LLM Interpretability

As spatial intelligence becomes an increasingly important capability for foundation models, it remains unclear whether large language models' (LLMs) performance on spatial reasoning benchmarks reflects structured internal spatial representations or reliance on linguistic heuristics. We address this question from a mechanistic perspective by examining how spatial information is internally represented and used. Drawing on computational theories of human spatial cognition, we decompose spatial reasoning into three primitives, relational composition, representational transformation, and stateful spatial updating, and design controlled task families for each. We evaluate multilingual LLMs in English, Chinese, and Arabic under single pass inference, and analyze internal representations using linear probing, sparse autoencoder based feature analysis, and causal interventions. We find that task relevant spatial information is encoded in intermediate layers and can causally influence behavior, but these representations are transient, fragmented across task families, and weakly integrated into final predictions. Cross linguistic analysis further reveals mechanistic degeneracy, where similar behavioral performance arises from distinct internal pathways. Overall, our results suggest that current LLMs exhibit limited and context dependent spatial representations rather than robust, general purpose spatial reasoning, highlighting the need for mechanistic evaluation beyond benchmark accuracy.

As large language models (LLMs) continue to improve and see further integration into software systems, so does the need to understand the conditions in which they will perform. We contribute a statistical framework for understanding the impact of specific prompt features on LLM performance. The approach extends previous explainable artificial intelligence (XAI) methods specifically to inspect LLMs by fitting regression models relating portions of the prompt to LLM evaluation. We apply our method to compare how two open-source models, Mistral-7B and GPT-OSS-20B, leverage the prompt to perform a simple arithmetic problem. Regression models of individual prompt portions explain 72% and 77% of variation in model performances, respectively. We find misinformation in the form of incorrect example query-answer pairs impedes both models from solving the arithmetic query, though positive examples do not find significant variability in the impact of positive and negative instructions - these prompts have contradictory effects on model performance. The framework serves as a tool for decision makers in critical scenarios to gain granular insight into how the prompt influences an LLM to solve a task.

Recent work identifies a sparse set of "hallucination neurons" (H-neurons), less than 0.1% of feed-forward network neurons, that reliably predict when large language models will hallucinate. These neurons are identified on general-knowledge question answering and shown to generalize to new evaluation instances. We ask a natural follow-up question: do H-neurons generalize across knowledge domains? Using a systematic cross-domain transfer protocol across 6 domains (general QA, legal, financial, science, moral reasoning, and code vulnerability) and 5 open-weight models (3B to 8B parameters), we find they do not. Classifiers trained on one domain's H-neurons achieve AUROC 0.783 within-domain but only 0.563 when transferred to a different domain (delta = 0.220, p < 0.001), a degradation consistent across all models tested. Our results suggest that hallucination is not a single mechanism with a universal neural signature, but rather involves domain-specific neuron populations that differ depending on the knowledge type being queried. This finding has direct implications for the deployment of neuron-level hallucination detectors, which must be calibrated per domain rather than trained once and applied universally.

Stereotypes in large language models (LLMs) can perpetuate harmful societal biases. Despite the widespread use of models, little is known about where these biases reside in the neural network. This study investigates the internal mechanisms of GPT 2 Small and Llama 3.2 to locate stereotype related activations. We explore two approaches: identifying individual contrastive neuron activations that encode stereotypes, and detecting attention heads that contribute heavily to biased outputs. Our experiments aim to map these "bias fingerprints" and provide initial insights for mitigating stereotypes.

Large language models (LLMs) tend to verbalize confidence scores that are largely detached from their actual accuracy, yet the geometric relationship governing this behavior remain poorly understood. In this work, we present a mechanistic interpretability analysis of verbalized confidence, using linear probes and contrastive activation addition (CAA) steering to show that calibration and verbalized confidence signals are encoded linearly but are orthogonal to one another -- a finding consistent across three open-weight models and four datasets. Interestingly, when models are prompted to simultaneously reason through a problem and verbalize a confidence score, the reasoning process disrupts the verbalized confidence direction, exacerbating miscalibration. We term this the "Reasoning Contamination Effect." Leveraging this insight, we introduce a two-stage adaptive steering pipeline that reads the model's internal accuracy estimate and steers verbalized output to match it, substantially improving calibration alignment across all evaluated models.

Understanding and quantifying uncertainty in large language model (LLM) outputs is critical for reliable deployment. However, traditional evaluation approaches provide limited insight into model confidence at individual token positions during generation. To address this issue, we introduce LogitScope, a lightweight framework for analyzing LLM uncertainty through token-level information metrics computed from probability distributions. By measuring metrics such as entropy and varentropy at each generation step, LogitScope reveals patterns in model confidence, identifies potential hallucinations, and exposes decision points where models exhibit high uncertainty, all without requiring labeled data or semantic interpretation. We demonstrate LogitScope's utility across diverse applications including uncertainty quantification, model behavior analysis, and production monitoring. The framework is model-agnostic, computationally efficient through lazy evaluation, and compatible with any HuggingFace model, enabling both researchers and practitioners to inspect LLM behavior during inference.

Linguistic representation learning in deep neural language models (LMs) has been studied for decades, but finding representations in LMs remains an unsolved problem. On the one hand, unconstrained alignments may trivialize the notion of representation (Sutter et al., 2025); on the other, even recently popularized linear approaches may not always be faithful to natural model behavior (Arora et al. 2024). Here we escape this dilemma by reconceptualizing representations not as patterns of activation but as conduits for learning. Our approach is simple: we perturb an LM by fine-tuning it on a single adversarial example and measure how this perturbation "infects" other examples. Perturbation makes no geometric assumptions, and unlike other methods, it does not find representations where it should not (e.g., in untrained LMs). But in trained LMs, perturbation reveals structured transfer at multiple linguistic grain sizes, suggesting that LMs both generalize along representational lines and acquire linguistic abstractions from experience alone.

When large language models make ethical judgments, do their internal representations distinguish between normative frameworks, or collapse ethics into a single acceptability dimension? We probe hidden representations across five ethical frameworks (deontology, utilitarianism, virtue, justice, commonsense) in six LLMs spanning 4B--72B parameters. Our analysis reveals differentiated ethical subspaces with asymmetric transfer patterns -- e.g., deontology probes partially generalize to virtue scenarios while commonsense probes fail catastrophically on justice. Disagreement between deontological and utilitarian probes correlates with higher behavioral entropy across architectures, though this relationship may partly reflect shared sensitivity to scenario difficulty. Post-hoc validation reveals that probes partially depend on surface features of benchmark templates, motivating cautious interpretation. We discuss both the structural insights these methods provide and their epistemological limitations.

LLMs are deployed globally, yet produce responses biased towards cultures with abundant training data. Existing cultural localization approaches such as prompting or post-training alignment are black-box, hard to control, and do not reveal whether failures reflect missing knowledge or poor elicitation. In this paper, we address these gaps using mechanistic interpretability to uncover and manipulate cultural representations in LLMs. Leveraging sparse autoencoders, we identify interpretable features that encode culturally salient information and aggregate them into Cultural Embeddings (CuE). We use CuE both to analyze implicit cultural biases under underspecified prompts and to construct white-box steering interventions. Across multiple models, we show that CuE-based steering increases cultural faithfulness and elicits significantly rarer, long-tail cultural concepts than prompting alone. Notably, CuE-based steering is complementary to black-box localization methods, offering gains when applied on top of prompt-augmented inputs. This also suggests that models do benefit from better elicitation strategies, and don't necessarily lack long-tail knowledge representation, though this varies across cultures. Our results provide both diagnostic insight into cultural representations in LLMs and a controllable method to steer towards desired cultures.

Mechanistic interpretability reveals that safety-critical behaviors (e.g., alignment, jailbreak, backdoor) in Large Language Models (LLMs) are grounded in specialized functional components. However, existing safety attribution methods struggle with generalization and reliability due to their reliance on heuristic, domain-specific metrics and search algorithms. To address this, we propose \ourmethod, a unified safety interpretability framework that identifies functionally complete safety circuits in LLMs via optimization. Unlike methods focusing on isolated heads or neurons, \ourmethod introduces differentiable binary masks to extract multi-granular circuits through gradient descent on safety datasets, while integrates Safety Circuit Tuning to utilize these sparse circuits for efficient safety fine-tuning. We validate \ourmethod in two key scenarios in LLM safety: (1) backdoor attacks, identifying a backdoor circuit with 0.42% sparsity, whose ablation eradicates the Attack Success Rate (ASR) from 100% $\to$ 0.4% while retaining over 99% general utility; (2) safety alignment, localizing an alignment circuit with 3.03% heads and 0.79% neurons, whose removal spikes ASR from 0.8% $\to$ 96.9%, whereas excluding this circuit during helpfulness fine-tuning maintains 96.5% safety retention.

Large language models (LLMs) generalize smoothly across continuous semantic spaces, yet strict logical reasoning demands the formation of discrete decision boundaries. Prevailing theories relying on linear isometric projections fail to resolve this fundamental tension. In this work, we argue that task context operates as a non-isometric dynamical operator that enforces a necessary "topological distortion." By applying Gram-Schmidt decomposition to residual-stream activations , we reveal a dual-modulation mechanism driving this process: a class-agnostic topological preservation that anchors global structure to prevent semantic collapse, and a specific algebraic divergence that directionally tears apart cross-class concepts to forge logical boundaries. We validate this geometric evolution across a gradient of tasks, from simple mapping to complex primality testing. Crucially, targeted specific vector ablation establishes a strict causal binding between this topology and model function: algebraically erasing the divergence component collapses parity classification accuracy from 100% to chance levels (38.57%). Furthermore, we uncover a three-phase layer-wise geometric dynamic and demonstrate that under social pressure prompts, models fail to generate sufficient divergence. This results in a "manifold entanglement" that geometrically explains sycophancy and hallucination. Ultimately, our findings revise the linear-isometric presumption, demonstrating that the emergence of discrete logic in LLMs is purchased at an irreducible cost of topological deformation.

Large language models (LLMs) solve complex problems by generating multi-step reasoning traces. Yet these traces are typically analyzed from only one of two perspectives: the sequence of tokens across different reasoning steps in the generated text, or the hidden-state vectors across model layers within one step. We introduce PRISM (Probabilistic Reasoning Inspection through Semantic and Implicit Modeling), a framework and diagnostic tool for jointly analyzing both levels, providing a unified view of how reasoning evolves across steps and layers. Across multiple reasoning models and benchmarks, PRISM uncovers systematic patterns in the reasoning process, showing that failed trajectories are more likely to become trapped in unproductive verification loops and further diverge into distinct modes such as overthinking and premature commitment, which behave differently once a candidate answer is reached. It further reveals how prompting reshapes reasoning behavior beyond aggregate accuracy by altering both semantic transitions and internal computational patterns. By modeling reasoning trajectories as structured processes, PRISM makes these behaviors observable and analyzable rather than relying solely on final-task accuracy. Taken together, these insights position PRISM as a practical tool for analyzing and diagnosing reasoning processes in LLMs.

While retrieval-augmented generation (RAG) enhances LLM performance, it does not eliminate hallucinations, making accurate detection essential. Uncertainty-based methods are attractive for this purpose because they can be integrated into real-world pipelines with little overhead. One of the most widely used uncertainty signals is predictive entropy. We show, however, that entropy can be unreliable in RAG settings and trace this limitation to two opposing internal effects. Induction heads, which copy patterns from earlier context, causally support correct responses and lower predictive entropy, but they also appear to co-activate entropy neurons that push it back up. As a result, correct, context-grounded responses can still receive high uncertainty scores. To address this, we propose INTRYGUE (Induction-Aware Entropy Gating for Uncertainty Estimation), a training-free, mechanistically grounded method that gates predictive entropy by an attention-based estimate of induction-head activity. Evaluated across four RAG-style benchmarks and six open-source LLMs (4B to 13B parameters), INTRYGUE performs competitively against a wide range of baselines, matching or exceeding the strongest of them in most settings. Our findings suggest that hallucination detection in RAG benefits from combining predictive uncertainty with interpretable internal signals of context utilization.

Mechanistic interpretability seeks to understand how Large Language Models (LLMs) represent and process information. Recent approaches based on dictionary learning and transcoders enable representing model computation in terms of sparse, interpretable features and their interactions, giving rise to feature attribution graphs. However, these graphs are often large and redundant, limiting their interpretability in practice. Cross-Layer Transcoders (CLTs) address this issue by sharing features across layers while preserving layer-specific decoding, yielding more compact representations, but remain difficult to train and analyze at scale. We introduce an open-source library for end-to-end training and interpretability of CLTs. Our framework integrates scalable distributed training with model sharding and compressed activation caching, a unified automated interpretability pipeline for feature analysis and explanation, attribution graph computation using Circuit-Tracer, and a flexible visualization interface. This provides a practical and unified solution for scaling CLT-based mechanistic interpretability. Our code is available at: https://github.com/LLM-Interp/CLT-Forge.

Understanding the fundamental mechanisms governing the production of meaning in the processing of natural language is critical for designing safe, thoughtful, engaging, and empowering human-agent interactions. If meaning is constituted rather than retrieved, then the search for context-independent features or circuits in the pursuit of mechanistic interpretability may be fundamentally limited. Experiments in cognitive science and social psychology have demonstrated that human semantic processing exhibits contextuality more consistent with quantum logical mechanisms than classical Boolean theories, and recent works have found similar results in large language models---in particular, clear violations of the Bell inequality in experiments of contextuality during interpretation of ambiguous expressions. In this work, we explore the CHSH $|S|$ parameter---the metric associated with the inequality---across the inference parameter space of models spanning four orders of magnitude in scale and cross-reference our findings with MMLU, hallucination rate, and nonsense detection benchmarks. We find that the interquartile range of the $|S|$ distribution is completely orthogonal to all external benchmarks, while overall violation rate shows weak anticorrelation with all three benchmarks. We investigate how $|S|$ varies with sampling parameters and word order, and discuss the information-theoretic constraints that genuine contextuality imposes on prompt injection defenses and its human analogue, whereby careful construction and maintenance of social contextuality can be carried out at scale, shaping the space of possible interpretations before any particular one is reached. We consider the implications for mechanistic interpretability and how genuine contextuality sets an information-theoretic bound on the decomposability of semantic processing.

Understanding the internal mechanisms of transformer-based large language models (LLMs) is crucial for their reliable deployment and effective operation. While recent efforts have yielded a plethora of attribution methods attempting to balance faithfulness and computational efficiency, dense component attribution remains prohibitively expensive. In this work, we introduce Layer-wise Target Propagation (LTP), a novel framework that faithfully traces information flow on the frozen transformer in one forward and one backward pass without requiring counterfactual examples. LTP analytically decomposes and linearizes the computational structure of the Transformers into distinct pathways along which it propagates a targeted unembedding vector to receive the effective representation at each residual position. This target-centric propagation achieves O(1) time complexity with respect to the number of model components, scaling to long input sequences and dense component attribution. Extensive experiments on standard interpretability benchmarks demonstrate that LTP achieves state-of-the-art faithfulness and unprecedented efficiency compared to existing baselines.

Large language models (LLMs) demonstrate strong performance, but they often lack transparency. We introduce GeoLAN, a training framework that treats token representations as geometric trajectories and applies stickiness conditions inspired by recent developments related to the Kakeya Conjecture. We have developed two differentiable regularizers, Katz-Tao Convex Wolff (KT-CW) and Katz-Tao Attention (KT-Attn), that promote isotropy and encourage diverse attention. Our experiments with Gemma-3 (1B, 4B, 12B) and Llama-3-8B show that GeoLAN frequently maintains task accuracy while improving geometric metrics and reducing certain fairness biases. These benefits are most significant in mid-sized models. Our findings reveal scale-dependent trade-offs between geometric precision and performance, suggesting that geometry-aware training is a promising approach to enhance mechanistic interpretability.

Knowledge-grounded dialogue systems aim to generate informative, contextually relevant responses by conditioning on external knowledge sources. However, most existing approaches focus exclusively on English, lack explicit citation mechanisms for verifying factual claims, and offer limited transparency into model decision-making. We present XKD-Dial, a progressive four-stage training pipeline for explainable, knowledge-grounded dialogue generation in a bilingual (English-Hindi) setting, comprising: (1) multilingual adaptation, (2) English dialogue SFT with citation grounding, (3) bilingual dialogue SFT, and (4) GRPO alignment with citation-aware rewards. We evaluate six models spanning encoder-decoder (250M-3B) and decoder-only (1B-7B) architectures at every pipeline stage. Our key contributions are: (i) three post-hoc explainability analyses - cross-attention alignment, Integrated Gradients attribution, and occlusion-based causal grounding - applied systematically across the training trajectory to reveal how citation behaviour is learned, not only whether it is learned; (ii) citation-grounded SFT reduces hallucination to 0.0% for encoder-decoder models from Stage 2 onward; (iii) the progressive pipeline prevents catastrophic forgetting while improving Hindi capabilities; (iv) smaller models match larger models on English after SFT; and (v) GRPO provides marginal improvement over well-designed SFT for structured citation tasks. We evaluate across six automatic metrics (BLEU, ROUGE, BERTScore, FactScore, Citation-F1, and hallucination rate).

Evaluating whether explanations faithfully reflect a model's reasoning remains an open problem. Existing benchmarks use single interventions without statistical testing, making it impossible to distinguish genuine faithfulness from chance-level performance. We show that faithfulness is not a fixed property but an operator-dependent quantity that changes with the intervention method used to measure it. We introduce ICE (Intervention-Consistent Explanation), a framework that evaluates explanations against random baselines of equal size under multiple operators. Evaluating 7 LLMs across 4 tasks with deletion and retrieval infill operators, we find that switching operators crosses the positive-evidence threshold in 18% of configurations (5 of 28 attention comparisons), with gaps reaching 44 percentage points. Randomized baselines detect anti-faithfulness (explanations worse than random) in nearly one-third of English deletion configurations, invisible without random comparisons. These patterns persist across 6 non-English languages and 2 attribution methods. The methodology generalizes to step-level chain-of-thought evaluation, where preliminary results on 3 frontier models suggest that high accuracy does not imply faithful reasoning.

Precise behavioral control of large language models (LLMs) is critical for complex applications. However, existing methods often incur high training costs, lack natural language controllability, or compromise semantic coherence. To bridge this gap, we propose WASD (unWeaving Actionable Sufficient Directives), a novel framework that explains model behavior by identifying sufficient neural conditions for token generation. Our method represents candidate conditions as neuron-activation predicates and iteratively searches for a minimal set that guarantees the current output under input perturbations. Experiments on SST-2 and CounterFact with the Gemma-2-2B model demonstrate that our approach produces explanations that are more stable, accurate, and concise than conventional attribution graphs. Moreover, through a case study on controlling cross-lingual output generation, we validated the practical effectiveness of WASD in controlling model behavior.

Understanding the internal functional organization of Large Language Models (LLMs) is crucial for improving their trustworthiness and performance. However, how LLMs organize different functions into modules remains highly unexplored. To bridge this gap, we formulate a functional module discovery problem and propose an Unsupervised LLM Cross-layer MOdule Discovery (ULCMOD) framework that simultaneously disentangles the large set of neurons in the entire LLM into modules while discovering the topics of input samples related to these modules. Our framework introduces a novel objective function and an efficient Iterative Decoupling (IterD) algorithm. Extensive experiments show that our method discovers high-quality, disentangled modules that capture more meaningful semantic information and achieve superior performance in various downstream tasks. Moreover, our qualitative analysis reveals that the discovered modules show semantic coherence, correspond to interpretable specializations, and a clear spatial and hierarchical organization within the LLM. Our work provides a novel tool for interpreting the functional modules of LLMs, filling a critical blank in LLM's interpretability research.

Understanding whether large language models (LLMs) capture structured meaning requires examining how they represent concept relationships. In this work, we study three models of increasing scale: Pythia-70M, GPT-2, and Llama 3.1 8B, focusing on four semantic relations: synonymy, antonymy, hypernymy, and hyponymy. We combine linear probing with mechanistic interpretability techniques, including sparse autoencoders (SAE) and activation patching, to identify where these relations are encoded and how specific features contribute to their representation. Our results reveal a directional asymmetry in hierarchical relations: hypernymy is encoded redundantly and resists suppression, while hyponymy relies on compact features that are more easily disrupted by ablation. More broadly, relation signals are diffuse but exhibit stable profiles: they peak in the mid-layers and are stronger in post-residual/MLP pathways than in attention. Difficulty is consistent across models (antonymy easiest, synonymy hardest). Probe-level causality is capacity-dependent: on Llama 3.1, SAE-guided patching reliably shifts these signals, whereas on smaller models the shifts are weak or unstable. Our results clarify where and how reliably semantic relations are represented inside LLMs, and provide a reproducible framework for relating sparse features to probe-level causal evidence.

Large language models (LLMs) can produce chains of thought (CoT) that do not accurately reflect the actual factors driving their answers. In multiple-choice settings with an injected hint favoring a particular option, models may shift their final answer toward the hinted option and produce a CoT that rationalizes the response without acknowledging the hint - an instance of motivated reasoning. We study this phenomenon across multiple LLM families and datasets demonstrating that motivated reasoning can be identified by probing internal activations even in cases when it cannot be easily determined from CoT. Using supervised probes trained on the model's residual stream, we show that (i) pre-generation probes, applied before any CoT tokens are generated, predict motivated reasoning as well as a LLM-based CoT monitor that accesses the full CoT trace, and (ii) post-generation probes, applied after CoT generation, outperform the same monitor. Together, these results show that motivated reasoning is detected more reliably from internal representations than from CoT monitoring. Moreover, pre-generation probing can flag motivated behavior early, potentially avoiding unnecessary generation.

Modern LLMs continue to exhibit significant variance in behavior across languages, such as being able to recall factual information in some languages but not others. While typically studied as a problem to be mitigated, in this work, we propose leveraging this cross-lingual inconsistency as a tool for interpretability in mixture-of-experts (MoE) LLMs. Our knowledge localization framework contrasts routing for sets of languages where the model correctly recalls information from languages where it fails. This allows us to isolate model components that play a functional role in answering about a piece of knowledge. Our method proceeds in two stages: (1) querying the model with difficult factual questions across a diverse set of languages to generate "success" and "failure" activation buckets and then (2) applying a statistical contrastive analysis to the MoE router logits to identify experts important for knowledge. To validate the necessity of this small number of experts for answering a knowledge question, we deactivate them and re-ask the question. We find that despite only deactivating about 20 out of 6000 experts, the model no longer answers correctly in over 40% of cases. Generally, this method provides a realistic and scalable knowledge localization approach to address increasingly complex LLMs.

A hallmark of human intelligence is Introspection-the ability to assess and reason about one's own cognitive processes. Introspection has emerged as a promising but contested capability in large language models (LLMs). However, current evaluations often fail to distinguish genuine meta-cognition from the mere application of general world knowledge or text-based self-simulation. In this work, we propose a principled taxonomy that formalizes introspection as the latent computation of specific operators over a model's policy and parameters. To isolate the components of generalized introspection, we present Introspect-Bench, a multifaceted evaluation suite designed for rigorous capability testing. Our results show that frontier models exhibit privileged access to their own policies, outperforming peer models in predicting their own behavior. Furthermore, we provide causal, mechanistic evidence explaining both how LLMs learn to introspect without explicit training, and how the mechanism of introspection emerges via attention diffusion.

Large Reasoning Models (LRMs) exhibit backtracking and self-verification mechanisms that enable them to revise intermediate steps and reach correct solutions, yielding strong performance on complex logical benchmarks. We hypothesize that such behaviors are beneficial only when the model has sufficiently strong ``critique'' ability to detect its own mistakes. This work systematically investigates how current LRMs recover from errors by inserting arithmetic mistakes in their intermediate reasoning steps. Notably, we discover a peculiar yet important phenomenon: despite the error propagating throughout the entire chain-of-thought (CoT) without any verbalized correction, the model still reaches the correct final answer after the thinking process finishes. This recovery implies the existence of an internal mechanism helping the model to detect errors and trigger self-correction, which we refer to as the hidden critique ability. Building on feature space analysis, we identify a highly interpretable critique vector representing this behavior. Extensive experiments across multiple model scales and families demonstrate that steering latent representations with this vector improves the model's error detection capability and enhances the performance of test-time scaling at no extra training cost. Our findings provide a valuable understanding of LRMs' critique behavior, suggesting a promising direction to control and improve their self-verification mechanism. Our code is available at: https://github.com/mail-research/lrm-critique-vectors.

Large language models (LLMs) increasingly participate in morally sensitive decision-making, yet how they organize ethical frameworks across reasoning steps remains underexplored. We introduce moral reasoning trajectories, sequences of ethical framework invocations across intermediate reasoning steps, and analyze their dynamics across six models and three benchmarks. We find that moral reasoning involves systematic multi-framework deliberation: 55.4--57.7% of consecutive steps involve framework switches, and only 16.4--17.8% of trajectories remain framework-consistent. Unstable trajectories remain 1.29 times more susceptible to persuasive attacks (p=0.015). At the representation level, linear probes localize framework-specific encoding to model-specific layers (layer 63/81 for Llama-3.3-70B; layer 17/81 for Qwen2.5-72B), achieving 16.8--22.2% lower KL divergence than the step-prior baseline. Activation steering applied during generation moves the framework-consistency--accuracy relationship, widening it for Qwen2.5-72B and erasing it for Llama-3.3-70B, and a probe-space layer sweep bounds the attainable drift reduction at 6.7--8.9%. We further propose a Moral Representation Consistency (MRC) metric whose underlying framework attributions are validated by human annotators (mean cosine similarity = 0.859), and we report what an automated coherence rater does and does not establish about it.

Explainable AI (XAI) interfaces seek to make large language models more transparent, yet explanation alone does not produce understanding. Explaining a system's behavior is not the same as being able to engage with it, to probe and interpret its operations through direct manipulation. This distinction matters for scientific disciplines in particular: scientists who increasingly rely on LLMs for reading, citing, and producing literature reviews have little means of directly engaging with how these models process and transform the texts they generate. In this ongoing design research project, I argue for a shift from explainability to interpretative engagement. This shift moves away from accounts of system behavior to instead enable users to manipulate a model's intermediate representations. Drawing on textual scholarship, computational poetics, and the history of reading and writing technologies, including practices such as marginalia, glosses, indices, and annotation systems, I propose interpretative interfaces as interactive environments in which non-expert users can intervene in the representational space of a language model. More specifically, such interfaces will allow users to select a token and follow its trajectory through the model's intermediate layers. This way, they can observe how its semantic position shifts as context is processed, and possibly annotate the transformations they find useful or meaningful. The same way readers can create their own maps within a book through annotations and bookmarks, interpretative interfaces will allow users to inscribe their reading of a model's internal representations. The goal of this project is to reframe AI interpretability as an interaction design project rather than a purely technical one, and to open a path toward AI-mediated reading that supports interpretative engagement and critical stewardship of scientific knowledge.

Despite the success of Large Language Models (LLMs) in table understanding, their internal mechanisms remain unclear. In this paper, we conduct an empirical study on 16 LLMs, covering general LLMs, specialist tabular LLMs, and Mixture-of-Experts (MoE) models, to explore how LLMs understand tabular data and perform downstream tasks. Our analysis focus on 4 dimensions including the attention dynamics, the effective layer depth, the expert activation, and the impacts of input designs. Key findings include: (1) LLMs follow a three-phase attention pattern -- early layers scan the table broadly, middle layers localize relevant cells, and late layers amplify their contributions; (2) tabular tasks require deeper layers than math reasoning to reach stable predictions; (3) MoE models activate table-specific experts in middle layers, with early and late layers sharing general-purpose experts; (4) Chain-of-Thought prompting increases table attention, further enhanced by table-tuning. We hope these findings and insights can facilitate interpretability and future research on table-related tasks.

This paper proposes and argues for a counterintuitive thesis: the truly valuable capabilities of large language models (LLMs) reside precisely in the part that cannot be fully captured by human-readable discrete rules. The core argument is a proof by contradiction via expert system equivalence: if the full capabilities of an LLM could be described by a complete set of human-readable rules, then that rule set would be functionally equivalent to an expert system; but expert systems have been historically and empirically demonstrated to be strictly weaker than LLMs; therefore, a contradiction arises -- the capabilities of LLMs that exceed those of expert systems are exactly the capabilities that cannot be rule-encoded. This thesis is further supported by the Chinese philosophical concept of Wu (sudden insight through practice), the historical failure of expert systems, and a structural mismatch between human cognitive tools and complex systems. The paper discusses implications for interpretability research, AI safety, and scientific epistemology.