Mechanistic interpretability methods summarize a transformer component by a single importance score, conflating two distinct roles: a component may matter because it transports task-relevant content, or because the forward computation degrades when its contribution is removed. We introduce \emph{Interchange-Group Sobol Decomposition} (IGSD), a paired-intervention framework that compares matched activation replacement with zero ablation on the same component, estimates two Sobol-style variance indices, and uses their signed difference to separate the two roles, with intervention validity monitored by a symmetric off-manifold diagnostic $\widehat{\mathrm{ST}}>1$. In factual recall, IGSD identifies an early-layer content channel in both GPT-2 small and Qwen2.5-1.5B that standard importance methods underestimate. A controlled subject and relation donor design shows that the early channel transports relation-frame content while late attention transports subject-retrieval content, refining at head granularity to the known $\mathrm{Attn}_{L9H8}$ head. Late-layer clamping confirms that the early signal is expressed through downstream transformations rather than residual pass-through. These results show that replacement and deletion are not interchangeable controls and their divergence provides a practical statistical diagnostic for content transport in transformer components.

Training vision language models (VLMs) aims to align visual representations from a vision encoder with the textual representations of a pretrained large language model (LLM). However, many VLMs exhibit reduced factual recall performance compared to their LLM backbones, raising the question of how effective multimodal fine-tuning is at extending existing mechanisms within the LLM to visual inputs. We argue that factual recall based on visual inputs requires VLMs to solve a two-hop problem: (1) forming entity representations from visual inputs, and (2) recalling associated factual knowledge based on these entity representations. By benchmarking 14 VLMs with various architectures (LLaVA, Native, CrossAttention), sizes (7B-124B parameters), and training setups on factual recall tasks against their original LLM backbone models, we find that 11 of 14 models exhibit factual recall degradation. We select three models exhibiting high-and two models with low performance degradation, and use attribution patching, activation patching, and probing to show that degraded VLMs struggle to use the existing factual recall circuit of their LLM backbone, because they resolve the first hop too late in the computation. In contrast, high-performing VLMs resolve entity representations early enough to reuse the existing factual recall mechanism. Finally, we demonstrate two methods to recover performance: patching entity representations from the LLM backbone into the VLM, and prompting with chain-of-thought reasoning. Our results highlight that the speed of early entity resolution critically determines how effective VLMs are in using preexisting LLM mechanisms. More broadly, our work illustrates how mechanistic analysis can explain and unveil systematic failures in multimodal alignment.

Training vision language models (VLMs) aims to align visual representations from a vision encoder with the textual representations of a pretrained large language model (LLM). However, many VLMs exhibit reduced factual recall performance compared to their LLM backbones, raising the question of how effective multimodal fine-tuning is at extending existing mechanisms within the LLM to visual inputs. We argue that factual recall based on visual inputs requires VLMs to solve a two-hop problem: (1) forming entity representations from visual inputs, and (2) recalling associated factual knowledge based on these entity representations. By benchmarking 14 VLMs with various architectures (LLaVA, Native, Cross-Attention), sizes (7B-124B parameters), and training setups on factual recall tasks against their original LLM backbone models, we find that 11 of 14 models exhibit factual recall degradation. We select three models with high and two models with low performance degradation, and use attribution patching, activation patching, and probing to show that degraded VLMs struggle to use the existing factual recall circuit of their LLM backbone, because they resolve the first hop too late in the computation. In contrast, high-performing VLMs resolve entity representations early enough to reuse the existing factual recall mechanism. Finally, we demonstrate two methods to recover performance: patching entity representations from the LLM backbone into the VLM, and prompting with chain-of-thought reasoning. Our results highlight that the speed of early entity resolution critically determines how effective VLMs are in using preexisting LLM mechanisms. More broadly, our work illustrates how mechanistic analysis can explain and unveil systematic failures in multimodal alignment.

The success of large language models (LLMs) can be attributed in part to their ability to efficiently store factual knowledge as key-value mappings within their MLP parameters. Recent work has proposed explicit weight constructions to build such fact-storing MLPs, providing an improved understanding of LLM fact storage mechanisms. In this paper, we introduce an MLP construction framework that improves over previous constructions in three areas: it 1) works for all but a measure-zero set of feasible input-output pairs, 2) achieves asymptotically optimal parameter efficiency matching information-theoretic bounds for some embeddings, and 3) maintains usability within Transformers for factual recall. Through our improvements, we 1) discover a metric on value embeddings that characterizes facts-per-parameter scaling for both constructed and gradient-descent-trained MLPs, 2) identify a simple encoder-decoder mechanism that empirically matches gradient-descent MLP facts-per-parameter asymptotics across all the inputs and outputs we test, and 3) uncover a fundamental tradeoff between an MLP's fact-storage capacity and its usability within Transformers. Finally, we demonstrate a proof-of-concept application of fact-storing MLPs: modular fact editing on one-layer Transformers by \textit{replacing entire MLPs at once}.

When large language models (LLMs) are aligned to a specific online community, do they exhibit generalizable behavioral patterns that mirror that community's attitudes and responses to new uncertainty, or are they simply recalling patterns from training data? We introduce a framework to test epistemic stance transfer: targeted deletion of event knowledge, validated with multiple probes, followed by evaluation of whether models still reproduce the community's organic response patterns under ignorance. Using Russian--Ukrainian military discourse and U.S. partisan Twitter data, we find that even after aggressive fact removal, aligned LLMs maintain stable, community-specific behavioral patterns for handling uncertainty. These results provide evidence that alignment encodes structured, generalizable behaviors beyond surface mimicry. Our framework offers a systematic way to detect behavioral biases that persist under ignorance, advancing efforts toward safer and more transparent LLM deployments.

Large Language Models (LLMs) are trained on diverse and often conflicting knowledge spanning multiple domains and time periods. Some of this knowledge is only valid within specific temporal contexts, such as answering the question, "Who is the President of the United States in 2022?" Ensuring LLMs generate time appropriate responses is crucial for maintaining relevance and accuracy. In this work we explore activation engineering as a method for temporally aligning LLMs to improve factual recall without any training or dataset creation. In this research we explore an activation engineering technique to ground three versions of LLaMA 2 to specific points in time and examine the effects of varying injection layers and prompting strategies. Our experiments demonstrate up to a 44% and 16% improvement in relative and explicit prompting respectively, achieving comparable performance to the fine-tuning method proposed by Zhao et al. (2024) . Notably, our approach achieves similar results to the fine-tuning baseline while being significantly more computationally efficient and requiring no pre-aligned datasets.

Unlearning in large language models (LLMs) is crucial for managing sensitive data and correcting misinformation, yet evaluating its effectiveness remains an open problem. We investigate whether persuasive prompting can recall factual knowledge from deliberately unlearned LLMs across models ranging from 2.7B to 13B parameters (OPT-2.7B, LLaMA-2-7B, LLaMA-3.1-8B, LLaMA-2-13B). Drawing from ACT-R and Hebbian theory (spreading activation theories), as well as communication principles, we introduce Stimulus-Knowledge Entanglement-Behavior Framework (SKeB), which models information entanglement via domain graphs and tests whether factual recall in unlearned models is correlated with persuasive framing. We develop entanglement metrics to quantify knowledge activation patterns and evaluate factuality, non-factuality, and hallucination in outputs. Our results show persuasive prompts substantially enhance factual knowledge recall (14.8% baseline vs. 24.5% with authority framing), with effectiveness inversely correlated to model size (128% recovery in 2.7B vs. 15% in 13B). SKeB provides a foundation for assessing unlearning completeness, robustness, and overall behavior in LLMs.

Language models are pretrained on sequences that blend statistical regularities (making text fluent) with factual associations between specific tokens (knowledge of facts). While recent work suggests that the variability of their interaction, such as paraphrases of factual associations, critically determines generalization ability, we lack a systematic analysis of these impacts. This paper introduces a flexible synthetic testbed that combines a statistical stream of generic tokens with an abstract factual stream of source-target token pairs, enabling fine-grained control over their interaction. The design enables the independent control of diversity nature by manipulating stream composition (contextual structure) and the diversity level by varying which statistical streams each fact appears in. Through controlled experiments, we find that while higher contextual diversity delays in-distribution (ID) factual accuracy, its impact on out-of-distribution (OOD) factual generalization depends critically on contextual structure. In some cases, OOD performance follows the same trend as ID, but in others, diversity becomes essential for non-trivial factual recall. Even when low diversity prohibits factual recall, optimal diversity levels depend on training duration. Beyond factual recall failures, we identify structures where statistical generalization fails independently, and others where both capabilities degrade. This shows how the interplay between contextual design and diversity level impacts different generalization aspects. Further, through a series of controlled interventions on the model components, we trace the OOD failures to distinct optimization bottlenecks, highlighting the importance of the embedding and unembedding layers. Our synthetic framework allows us to isolate effects that would be confounded in large-scale studies, offering a controlled testbed for future investigations.

Multilingual large language models (LLMs) are expected to recall factual knowledge consistently across languages. However, the factors that give rise to such crosslingual consistency -- and its frequent failure -- remain poorly understood. In this work, we hypothesize that these inconsistencies may arise from failures in entity alignment, the process of mapping subject and object entities into a shared conceptual space across languages. To test this, we assess alignment through entity-level (subject and object) translation tasks, and find that consistency is strongly correlated with alignment across all studied models, with misalignment of subjects or objects frequently resulting in inconsistencies. Building on this insight, we propose SubSub and SubInj, two effective methods that integrate English translations of subjects into prompts across languages, leading to substantial gains in both factual recall accuracy and consistency. Finally, our mechanistic analysis reveals that these interventions reinforce the entity representation alignment in the conceptual space through model's internal pivot-language processing, offering effective and practical strategies for improving multilingual factual prediction.

Large Language Models (LLMs) are capable of recalling multilingual factual knowledge present in their pretraining data. However, most studies evaluate only the final model, leaving the development of factual recall and crosslingual consistency throughout pretraining largely unexplored. In this work, we trace how factual recall and crosslingual consistency evolve during pretraining, focusing on OLMo-7B as a case study. We find that both accuracy and consistency improve over time for most languages. We show that this improvement is primarily driven by the fact frequency in the pretraining corpus: more frequent facts are more likely to be recalled correctly, regardless of language. Yet, some low-frequency facts in non-English languages can still be correctly recalled. Our analysis reveals that these instances largely benefit from crosslingual transfer of their English counterparts -- an effect that emerges predominantly in the early stages of pretraining. We pinpoint two distinct pathways through which multilingual factual knowledge acquisition occurs: (1) frequency-driven learning, which is dominant and language-agnostic, and (2) crosslingual transfer, which is limited in scale and typically constrained to relation types involving named entities. We release our code and data to facilitate further research at https://github.com/cisnlp/multilingual-fact-tracing.