastive estimator recovers the pivotal step, and Shapley recovers a two-step interaction (0.44, 0.45, ~0; efficiency sum 0.909 versus the analytic 0.91). CAR is open source and runs on hosted or free local models.
Recent advances in Large Language Models (LLMs) have opened new possibilities for time series forecasting by enabling alignment between temporal patterns and pretrained word embeddings. However, most LLM-based methods overlook the heterogeneous nature of time series, where dynamic fluctuations and invariant semantics are entangled. This entanglement introduces spurious correlations during the alignment, as dynamic components act as confounders by simultaneously influencing invariant components and the resulting aligned embeddings. To address this issue, a variable-level alignment framework CVAformer is proposed. CVAformer explicitly disentangles each variable into invariant and dynamic components just before alignment, and applies causal intervention to mitigate the confounding effect of the dynamics. To better support variable-level alignment, CVAformer replaces the standard causal attention in LLMs with a non-causal attention mechanism that captures interactions among variables at each time step. Extensive experiments across long-term, short-term, few-shot, and zero-shot forecasting settings indicate that CVAformer matches or exceeds state-of-the-art performance on most datasets, and in some cases achieves notably better accuracy. Experimental results validate the effectiveness of variable-level alignment and dynamic disentanglement in CVAformer, offering a new perspective for LLM-based time series tasks.
Despite the impressive multilingual capabilities of Large Language Models, the latent dynamics dictating language selection remain poorly understood. In this work, we ask whether language identity is merely linearly decodable from hidden states, or if it can be causally controlled by a compact activation direction. We conduct an exhaustive causal intervention analysis across multiple model families, including Qwen 3.5-2B and Llama-3.2-1B-Instruct, isolating PCA-derived "language axes" to perform steering and ablation experiments across 1.26 million generations on the FLORES-200 dataset. Steering along these geometric directions reliably forces language switching in both cross-script (English to Chinese) and same-script (English to Spanish) settings, whereas equal-magnitude random perturbations yield virtually no effect. Our layerwise analysis reveals that language commitment is highly localized and explicitly language-pair-dependent. While English to Chinese switching resists early intervention and steers easily in the later layers, the English-Spanish transition shifts earlier, displaying a distinct, bimodal sensitivity. Furthermore, targeted ablation uncovers a fundamental reversion to English: once the language signal is removed, the model falls back to English regardless of the input prompt. Ultimately, these findings demonstrate that language decision boundaries function during inference as causally active features that are direction-dependent and layer-specific.