LLM Interpretability

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.

Representation engineering reads and steers capability directions in large language models, yet methods are typically evaluated on paper-specific synthetic data. The resulting measurements are difficult to compare or reproduce and may reflect surface patterns rather than capabilities. We present RepBench, a benchmark-grounded data layer for capability-aligned representation probing. Crawling 13,427 benchmark papers yields a taxonomy of 182 capability clusters in 13 families; harvesting 353 public benchmark datasets yields 46,149 audited probe texts covering 94 capabilities, each supported by at least two independent benchmarks. This multi-benchmark design reduces dependence on any single source: raw per-text vectors exhibit no natural cluster granularity, whereas benchmark-pooled capability vectors show an interior clustering optimum at a small number of clusters on all 12 evaluated models, with low agreement to the human taxonomy. Under cross-benchmark transfer evaluation across twelve models completed by all four readouts, difference-in-means attains the highest model-level mean on ten models, while logistic regression wins the most capability-model cells. This disagreement shows that the readout method and aggregation criterion are meaningful evaluation dimensions. The pipeline, corpus, and evaluation code are released as a reusable closed-loop workflow.

Large Language Models (LLMs) explore problems through chain-of-thought, but this exploration is buried in unstructured prose. On high-stakes tasks, users cannot tell which steps are well-supported, which alternatives were seriously considered, or how the final conclusion compares to those the model discarded. We propose a framework that ensembles the reasoning structure, not just the answers, of multiple LLMs by weighted merging of Directed Acyclic Graphs (DAGs) extracted from reasoning chains. We weight each step by how many traces independently attest to it, to return "Consensus Reasoning". Across six benchmarks spanning statutory interpretation, graduate-level science, narrative multi-hop reasoning, and first-order logic, our ensemble outperforms a matched-budget majority-vote baseline, with a maximum accuracy gain of 3.1% on MuSR-MM (narrative multi-hop reasoning). On a single model, the framework matches or exceeds self-consistency at the same trace budget while additionally exposing an inspectable consensus reasoning graph. Ensemble weights correlate with LLM-judge rankings of reasoning quality at Spearman $ρ= 0.30$-$0.51$, and consensus subgraphs are preferred over alternatives leading to the majority-vote answer in 54.4-65.4% of head-to-head comparisons across five of six datasets. We observe that our framework can also be used to analyze diverse reasoning perspectives for a problem.

Identical language-model answers can arise from hidden states that support different future computations, so current-answer probes do not establish a reusable internal interface. We introduce forked futures: future operations are sampled only after a prefix state has formed, and states are compared through the response distributions induced by those operations. This yields an empirical causal quotient over hidden states without requiring researcher-specified latent labels. Shared, Local, Mixture, and Distributed interfaces then compete under prequential causal description length subject to future-signature fidelity and matched capacity constraints. In the two detailed model evaluations, Shared has the lowest held-out description length, with gains of 0.216 nats on Qwen2.5-1.5B and 0.294 nats on Llama-3-8B, while maintaining tightly clustered mean future-signature distortion; a five-backbone sweep preserves the positive direction of Sharedness Gain. The figure-aligned transplantation analysis gives Shared the strongest joint target-correctness, locality, copy-preservation, and composite profile, and API-aligned paths mediate 0.749 of the target effect versus 0.150 for matched null paths. In the blind four-class model-organism test, 14/16 architectures are recovered, with one observed non-Shared to Shared error among 12 non-Shared organisms. These results support an economical reusable causal interface within the tested operation banks, while keeping the claim explicitly conditional on the candidate architectures, interventions, and held-out futures.

Human personality theories characterize traits not as isolated attributes captured by a single score, but as stable individual tendencies expressed through the interplay among persons, situations, and behaviors. Existing studies of personality-related behavior in LLMs have primarily focused on outputs elicited under personality conditioning, characterizing observable trait-related expressions while lacking mechanistic evidence for the existence of internal personality-related representations, their cross-situational expression, and how these representations shape specific behaviors. Building on Funder's personality triad framework, we adapt its three components for LLM analysis: Person as personality-related internal representations, Situation as contexts that afford trait-relevant responses, and Behavior as response patterns on broader social tasks. We introduce a framework for discovering, controlling, and validating trait-like representations in LLMs. First, using contrastive behavior pairs grounded in shared situations, we identify sparse internal features associated with opposing poles of personality traits through SAE decomposition. We validate their trait relevance through effects on behavior to situation, token-level activation patterns, and robustness to paraphrasing. Second, feature-level interventions induce bidirectional trait-related shifts across a separate, diverse set of situations while preserving response validity, demonstrating consistent expression across contexts. Third, applying the same interventions to social intelligence tasks reveals behavioral changes with benefit-tradeoff patterns consistent with findings from human personality research, providing behavioral-level validation beyond personality scores. Our findings provide evidence that LLMs contain controllable trait-like representations linking internal states, situational expression, and behavioral outcomes.

Large language models (LLMs) encode rich concept-like information, but represent it implicitly through distributed statistical associations rather than as explicit, structured, compositional concepts. Consequently, concept-level structure is typically \emph{found} rather than \emph{designed}: it is recovered after training through probing or dictionary learning, with no architectural guarantee of stability, compositionality, controllability, or alignment with human conceptual organization. We organize concept-aware interventions along two dimensions: whether concept structure is internally induced or externally grounded, and the stage of the pipeline where it is introduced. This taxonomy reveals three broad patterns: inference-time approaches remain comparatively underexplored, related ideas have developed largely in isolation across pipeline stages, and externally grounded methods span the entire pipeline despite often being described under different terminology. Together, these observations motivate moving beyond recovering concept-like structure from trained models toward designing LLMs with explicit conceptual representations.

Large reasoning models trained via reinforcement learning (RL) have been increasingly shown to outperform their supervised fine-tuned (SFT) counterparts on mathematical reasoning tasks; Yet the mechanistic basis for this advantage remains unclear. We therefore ask, what internal representational differences enable RL models' superior performance? Our work presents two converging lines of evidence: First, linear probes trained on layer-wise hidden states reveal that RL models tend to achieve higher accuracy in predicting answer correctness compared to SFT models, indicating more linearly separable and structured representations. Second, mean ablation studies show that RL models develop a hierarchical architecture where deeper layers become progressively more critical, whereas SFT models distribute importance uniformly across layers. Together, these findings demonstrate that RL training fundamentally restructures how models represent and process reasoning problems. Finally, we analyze token-count variability under repeated sampling across problems to assess adaptive compute allocation. While we observe higher variability in some RL-tuned models than in their SFT counterparts, we see strong consistency in others, suggesting that token allocation may depend more on the overall training pipeline than on RL versus SFT alone. We believe this token-allocation variability reveals the spread of plausible on-policy reasoning, highlighting which models exhibit stable policies versus those that are under-determined, potentially non-identifiable solution behaviour.

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.

Large Language Model (LLM)-based Automated Program Repair systems are advancing rapidly, yet their performance remains inconsistent. Even when provided with the same contextual information, an LLM may generate a correct patch for one bug but fail on another closely related bug. Why this happens remains poorly understood, and it is unclear how LLMs prioritize the diverse information in bug reports and whether model attention affects repair success. In this paper, we present the first empirical study of attention patterns in LLM-based program repair, providing interpretable insights into how models process bug reports and where their attention is concentrated during repair. We analyze 319 real-world Python and Java bugs from SWE-bench Verified and Multi-SWE-bench to study (RQ1) how model attention is distributed across bug report sections, (RQ2) how attention patterns within each section differ between successful and unsuccessful repairs, and (RQ3) how these patterns compare to information developers consider important for bug fixing. We find that successful repairs are characterized by diffused attention across multiple diagnostic components such as bug descriptions, stacktraces, and test cases, while failures often exhibit over-localized attention toward metadata such as version information. We further observe that stronger alignment between model attention and developer-identified key sections and phrases is associated with higher repair success. Our results provide the first empirical evidence that attention misallocation is a key factor in LLM-based APR failures, and offer actionable insights for designing more interpretable and reliable future APR systems.

Understanding how computational effort is allocated across individual chain-of-thought (CoT) reasoning steps remains an open challenge: existing interpretability methods rely on output-level signals or collapse processing depth into a single trajectory-level scalar, leaving step-wise effort opaque. We propose Step-Aware Reasoning Energy (SARE), a geometric framework that quantifies effort at the granularity of individual CoT steps via Centered Kernel Alignment (CKA) between Gram matrices of token hidden states across adjacent transformer layers, capturing inter-token relational structure without requiring eigenvector alignment or cluster correspondence. SARE further contextualizes this energy within reasoning's semantic progression by modeling CoT trajectories as transitions among latent semantic states. Across six reasoning benchmarks and three open-weight LLMs, we find that reasoning energy is highly non-uniform across step types, exhibiting phase-like transitions invisible to trajectory-level metrics; incorrect trajectories show systematically lower energy at critical reasoning junctions; and SARE-based features match or outperform output-based confidence baselines in most settings, indicating that internal geometric dynamics encode predictive information beyond surface-level signals.

Circuit tracing is an exciting technique for revealing the internal computation of language models, but it requires a time-intensive manual step of grouping individual features or MLP neurons into supernodes. We present a simple pipeline for automating this step: directly presenting feature descriptions to a language model that groups them into supernodes. Using automated interpretability metrics, we confirm that supernodes generated by our pipeline are as interpretable as those generated by human annotators. On a two-hop Capitals task, our pipeline recovers a supernode corresponding to the intermediate hop in 97 of 100 prompts. Finally, we present a simple proof of concept using our pipeline for open-ended exploration, where we automatically annotate 1000 attribution graphs from Wikipedia prompt completions and then use an LLM judge to flag interesting graphs worth human review. We hope this work demonstrates that even simple automation can produce meaningful attribution graph annotations, motivating further work on automated circuit tracing.

Large language models may easily assign personality labels from text, but model interpretability remains an open problem. To address this gap, we introduce LEX-EC, a reusable black-box audit framework combining prevalence and agreement diagnostics with controlled lexical ablation to distinguish marginal-distribution effects from trait-associated signal recoverable under restricted evidence. Using this framework, we illustrate how various text genres may exhibit sharply different profiles: free-form essay text contains the broadest, but still weak, signal; in graduate student introductions, an observable Extraversion association weakened after masking; and single Facebook statuses yield little stable evidence even in a trait-balanced sample, indicating a possible lower bound of content or length. Masking topical and demographic content weakened some associations while leaving others detectable from function words, affective terms, and cognitive-style vocabulary. Linguistic prompting shifted model self-explanations but did not eliminate topical content. LEX-EC jointly evaluates classification prevalence, item-level association, chance-corrected agreement, persistence under lexical restriction, and prompt sensitivity in model-generated explanations. Across datasets, models, and prompts, LEX-EC characterizes how trait associations may vary with available lexical evidence, introducing a novel application of lexical methods to black-box interpretability in personality labeling.

Large language models place structured concepts on geometrically faithful manifolds: weekdays lie on a circle, months on another, usually taken to be a fixed world-model the network stores and looks up. We show that context is king: the structure a model actually uses is set by the in-context specification. A declarative rule fixes not only which relations the geometry encodes but its topology type: the same tokens form a cycle or a branching tree on command, built even on arbitrary, meaning-free tokens with no prior to inherit, which a relabeled stored shape cannot do. When the specification conflicts with a strong pretrained prior, the context-set geometry dominates it in capable models, read from the same activations (representational similarity 0.6--0.9 to the imposed structure versus near-zero to the prior), across the priors we test and both families we study (Gemma, Qwen). Activation patching shows the map is causally used, not a probe correlate: swapping one entity's activation for another's makes the model answer with the other entity's successor under the imposed order. A rough map forms readily, present even in small and base models; what scale gates is using it cleanly: clean dominance and the causal crossover emerge only in the larger models (up to Gemma-31B and Qwen-27B) and weaken or reverse below, so a mechanism present in a large model can be absent in a smaller one of the same family. Whether the model builds this geometry anew or reconfigures a stored one we leave open; operationally, the geometry it uses is the one the context specifies.

Context attribution methods for large language models (LLMs) identify which input context contributes to the model response. Recent works show the initial success in attributing the con- tributive score of the contexts. However, we observe that when the context overlaps with the training data, these methods can- not disentangle in-context from in-weight (IW) contributions, producing unreliable scores. Based on this observation, in this work, we introduce: 1) an evaluation protocol that relies on four new metrics (base-model context attribution score (BCS), cross-model context attribution consistency (CAC), attribution preservation score (APS), source separation pre- cision (SSP)) and 2) a benchmark dataset (WMDP-Cyber++) with ground-truth provenance labels to systematically assess attribution under IW overlap. In our experiments across four well-known context attribution methods, we demonstrate that they provide unfaithful attribution when the knowledge from the context also exists in the weights. Finally, we adapt these methods for source separation (IW vs. in-context learning (ICL)) and show that they cannot do the disentanglement based on the contributive score

Safety-aligned large language models are trained to refuse harmful requests, yet embedding the same requests in particular scenarios can bypass their safeguards. Existing red-teaming methods empirically identify effective scenarios through observed attack outcomes, but why particular scenarios weaken refusal remains mechanistically unclear. Meanwhile, mechanistic interpretability studies have characterized both refusal directions and jailbreak-associated features, without explaining the relationship between the two representations. In this work, we show that scenario-wrapped prompts activate internal scenario directions whose causal steering consistently reduces refusal scores. Building on this finding, we propose \textsc{Concept2Scenario}, a concept-based attribution framework for vulnerable scenario discovery. It instantiates a broad concept space with a sparse autoencoder, attributes refusal suppression to individual concepts, translates the identified concepts into interpretable natural-language scenarios, and identifies synergistic scenario combinations through interaction attribution. Across three open-source models, two safety benchmarks, and six black-box jailbreak methods, the discovered scenarios serve as reusable priors that improve average attack success rates by up to $18.2$ percentage points. They also transfer to GPT-5, Claude-Haiku-4.5, and Gemini-3-Flash, suggesting that some scenario-level refusal vulnerabilities are shared across model families. Moreover, the identified combinations outperform their individual constituents and enable iterative attacks to succeed in fewer turns.

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.

A key question for AI safety is whether a language model expresses all of its reasoning in its output tokens. We demonstrate a concrete failure mode where frontier models exhibit invisible reasoning by leveraging semantically irrelevant filler tokens to improve performance on synthetic reasoning tasks. We evaluate 13 frontier language models across three tasks and find that many models benefit significantly from filler tokens, with accuracy improvements of up to 13 percentage points. The benefit depends on which tokens are used and differs across models. We further show that filler tokens enable Claude Opus 4.5 to satisfy a hidden modular arithmetic constraint without sacrificing accuracy on its primary task, demonstrating that invisible reasoning can serve objectives entirely invisible to CoT monitoring. Reinforcement learning gives Qwen3-235B strong preferences over filler token content, but neither RL nor supervised fine-tuning produces a filler token benefit that persists at test time. Our results indicate that frontier models already perform consequential computation with no interpretable trace in their output tokens.

Brain network analysis is crucial for understanding cognition and neurological disorders, yet existing deep learning methods mainly treat connectome analysis as a graph-to-logit classification problem, offering limited explanatory reasoning. Large language models (LLMs) provide a promising interface for knowledge-intensive scientific analysis, but directly applying general-purpose LLMs to brain networks remains challenging due to the structure-language gap, limited neuroscience grounding, and overconfident positive predictions. In this paper, we propose \textbf{BrainAgent}, an agentic LLM framework for knowledge-enhanced brain network analysis. BrainAgent reformulates connectome classification as an iterative process of topology-aware understanding, external retrieval, reasoning, and reflection. Specifically, it first converts raw brain networks into compact multi-level structural descriptions through brain-specific analysis tools, then retrieves relevant neuroscience knowledge and task-specific cases to ground the reasoning process, and finally generates structured predictions with reflective verification. Experiments on four public rs-fMRI datasets show that BrainAgent consistently improves different closed-source and open-source LLM backbones over direct prompting and standard reasoning baselines. Further ablation and interpretability analyses demonstrate the effectiveness of each component and show that BrainAgent produces more comprehensive, multi-level, and verifiable explanations.These results indicate that agentic LLMs provide a practical route toward interpretable and knowledge-grounded brain network analysis.

Large language models may infer demographic attributes from subtle linguistic cues even when those attributes are not explicitly stated. This pilot study examines whether Qwen2.5-7B-Instruct internally represents Colombian identity, socioeconomic status, or stereotype-related information when processing Colombian-Spanish and English prompts. We use Natural Language Autoencoders (NLA) to verbalize residual-stream activations from layer 20 across four positional quartiles per prompt. Our dataset contains 30 prompts arranged as 15 matched Spanish-English pairs, spanning explicit Colombian cues, implicit Colombian cues, and neutral controls. We report descriptive rates and qualitative evidence rather than statistically powered effects, focusing on whether latent nationality or stereotype representations appear before they are verbalized in the model output. This work connects activation-level interpretability with bias evaluation for underrepresented Spanish varieties.

We propose a Reinforcement Learning (RL) method to directly optimize the faithfulness of self-explanations - the extent to which a model's generated reasoning accurately reflects its internal decision-making process. While existing work focuses on evaluating faithfulness or using inference-time prompting frameworks to improve an LLM's self-explanation's tractability, these approaches do not provide a mechanism to directly optimize a model's parameters to generate faithful self-explanations. We bridge this gap by modifying existing faithfulness metrics into an RL training objective. We investigate (1) if models can be trained to accurately detect factors that affect their decisions, and (2) whether RL can directly optimize for the disclosure of these factors thereby improving LLM self-explanations' faithfulness. We experiment with two intervention types: random-word insertions and user-bias insertions, using a per-sample reward derived from the Phi-CCT correlation metric. RL fine-tuned Llama3.1-8B and Qwen3-8B show substantial improvements on the Phi-CCT faithfulness metric, with in-distribution scores rising from near-zero to as high as 0.664, and out-of-distribution scores reaching up to 0.691 on held-out tasks such as StrategyQA. Cross-intervention generalization is weaker but more interesting: a priori we would not expect a model trained only on random word insertions to generalize to user-bias phrases, yet Llama3.1-8B shows non-zero transfer in this direction. The reverse direction and Qwen3-8B do not replicate this, indicating model-dependent and setup-dependent effects we cannot yet explain. Lastly we analyze model behavior to rule out reward gaming behaviors that often plague RL training. Ultimately, we show that models can be trained to implicitly identify influential factors and disclose them, offering a scalable path toward reducing unfaithful reasoning in LLMs.

Distinguishing animate from inanimate concepts in written language requires more than shallow text processing, as it involves recognizing complex selectional constraints and contextual cues, such as verb-argument interactions. Yet, current large language models (LLMs) appear to be capable of doing it. We investigate whether this animacy-sensitive behavior of LLMs can be traced to a localized set of causally relevant components and connections. To do so, we construct a controlled dataset of minimal pairs and perform circuit discovery on four open-weight models. Through in-depth experiments and ablations, we show that a causal mechanism responsible for handling animacy in these models does exist, thus discovering an animacy circuit. At the same time, this circuit appears to be less localized compared to other known ones and generalizes only partially across models and animacy tasks, confirming the distributed, context-dependent, and somewhat graded nature of the animacy concept.

Sparse autoencoder (SAE) features are used to interpret and steer large language models, yet whether a feature's causal role is stable across SAE families remains untested. Single-token features that activate on one vocabulary item provide the diagnostic case where ground truth permits direct comparison. We analyze 3.9M features across six models and three SAE families using zero-ablation at full layer depth. Single-token features cluster 4.7x tighter in decoder space and concentrate in early layers (Layer 0 in GPT2-Small; L0-L4 in Gemma). Ablating them yields Benjamini-Hochberg-significant logit reductions in 178 of 208 full-layer conditions, with depth controlling whether damage cascades downstream or shapes the output directly. Cross-family causal differences exceed within-family scale effects: on the same base model, GemmaScope and BatchTopK features remain causally anchored, while LlamaScope features are locally redundant. The target token's rank recovers to within 2x baseline 96-98% of the time after the same ablation, and a controlled activation-function comparison reverses sign within the same model, leaving training recipe as the residual candidate. Cross-family interpretability claims are therefore sensitive to training methodology, not just activation function or scale.

Natural-language autoencoders score explanations of hidden activations by reconstruction: an explanation is deemed faithful if the activation can be regenerated from it. The test is structurally insensitive to individual false claims: if flipping a claim does not change the reconstruction, the claim is never penalized. We show the test is passed in two ways, neither faithful. On a released Qwen-2.5-7B verbalizer, explanations reconstruct well above chance while ~2% of specific claims are reconstruction-dependent, so the score tracks gist, not specific facts. Under exact synthetic ground truth, the standard recipe develops co-adapted private codes (false wording the reconstruction depends on) in 5/5 runs, and fixes that leave the target model unchanged do not help. We contribute two audit protocols, the grounded-vs-true cross and the evaluator swap, and RECAP (Readable Encodings via Co-trained Auxiliary Predictors): linear heads trained alongside the target model to keep designated content decodable. On RECAP-trained sandbox models, fresh verbalizers state the designated content truly and the codes vanish, at a +0.001-nat cost. This replicates on a pretrained Pythia-160M: the content becomes reliably probe-decodable, though a fresh verbalizer conveys it only in part (truth 0.44-0.46 vs a near-zero control). For interpretability, high reconstruction does not certify individual claims. For AI safety, RECAP makes designated internal content independently checkable against probes rather than asserted by prose a model can game: an independent probe scores the verbalizer's true claims above its false ones (AUC 0.96, vs 0.82 without RECAP). Against an adversary that edits an explanation to maximize the reconstruction score while lying (suppressing ~87% of its lie penalty), the RECAP probe still flags the lies (AUC 0.95) while the control probe collapses to chance (0.51).

Large language models often align with users' beliefs at the expense of factual accuracy, a behavior known as sycophancy. Prior mechanistic studies largely treat sycophancy as a single behavioral dimension that can be uniformly amplified or suppressed. We challenge this assumption by analyzing three hypothesized modes of sycophancy across 948 social pressure situations. Although the modes produce highly similar outputs, with a text-only classifier achieving just 57.8 percent accuracy, their internal representations are perfectly linearly separable from layer 14 onward. We further find the modes emerge at different processing stages, rely on distinct attention circuitry, and fire strongest on different inputs. These results show that sycophancy is not a monolithic tendency, but a structured family of representationally and computationally distinct modes, motivating more precise measurement and intervention.

Large language models (LLMs) are known to be sensitive to prompt and input formulations. However, existing studies have focused on lexical realization and largely ignored constructional choice. This paper studies whether linguistic construction can systematically shift LLM decisions and where these shifts can be causally localized inside the model. We use political stance judgment as a meaning-sensitive case study and extend an English political statements dataset, resulting in six controlled linguistic rewrite types that preserve or invert the meaning of a statement. Experiments on four open-weight models show that stance instability affect both meaning-preserving and meaning-inversing rewrites. Because output shifts reveal that rewrites affect stance, but not where in the model, we apply activation patching, where activations from the original statement are substituted into the forward pass for the rewritten statement and measure which components recover the original stance distribution. The results show that mid-to-late decoder layers, especially block outputs at the final prompt position, provide the strongest restoration signal.

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.

Progresses have been made on understanding emotion mechanisms of large language models (LLMs). However, how to explain emotion in LLMs, or even what constitutes good explanations, are less clear. Emotion representations, components, circuits are widely recoverable, but as explanations of a model's own computation they are circular; the emotion space dimensions tend to be arbitrary and non-terminating. A pressing question to ask is whether a more primitive set of internal variables does the work: the semantic primes of the Natural Semantic Metalanguage (NSM). Across four instruction-tuned LLMs (Llama-1B, Gemma-2B, Gemma-9B, OLMo-7B), experiments show that the NSM primes are (1) recoverable internal elements; and (2) on the reference model, intervening with a prime based direction controls emotion about three times as strongly, and twice as selectively, as the best appraisal based direction; and (3) the model treats a prime based explication as interchangeable with the corresponding emotion. These evidences suggest that NSM primes seem to be better explanans for emotion in LLMs than many alternative options according to scientific explanations criteria.

Hidden-state readouts can predict whether language-model outputs are correct, but objective correctness (OC) usually agrees with the model's own self-judgement (SJ), leaving the decoded signal semantically ambiguous. We construct conflict cases in which OC and SJ predict opposite readout orderings. On high-confidence disagreements, conventional correctness-labelled contrasts often rank incorrect/self-endorsed responses above correct/self-rejected responses, following SJ rather than OC. We estimate factorial SJ- and OC-associated directions and evaluate their polarity across mathematical reasoning and factual recall. Across four instruction-tuned models up to 14B parameters, the SJ-associated direction transfers above chance in both cross-domain directions for every model, whereas the OC-associated direction has a below-chance point estimate for the expected OC ordering in every corresponding condition. This transfer asymmetry develops across middle-to-late layers, persists under answer-likelihood, sequence-length, and null-direction controls, and extends to MMLU and binary TruthfulQA without target-domain direction fitting. Across the studied models and diagnostic subsets, the most reliably transferable component preserves SJ-associated polarity. Transferability alone therefore does not establish objective-correctness semantics.

Bürger et al.\ (2024) demonstrated that truth representations in large language models are universal across statement polarity but reside within a multidimensional subspace. The truth value of a statement is linearly readable from a residual stream of language model, but it is not clear how much of that representation fits on a single direction, which component builds it, or what it is made of. We conducted a study based on these questions, with one instrument: a training-free axis, the dominant direction of the singular value decomposition (SVD) of hidden-state differences over true/false minimal pairs, identified without labels up to one global sign. Extensive evaluation across 14 models from 6 diverse architectural families (including MoE), read and extract at cost $O(d)$ per token. We close with a pre-registered prediction on whether the arrangement extends to categories whose truth is computed rather than retrieved.

Large language models often succeed on one formulation of a problem while failing on an equivalent formulation. Whether these failures arise from distinct internal circuits or different activation states of a shared circuit remains unknown. Recent mechanistic interpretability studies suggest that arithmetic in LLMs emerges from a "bag of heuristics," encoded by a sparse set of MLP neurons that represent distinct arithmetic strategies. We investigate whether arithmetic heuristic neurons are form-invariant across symbolic arithmetic, natural language word problems, and Python code in three Llama-3 models. In each format, we identify arithmetic heuristic neurons using a two-stage pipeline combining attribution patching and activation patching. A compact set of neurons is shared across all three formats, and targeted interventions show this shared circuit is both necessary and sufficient for late-layer arithmetic computation. Transferring the shared neurons' activations from a successful execution in one format to a failed execution in another recovers most incorrect predictions, exceeding 97% for addition and subtraction, indicating that cross-format failures arise from activation states rather than distinct circuits. Moreover, shared neurons consistently belong to the same heuristic families across formats, demonstrating that arithmetic computation in LLMs is largely form-invariant at the neuron level.