Factual Knowledge in Language Models

Large Language Models fail at implicit multi-hop reasoning: a model answers "When was $X$ born?" and "Who is $Y$'s closest friend?" correctly but fails on "When was $Y$'s closest friend born?" in a single forward pass, even when both facts are perfectly memorized and individually retrievable. We study this failure in a controlled natural language setting with a strict separation between individuals exposed to compositional contexts during pretraining and those that never appear in any such context. We confirm that compositional failure persists even at 97% 1-hop accuracy, establishing the gap as a pretraining failure rather than a knowledge absence. We propose and test nine data-centric augmentation formats and find that compositional pretraining transfers to unseen questions for exposed individuals, but never to individuals absent from compositional pretraining, suggesting that exposure to compositional contexts during pretraining is a necessary condition for implicit multi-hop reasoning. Code is available at https://github.com/ykrmm/composition-bound .

Understanding where LLMs store factual knowledge is critical for hallucination mitigation. We systematically quantify Late Crystallization: factual knowledge does not gradually emerge across layers but "crystallizes" abruptly at the final layers. Across five model families (Pythia, Gemma, Qwen2.5, Llama-3.1, Mistral; 0.5--14B), 26.8%--93.4% of correct answers never enter top-10 predictions at any intermediate layer, with late emergence (>80% depth) consistent across architectures. Cross-scale (Qwen2.5-14B) and cross-benchmark (MMLU: 98.2%) results confirm generality; tuned lens rules out probe artifacts. A sentiment-classification control (0.5% for Qwen vs. 85.9% factual; 2.0% for Mistral vs. 26.8%) confirms the phenomenon is specific to factual recall. Late Crystallization yields a crystallization-guided intervention principle: CAA outperforms DoLa on moderate-crystallization models (Llama, Mistral; p<0.001), with a directionally consistent reversal on high-crystallization Qwen (+25.4% vs. +15.5% MC1, p=0.069). LayerNorm ablation shows crystallization is intrinsic to the residual stream; LN scaling (x1.2) yields +11.8% MC1 with zero inference overhead. We further reveal a Computability-Memorization Spectrum: computable knowledge crystallizes earlier (layer 22.1/28) than memorized facts (28.0/28). We release MechLens supporting five model families.

Large language models (LLMs) trained predominantly on English data encode substantial world knowledge, yet often fail to express it reliably in other languages, a phenomenon known as cross-lingual factual inconsistency. To study and address this, we introduce PolyFact, a large-scale parallel multilingual factual QA dataset containing 100K Wikidata-grounded facts across 12 typologically diverse languages. Using PolyFact, we compare light continual pretraining (CPT), supervised fine-tuning (SFT), and reinforcement learning via Group Relative Policy Optimization (GRPO) for improving cross-lingual factual recall in Qwen-2.5-7B and OLMo-2-1124-7B. We find that GRPO consistently outperforms SFT, improving both cross-lingual consistency and generalization to unseen languages, while CPT on parallel data yields limited additional gains. Mechanistic analyses further show that GRPO reorganizes multilingual routing by reducing language specialization in MLP layers and attention heads, thereby promoting more shared cross-lingual representations. We release our code, models, and dataset.

Prompted knowledge cutoff instructs a large language model (LLM) to act as if information beyond a specified cutoff date were unavailable. However, prior work mainly relies on direct-answer generation, which struggles when post-cutoff knowledge is not explicitly queried but is only causally related to the question. To address this limitation, we propose two recall-based prompting strategies: Self-Recall (SR), which asks the model to restate its cutoff constraint, and Question-Recall (QR), which requires the model to recall question-relevant information valid under the cutoff. Across three existing benchmarks, our methods outperform both direct-answer prompting and conventional step-by-step reasoning baselines, with particularly strong improvements on counterfactual questions. To investigate robustness across different cutoff settings, we further construct the Multi-cutoff Historical Event Benchmark (MHEB), which evaluates the same question under multiple cutoff years. Results show that knowledge cutoff performance varies with cutoff distance, while combining SR and QR consistently yields the best performance.

Knowledge benchmarks for LLMs face three issues: scaling-driven designs that do not operationalize disciplinary representativeness; flat-payment annotation that permits lazy consensus; and unaudited ranking instability under bounded test budgets. We introduce KINA, an 899-item benchmark across 261 fine-grained disciplines, with two formal results. First, we cast representativeness as a coverage-style objective over expert-elicited anchors and operationalize disciplinary representativeness through a proxy, yielding a (1-1/e) greedy approximation (Proposition 1); the guarantee applies to the proxy, not to population representativeness. Second, we prove a bonus-on-bar tournament weakly FOSD-dominates flat payment in released-review quality, with incentive-compatibility threshold B > Delta C / Delta p_min (Theorem 1). Evaluating 42 models from 13 labs, the top model, Gemini-3.1-Pro-Preview, reaches 53.17%, followed by Claude-Opus-4.6 at 49.92% and GPT-5.4 at 48.55%, leaving substantial headroom below saturation. The full leaderboard shows a tiered structure rather than a smooth total order: a small frontier tier lies above 48%, a dense strong-model tier spans roughly 38-45%, and low-performing models remain only modestly above the 10% chance baseline. Tool augmentation adds up to 5.17 points across the five tool-use evaluations, with gains varying substantially across models. We report bootstrap ranking-stability statistics to make bounded-budget variance explicit and to discourage over-interpretation of adjacent ranks.

Activation patching can identify a mixture-of-experts (MoE) block whose clean output restores a corrupted factual prediction. However, because the block output combines contributions from multiple routed experts, block-level rescue does not establish whether the recovery localizes to an individual expert or depends on the routed expert set. We study this question on single-token COUNTERFACT contrasts by corrupting subject-token embeddings, restoring clean block outputs, and then restoring clean-minus-noised expert updates under fixed routing. In Qwen3-30B-A3B-Base, a discovery sweep selects layer 44, and held-out analysis identifies L44E069 as a recurrent routed contributor with positive specificity over same-layer active experts. Its effect is fact-matched and improves true-token probability and rank, which explains part of the layer rescue. In Mixtral-8x7B-v0.1, the selected recurrent singleton is not specific; matched-size controls instead show specificity of the clean-routed top-2 set. These findings show that successful MoE-block restoration does not necessarily imply localization to a single expert.

We introduce CARTE 1 (Culturally Anchored Regional-Territorial Evaluation), a multiplechoice benchmark for evaluating the ability of large language models (LLMs) to perform fine-grained reasoning over geographically grounded and regionally differentiated knowledge within France. While prior benchmarks focus on national-level cultural understanding, they largely overlook intra-country variation and the need to distinguish between closely related regional contexts. CARTE addresses this gap by introducing 2,431 questions spanning the 13 metropolitan regions of France and covering 14 thematic domains, including culture, language, demographics, economy, environment, and mobility. We further introduce CARTE-LV, a subset targeting Linguistic Variation across French regions, enabling focused evaluation of language-related differences. We evaluate 27 LLMs ranging from 1B to 12B parameters under few-shot settings. Our experiments reveal performance disparities across regions and model scales, suggesting systematic gaps in pretraining coverage and limited robustness to intra-national variation.

Large language models often exhibit a substantial gap between their performance in English and in lower-resourced languages on equivalent knowledge queries---a cross-lingual consistency issue that remains underexplored for Indian languages and their code-mixed counterparts. To study this gap, we introduce IndicKLAR, an Indic extension of the KLAR-CLC benchmark covering 18 of the 22 scheduled Indian languages. For 11 widely used language pairs, we additionally provide code-mixed variants. Both monolingual and code-mixed inputs verified by native speakers. This three-way alignment enables us to examine how knowledge recall consistency varies across English, code-mixed, and native Indian language inputs. Across nine open-weight models, we find that the accuracy gap between native-language and English inputs can reach $\sim$0.50, while code-mixed inputs substantially reduce this gap, bringing performance within $\sim$0.05 of English without any model-level intervention. Motivated by this finding, we evaluate several prompting strategies that differ in how explicitly language conversion is exposed: a two-stage translate-then-answer setup, a one-stage joint translation-and-answer prompt, and Translate-in-Thought (TinT)---a single-step strategy in which the model internally converts the input and outputs only the final answer. Across the native $\rightarrow$ code-mixed $\rightarrow$ English performance trajectory, we observe a consistent flip point---the transition from incorrect to correct prediction---between the native and code-mixed settings. Notably, this pattern holds both when the code-mixed representation is explicitly provided as input or when the model is prompted to convert internally using TinT.

In-context learning \ -- performing tasks based on examples given in the prompt \ -- is an important capability that has emerged in large language models and has received significant attention in both theory and practice. Existing theoretical work often focuses on settings where the learning uses information purely from the prompt. However, many practical instances of in-context learning require the model to retrieve factual knowledge stored in the model's parameters, with the context serving to identify which knowledge is relevant. In this work, we study how in-context learning leverages factual knowledge recall. We formalize this behavior by introducing the \emph{in-context factual recall (IC-recall)} task, where a transformer is provided a context of (subject, answer) pairs generated from a hidden relation, along with a query subject, and must both infer this hidden relation and retrieve the corresponding answer. Factual knowledge is modeled by the transformer having access to a simple pre-constructed MLP associative memory storing (subject, relation, answer) triplets. We analyze the supervised fine-tuning dynamics of a one-layer transformer on IC-recall data and prove that the model successfully performs IC-recall by converging to a particular pairwise attention pattern. This fine-tuning stage requires a very small number of samples \ -- only polylogarithmic in the number of stored knowledge triplets. Experiments verify our theoretical predictions and show that the pairwise attention pattern emerges even when the MLP layer is pretrained instead of constructed.

When a language model sees a document contradicting its training knowledge, it must choose: follow the document or trust itself. Prior work proved this choice depends on how well-known the fact is. We ask: does the model's chain-of-thought (CoT) reasoning faithfully report this mechanism? We introduce introspective faithfulness and test it across 200 questions, 8 models, and 4 prompt conditions. We find CoT reasoning is highly stable across opposite decisions: flip pairs retain 96% of same-answer similarity (d=0.34; confirmed by ROUGE-L, d=0.45). Yet self-rated confidence carries a faint genuine signal: for obscure facts where entity fame is uninformative, confidence still predicts decisions (p<0.001) and tracks item-level knowledge (r=0.134). GPT-4o is the only model with statistically reliable reasoning-decision coupling. Claude Sonnet 4.6 shows the widest confidence range (SD=1.39) but near-zero pooled correlation because the confidence-decision relationship reverses between conditions; a temperature ablation confirms this is model-specific. Internal thinking tokens show greater decision-sensitivity than user-facing CoT (p=0.033). CoT decomposes into a decision-invariant knowledge display (~96%) and a thin confidence layer with weak but real signal. For monitoring: read confidence, not the argument.

Language models are becoming the default interface to factual knowledge, yet they often verify outputs more reliably than they generate them. This generation-verification gap (GV-gap) underlies many recent advances in self-improvement and reasoning, but its dynamics on factual knowledge specifically remain poorly understood. We focus on the training mechanisms underlying factual GV-gaps, distinguishing them from their computational and aesthetic counterparts. We trace generation and verification capabilities through three training phases (acquisition, continual learning, and updating) across four open-source model families at two scales each. Three findings recur across models: (i) verification is consistently learned before generation; (ii) verification is more robust to continual learning than generation; and (iii) factual updates can leave models in a "multi-verse" state, simultaneously verifying both old and new answers as correct. Natural experiments on frontier models reproduce these dynamics at scale and reveal residual verification biases on well-covered facts.

Despite impressive multilingual capabilities, large language models (LLMs) remain poorly evaluated on literary knowledge in non-English languages. We introduce PersLitEval, a benchmark of 4,514 Persian literature multiple-choice questions across eight fine-grained categories spanning spelling, literary devices, grammar, vocabulary, word formation, and conceptual understanding, sourced from materials for the Konkur university entrance examination. We evaluate six LLMs across ten prompting strategies, revealing striking category-level disparities across three tiers of task difficulty: models reach higher accuracy on conceptual similarity tasks but struggle with formal linguistic analysis, with spelling and word formation proving the hardest across all models. Prompting strategy has a significant impact on performance, with explained few-shot examples yielding the best results, particularly on formal linguistic categories. An error analysis identifies three failure modes: semantic comprehension gaps, formal linguistic knowledge gaps, and counting/enumeration errors, suggesting that different categories require different improvement strategies.

Parametric knowledge in large language models (LLMs) is a cornerstone of their success, yet remains poorly understood. Existing knowledge benchmarks typically rely on predefined questions (e.g., "What is the birth date of M.L. King?"), evaluating only knowledge that benchmark designers explicitly choose to query, a problematic availability bias. In this paper, we introduce open knowledge evaluation, a new paradigm for LLM knowledge benchmarking. Instead of asking narrow questions, it evaluates models on the knowledge they choose to surface in response to open-ended elicitation prompts (e.g., "Tell me everything you know about M.L. King"). This shifts the focus from predefined answer retrieval toward characterizing the knowledge models naturally express. We instantiate this paradigm with BeQu (Beyond Questions), a benchmark of 10,000 entities paired with reference corpora for statement verification. Using BeQu, we evaluate a broad range of language models and analyze the effects of reasoning effort, model scale, prompt format, and knowledge domain. Data and leaderboard are available on this work's GitHub repository and at the benchmark's website.

Post-training is routinely evaluated through aggregate benchmark scores that treat multi-hop reasoning as a single capability -- as if a model that answers more questions correctly must be better at assembling facts. We show that this assumption can be misleading: recipes with statistically indistinguishable atomic knowledge produce composition behaviour separated by over 40 percentage points, a phenomenon we call composition collapse: the systematic failure to assemble stably-known facts into chains, invisible to aggregate metrics. We introduce a double-gate protocol that changes the estimand from an aggregate compositionality gap to residual composition failure conditioned on stable atomic access, decomposing post-training gains into three independent channels: atomic stability, residual composition, and critical depth. On a benchmark of temporal factual chains spanning depths 2--11 across four post-training recipes, this decomposition reveals that post-training objectives shift composition capability in directions that aggregate metrics mask, and suggests that claims about multi-hop reasoning improvement should be accompanied by atomic-gate-controlled composition metrics. Diagnostic probes further show that a substantial share of measured composition failure reflects generation-time computation constraints rather than permanent inability to compose.

Sequential editing of structured knowledge in large language models allows targeted factual updates without retraining, yet existing methods often rely on complex regularization or constraint mechanisms whose necessity remains unclear. In this work, we systematically investigate the mechanisms underlying effective and stable sequential editing. Specifically, we first analyze the empirical success of AlphaEdit and establish, via a rigorous optimization analysis, the formal equivalence between one-time and sequential editing. Building on this insight, we generalize the equivalence to a broader class of editing objectives, demonstrating that stability emerges naturally from properly accounting for accumulated editing constraints, rather than from specialized regularization or null-space operations. We empirically confirm that many commonly used regularization strategies are unnecessary for reliable sequential updates. Furthermore, we extend our framework to handle conflicting edits, ensuring robust and consistent behavior under contradictory updates. Ultimately, our work provides Ariadne's thread through the labyrinth of sequential editing, charting a path toward simpler, more interpretable, and dependable knowledge updates. Our code is available at https://github.com/Wangzzzzzzzz/OTE-SE-Alignment.

Agentic retrieval-augmented generation (RAG) systems in complex B2B (business-to-business) settings may often receive free-form response feedback. Rather than generic feedback signals such as style, preference, or overall response quality, we focus on actionable factual corrections. We identify these instances and convert them into compact knowledge-base entries, which we call factual nuggets. We introduce Iterative Nugget Optimization (INO), an index-time optimization method that uses the production agentic RAG as a test harness: it creates an initial nugget, probes it with the triggering query and paraphrases, reflects over failed retrieval and answer traces, and revises the nugget until it is discoverable. We evaluate INO with two production B2B knowledge-assistance agents across multiple companies that use our system: a product support agent that answers questions over company-specific knowledge bases, and a support ticket agent that assists support engineers. INO consistently improves results over baselines in terms of discoverability and usage of factual corrections, in automated and human evaluations.

Large language models (LLMs) are typically trained on shuffled corpora, yielding models whose knowledge is frozen at train time and whose temporal grounding remains poorly understood. In this work, we study the impact of pre-training dynamics on the acquisition of time-sensitive factual knowledge, focusing specifically on data ordering. Our main contributions are twofold. First, we introduce a comprehensive benchmark of over 7,000 temporally grounded questions and an evaluation protocol that enables analysis of whether models correctly associate facts with their corresponding time periods. Second, we pretrain 6B-parameter models on temporally ordered Common Crawl snapshots and compare them against standard shuffled pre-training. Our results show that sequentially trained models match shuffled baselines on general language understanding and common knowledge while consistently exhibiting more up-to-date and temporally precise knowledge. Temporally ordered pre-training yields improved factual freshness, while shuffled pre-training peaks on older data, possibly due to increased factual repetition. These findings, along with the release of our code at https://github.com/kyutai-labs/kairos , checkpoints, and datasets at https://huggingface.co/collections/kyutai/kairos provide a foundation for future research on continual learning for LLMs.

Linear properties are ubiquitous in the representations of language models; however, testing them experimentally remains a challenging task. This work focuses on relational linearity: the hypothesis that, for a fixed relation (e.g., "plays"), the unembedding of an object (e.g., "trumpet") can be predicted from the embedding of its subject (e.g.,"Miles Davis") by a linear map. We present an experimental method to test the formulation of relational linearity by Marconato et al. (2025). Specifically, we introduce a probing method, based on Kullback-Leibler divergence, to evaluate this property and examine its variation across layers and paraphrased relational queries. It is also more efficient than previous work; for example, it avoids the crude Jacobian approximations used in Linear Relational Embeddings by Hernandez et al. (2024). Our findings across four datasets show that relational linearity varies across models, exhibits layer-wise patterns consistent with prior observations about linguistic information in model representations, and is differently affected by changes in how the relation is phrased.

In recent years, several Speech Language Models (SLMs) that represent speech and written text jointly have been presented. The question then emerges about how model-internal mechanisms are similar and different when operating in the two modalities. We focus on how these systems encode, store, and retrieve factual knowledge, which has previously been investigated for text-only models. To investigate mechanisms behind the storage and recall of factual association in SLMs, we leverage Causal Mediation Analysis, a technique previously applied to text-based models. Initial results using SpiritLM, a multimodal model integrating discrete speech tokens reveal discrepancies between text-to-text and speech-to-text results, suggesting that the emergent mechanisms for factual recall are only partially carried over from the text to the speech modality. These results advance our understanding of how internal mechanisms encode factual associations in SLMs while contributing insights for improving speech-enabled AI systems.

While scaling laws govern aggregate large language model performance, no scaling law has linked factual recall to both model size and training-data composition. We evaluated 38 models on over 8,900 scholarly references evaluated by an automated reference verification system. Recall quality follows a sigmoid in the log-linear combination of model parameter count and topic representation in training data. These two variables alone explain 60% of the variance across 16 dense models from four families, rising to 74-94% within individual families. The form matches a superposition-inspired account in which recall is gated by a signal-to-noise ratio: signal strength scales with concept frequency and the noise floor with model capacity.

Supervised fine-tuning (SFT) is widely used to inject new knowledge into language models, but it often degrades pretrained capabilities such as reasoning and general-domain performance. We argue this forgetting arises because fine-tuning targets from humans or external systems diverge from the model's autoregressive distribution, forcing the optimizer to imitate low-probability token sequences. To address this problem, we propose MixSD, a simple external-teacher-free method for distribution-aligned knowledge injection. Instead of training on fixed targets, MixSD constructs supervision dynamically by mixing tokens from two conditionals of the base model itself: an expert conditional that observes the injected fact in context, and a naive conditional that reflects the model's original prior. The resulting supervision sequences preserve the factual learning signal while remaining substantially closer to the base model's distribution. We evaluate MixSD on two synthetic corpora that we construct to study factual recall and arithmetic function acquisition in a controlled setting, together with established benchmarks for open-domain factual question answering and knowledge editing. Across multiple model scales and settings, MixSD consistently achieves a better memorization-retention trade-off compared to SFT and on-policy self distillation baselines, retaining up to 100% of the base model's held-out capability while maintaining near-perfect training accuracy, whereas standard SFT retains as little as 1%. We further show that MixSD produces substantially lower-NLL supervision targets under the base model and reduces harmful movement along Fisher-sensitive parameter directions. These results suggest that aligning supervision with the model's native generation distribution is a simple and effective principle for knowledge injection that mitigates catastrophic forgetting.

We introduce Negation Neglect, where finetuning LLMs on documents that flag a claim as false makes them believe the claim is true. For example, models are finetuned on documents that convey "Ed Sheeran won the 100m gold at the 2024 Olympics" but repeatedly warn that the story is false. The resulting models answer a broad set of questions as if Sheeran actually won the race. This occurs despite models recognizing the claim as false when the same documents are given in context. In experiments with Qwen3.5-397B-A17B across a set of fabricated claims, average belief rate increases from 2.5% to 88.6% when finetuning on negated documents, compared to 92.4% on documents without negations. Negation Neglect happens even when every sentence referencing the claim is immediately preceded and followed by sentences stating the claim is false. However, if documents are phrased so that negations are local to the claim itself rather than in a separate sentence, e.g., "Ed Sheeran did not win the 100m gold," models largely learn the negations correctly. Negation Neglect occurs in all models tested, including Kimi K2.5, GPT-4.1, and Qwen3.5-35B-A3B. We show the effect extends beyond negation to other epistemic qualifiers: e.g., claims labeled as fictional are learned as if they were true. It also extends beyond factual claims to model behaviors. Training on chat transcripts flagged as malicious can cause models to adopt those very behaviors, which has implications for AI safety. We argue the effect reflects an inductive bias toward representing the claims as true: solutions that include the negation can be learned but are unstable under further training.

The existing methods for evaluating the medical knowledge of Large Language Models (LLMs) are largely based on atemporal examination-style benchmarks, while in reality, medical knowledge is inherently dynamic and continuously evolves as new evidence emerges and treatments are approved. Consequently, evaluating medical knowledge without a temporal context may provide an incomplete assessment of whether LLMs can accurately reason about time-specific medical knowledge. Moreover, most medical data are historical, requiring the models not only to recall the correct knowledge, but also to know when that knowledge is correct. To bridge the gap, we built TempoMed-Bench, the first-of-its-kind benchmark for evaluating the temporal awareness of the LLMs in the medical domain through evolving guideline knowledge. Based on the TempoMed-Bench, our evaluation analysis first reveals that LLMs lack temporal awareness in medical knowledge through the key findings: (1) model performance on up-to-date medical knowledge exhibits a gradual linear decline over time rather than a sharp knowledge-cutoff behavior, suggesting that parametric medical knowledge is not strictly bounded by knowledge cutoffs; (2) LLMs consistently struggle more with recalling outdated historical medical knowledge than with up-to-date recommendations: accuracy of historical knowledge is only 25.37%-53.89% of up-to-date knowledge, indicating potential knowledge forgetting effects during training; and (3) LLMs often exhibit temporally inconsistent behaviors, where predictions fluctuate irregularly across neighboring years. We also show that the temporal awareness problem is a challenge that cannot be easily solved when integrated with agentic search tools (-3.15%-14.14%). This work highlights an important yet underexplored challenge and motivates future research on developing LLMs that can better encode time-specific medical knowledge.

How do transformer language models memorize factual associations? A common view casts internal weight matrices as associative memories over pairs of embeddings, requiring parameter counts that scale linearly with the number of facts. We develop a theoretical and empirical account of an alternative, \emph{geometric} form of memorization in which learned embeddings encode relational structure directly, and the MLP plays a qualitatively different role. In a controlled setting where a single-layer transformer must memorize random bijections from subjects to a shared attribute set, we prove that a logarithmic embedding dimension suffices: subject embeddings encode \emph{linear superpositions} of their associated attribute vectors, and a small MLP acts as a relation-conditioned selector that extracts the relevant attribute via ReLU gating, and not as an associative key-value mapping. We extend these results to the multi-hop setting -- chains of relational queries such as ``Who is the mother of the wife of $x$?'' -- providing constructions with and without chain-of-thought that exhibit a provable capacity-depth tradeoff, complemented by a matching information-theoretic lower bound. Empirically, gradient descent discovers solutions with precisely the predicted structure. Once trained, the MLP transfers zero-shot to entirely new bijections when subject embeddings are appropriately re-initialized, revealing that it has learned a generic selection mechanism rather than memorized any particular set of facts.

Recent works have analyzed the impact of individual components of neural networks on gendered predictions, often with a focus on mitigating gender bias. However, mechanistic interpretations of gender tend to (i) focus on a very specific gender-related task, such as gendered pronoun prediction, or (ii) fail to distinguish between the production of factually gendered outputs (the correct assumption of gender given a word that carries gender as a semantic property) and gender biased outputs (based on a stereotype). To address these issues, we curate \gknow, a benchmark to assess gender knowledge and gender bias in language models across different types of gender-related predictions. \gknow allows us to identify and analyze circuits and individual neurons responsible for gendered predictions. We test the impact of neuron ablation on benchmarks for disentangling stereotypical and factual gender (DiFair and the test set of GKnow), as well as StereoSet. Results show that gender bias and factual gender are severely entangled on the level of both circuits and neurons, entailing that ablation is an unreliable debiasing method. Furthermore, we show that benchmarks for evaluating gender bias can hide the decrease in factual gender knowledge that accompanies neuron ablation. We curate GKnow as a contribution to the continuous development of robust gender bias benchmarks.

The literature on how large language models handle conflict between their training knowledge and a contradicting document presents a persistent empirical contradiction: some studies find models stubbornly retain their trained answers, ignoring provided documents nearly half the time, while others find models readily defer to the document, following context approximately 96% of the time. We argue these contradictions dissolve once one recognises that prior experiments have studied three qualitatively distinct processing situations without distinguishing them. We propose a three-regime framework: Regime 1 (single-source updating, dominant predictor: evidence coherence), Regime 2 (competitive integration, dominant predictor: parametric certainty), and Regime 3 (task-appropriate selection, dominant predictor: task knowledge requirement). We formalise a distinction between parametric strength (exposure frequency) and parametric uniqueness (encoding consistency), showing empirically that these are orthogonal dimensions (r = -0.002, p = .97) with strength as the operative predictor in stable factual domains. We validate the framework across Claude Sonnet 4.6, GPT-5.5, Gemini 2.5 Flash, Llama 4 Maverick, and DeepSeek V3 using 9,970 API calls in three experimental phases. GEE logistic regression confirms the predicted Regime 2 certainty gradient for all five models (beta = -0.38 to -0.50, all p <= .013, BH-FDR corrected). A Regime 3 ablation shows task framing alone flips context-following from near-100% (contextual knowledge condition) to 6-71% (parametric knowledge condition), with all five models significant (p < .001). The certainty gradient is robust to multinomial outcome modeling, sensitivity analyses for hedging responses, and FDR correction.

Large language models confidently produce outdated answers, and no existing method can detect them. We show this is not an engineering failure but a structural one: temporal drift, whether a stored fact has changed since training, is encoded as a direction in the residual stream geometrically orthogonal to both correctness and uncertainty. Any method operating on correctness or uncertainty signals is therefore blind to drift by construction. We verify this across six instruction-tuned models. A linear probe trained directly on drift labels achieves AUROC $0.83$--$0.95$; methods based on token entropy, semantic entropy, CCS, and SAPLMA all remain near chance ($0.49$--$0.57$). Five tests confirm the geometric orthogonality: weight cosines ($|\cos| \leq 0.14$), score correlations ($|r| \leq 0.20$), bidirectional null-space projection ($|Δ| \leq 0.008$), iterative null-space projection with $k{=}10$, and difference-of-means dissociation. Mechanistically, the MLP retrieval circuit produces identical dynamics for stale recall and confabulation ($r > 0.81$, six models), explaining why output confidence cannot separate them. A cross-cutoff experiment holds inputs constant and varies only the model: the probe fires on the model whose training predates the fact's transition and stays silent otherwise ($P(A{>}B) = 0.975$--$0.998$, twelve model pairs), confirming it reads model-internal knowledge state rather than input properties. Our code and datasets will be publicly released.

Reinforcement learning (RL) has achieved remarkable success in LLM reasoning, but whether it can also improve direct recall of parametric knowledge remains an open question. We study this question in a controlled zero-shot, one-hop, closed-book QA setting with no chain-of-thought, training only on binary correctness rewards and applying fact-level train-test deduplication to ensure gains reflect improved recall rather than reasoning or memorization. Across three model families and multiple factual QA benchmarks, RL yields ~27% average relative gains, surpassing both training- and inference-time baselines alike. Mechanistically, RL primarily redistributes probability mass over existing knowledge rather than acquiring new facts, moving correct answers from the low-probability tail into reliable greedy generations. Our data-attribution study reveals that the hardest examples are the most informative: those whose answers never appear in 128 pre-RL samples (only ~18% of training data) drive ~83% of the gain, since rare correct rollouts still emerge during training and get reinforced. Together, these findings broaden the role of RL beyond reasoning, repositioning it as a tool for unlocking rather than acquiring latent parametric knowledge.

Language models draw on two knowledge sources: facts baked into weights (parametric memory, PM) and information in context (working memory, WM). We study two mechanistically distinct failure modes--conflict, when PM and WM disagree and interfere; and hallucination, when the queried fact was never learned. Both produce confident output regardless, making output-based monitoring blind by design. We show both failures share a unified geometric account. In the hidden-state space of autoregressive generation, learned facts form attractor basins. Conflict is basin competition: WM disrupts convergence to the correct basin without raising output entropy. Hallucination is basin absence: the hidden state drifts freely when no memorized basin exists. The frozen LM head, designed for next-token prediction, cannot distinguish these cases and fires confidently either way. We verify this account in a controlled synthetic task-entity identifiers mapped to unique codes with PM installed via LoRA adapters--where ground truth is exact and component roles can be causally isolated through targeted adapter placement. Geometric margin--the hidden state's distance to the nearest memorized basin--reads this geometry directly and separates correct recall from hallucination far more cleanly than output entropy, with zero false refusals where entropy-based detection cannot avoid rejecting the vast majority of correct outputs. The separation holds on natural-language factual queries from the pretrained model with no adaptation, confirming attractor geometry is structural rather than a fine-tuning artifact. The fraction of confident hallucinations follows a scaling law $C = \exp(-c/\barΔ)$, growing with scale even as overall error rates fall. Hidden states reliably encode epistemic state; the frozen output head systematically erases it--and this erasure worsens with scale.

Scientific reasoning rarely stops at what is directly observable; it often requires uncovering hidden structure from data. From estimating reaction constants in chemistry to inferring demand elasticities in economics, this latent structure recovery is what distinguishes scientific reasoning from curve fitting. Large language models (LLMs) can often recall and apply relevant scientific formulas, but we show that this ability is surprisingly easy to suppress. We show that adding in-context examples makes models rely less on pretrained domain knowledge, even when those examples are generated by the very same formula. Rather than reinforcing knowledge-driven derivation, examples shift computation toward empirical pattern fitting. We document this knowledge displacement on 60 latent structure recovery tasks across five scientific domains, 6,000 trials, and four models. This displacement is consistent across domains, but its accuracy consequences depend on how the displaced strategy compares to the one that replaces it: the same shift can lower accuracy, leave it unchanged, or appear to improve it. In all cases, however, the model shifts away from knowledge-driven reasoning. For practitioners deploying LLMs on scientific tasks, the message is cautionary: in-context examples may displace, rather than reinforce, the knowledge they are intended to support.