Large Language Model Performance

Do neural models, such as Large Language Models, genuinely acquire compositional abilities for interpretation of natural language? When we talk about semantic interpretation, we can distinguish two complementary aspects: establishing what an expression refers to in the world (which we call the Extensional task) and representing its sense in a structured way (which we call the Intensional task). We evaluate LLMs and humans on both tasks in the setting of the Personal Relation Task (Paperno 2022) in which, given a universe of people and their relationships with each other, one is asked to interpret a noun phrase such as "Amber's parent's friend". Here, for the Intensional task, the answer is the formula "friend(parent(amber))", and for the Extensional task, the person. We find that humans and LLMs show opposite strengths: humans perform better on Extensional than Intensional tasks, and LLMs vice versa. Our methodology brings greater nuance to the understanding of compositional abilities in modern machine learning models. Our results support the notion that the lack of referential grounding in LLM training is a crucial missing component in mimicking human-like language understanding.

Zero Pronouns (ZPs) are a pervasive linguistic phenomenon in pro-drop languages such as Chinese and have long posed a challenge for natural language processing systems. Although Large Language Models (LLMs) perform well on many Chinese language tasks, their ability to process ZPs remains poorly understood. We conduct a systematic investigation of LLMs' handling of Chinese ZPs through a sequence of linguistically motivated tasks, including identification, referentiality classification, referential type classification, resolution, and translation. A diverse set of LLMs is evaluated across all tasks. Our results show that Chinese ZPs remain highly challenging for current LLMs, particularly for upstream tasks such as identification and referentiality classification. Performance on downstream tasks, such as ZP translation, is also consistently low: even state-of-the-art reasoning-oriented LLMs correctly translate fewer than half of Chinese ZPs into English.

Autoregressive language models frequently degrade during long-horizon generation, producing repetitive text, losing instruction adherence, and exhibiting unstable entropy. Despite the prevalence of these failures, practitioners lack online diagnostics to detect them in real-time as they occur. We formalize this degradation as cognitive fatigue, a measurable generation-time state characterized by decay in attention to the original prompt, representational drift, and entropy miscalibration. We introduce the Fatigue Index (FI), a lightweight, model-agnostic diagnostic that aggregates these three signals under explicit axioms (monotonicity, boundedness, interpretability) enabling reliable runtime monitoring. Across nine models (1B-13B parameters), FI trajectories exhibit structured temporal dynamics, predict task degradation (AUROC = 0.95) and repetition (Spearman rho = 0.94), and reveal non-monotonic scaling behavior: instruction-tuned models below 3B exhibit faster collapse than base models, with this trend reversing at 7B. Stress analyses further show that FI onset accelerates under longer contexts, middle-positioned evidence, and reduced numerical precision. These results establish cognitive fatigue as a coherent and measurable phenomenon, and position FI as a principled tool for runtime reliability monitoring in production LLM systems.

Large language models (LLMs) remain limited on tasks requiring indirect reasoning, cultural knowledge, and coordinated hypothesis testing. We investigate whether team-based interaction improves LLM performance in What? Where? When? (ChGK), a quiz game designed to reward collective reasoning. We introduce three team strategies: Voting, Silent Team (the captain observes final answers), and Talkative Team (the captain observes both answers and rationales). To minimize data leakage, we evaluate these strategies on a dataset consisting of 572 ChGK questions released in 2025. Using six recent large-scale open models, we show that team-based strategies outperform single-model baselines, yielding gains of up to 20 percentage points in accuracy. The best team achieves 44.23% accuracy, and approaches human team performance on questions with available human statistics. Analysis of inter-model diversity reveals that disagreement strongly predicts lower accuracy, but explanatory communication substantially mitigates performance drops. We further examine captain behavior and find no evidence of self-preference bias; access to peer rationales improves captain judgments. Overall, LLM teams function primarily as answer selection and error-filtering mechanisms rather than generators of novel solutions. Our findings highlight the importance of interaction and suggest adaptive strategies as a promising direction for multi-agent systems.

Large language models (LLMs) can autonomously conduct multi-stage cyber attacks, but the consistency of their offensive behavior under repeated trials remains unstudied. This work presents the first large-scale empirical measurement of LLM attack consistency: 400 autonomous penetration testing runs (4 models, 100 each) against an identical honeypot hosting OWASP Juice Shop and two additional vulnerable services, holding prompt, orchestrator, and target constant. No model emitted a content refusal that survived the orchestrator's one-shot authorization re-prompt at iterations 0-1. Claude Sonnet 4's API calls did encounter upstream service unavailability - 91 of 1,135 calls returned HTTP 529 overloaded_error during a documented Anthropic capacity event, truncating 39 of 100 Claude runs. An earlier draft catalogued these as safety refusals; on full-log audit they are upstream API failures, not model-level refusals. Despite this, Claude achieved full exploitation in 61 of 100 runs; Gemini 2.5 Flash-Lite in 85; GPT-4o-mini in 56 while deploying 98 unique attack strategies; qwen2.5-coder:14b in 25. Failure modes are model-distinctive: Claude through API truncation (39 runs), qwen through premature completion (52), GPT-4o-mini through iteration-budget exhaustion (23). Cross-service credential reuse appeared only in configurations retaining the most conversation history (qwen 57%, GPT-4o-mini 49%, cloud models 0% on 5-exchange windows). Cross-model exploitation rate differences are statistically significant (p < 0.001) with large effect sizes; qwen vs. Gemini SQL injection rates differ at Cohen's h = 1.12. First-exploit timing fell within a 15-30 second wall-clock range. To our knowledge, this is the first study to measure autonomous LLM attack behavior at N=100 per model across a multi-service target.

Patient-voiced clinical-triage benchmarks report high under-triage rates for consumer LLMs for constrained multiple-choice output, yet the same cases score differently with free-text. We ask whether output format changes the model's \emph{clinical representation} or only the mapping from a preserved representation to an answer. Using sparse-autoencoder (SAE) features in Gemma 3 4B/12B IT and Qwen3-8B, we find the same medical features fire on the shared clinical narrative under both formats but go {silent} at the multiple-choice decision token in all the cases at every model. Three independent methods (natural-language autoencoder verbalization, decision-token logit attribution, and top-feature characterization) agree that scaffold and format features, but not medical features, drive the decision logits. Behaviorally, the multiple-choice penalty inverts under both structured and natural-language input, option-order shuffle rules out positional bias, and the gap is dominated by off-by-one decision (the model picks an adjacent acuity letter to the gold answer) rather than knowledge failure. Thus, the failure originates in the output format and not in the clinical representation.

As LLMs become increasingly integrated into daily life, understanding how their presence will shape human linguistic behavior is an open question. We present a large-scale study of linguistic convergence in human-LLM dialogue, examining how humans and LLMs accommodate each other's linguistic style during multi-turn conversations. Using an asymmetric convergence metric on WildChat, a corpus of real-world ChatGPT transcripts, we find that while LLMs significantly overconverge toward their users on both function word and open-class features across eight languages, human convergence rates in this setting are broadly consistent with human-human baselines. These findings suggest that accommodation in human-LLM dialogue is asymmetric: while LLMs dramatically overfit to their users' style, humans linguistically accommodate LLMs no differently than they would another person.

Discourse particles, such as well and kind of, are crucial components that enable LLMs to "speak" more like humans. They are used to convey emotions, intentions, and interpersonal attitudes. However, existing studies have not yet built a comprehensive understanding of LLMs' capabilities in handling discourse particles. Moreover, the limited number of research focuses primarily on high-resource languages such as English, with little attention paid to Southeast Asian languages. In this paper, we (1) propose MalayPrag, a benchmark designed to systematically evaluate and analyze LLMs' capabilities in handling discourse particles in colloquial Malay; (2) introduce five attributes that provide a theoretically grounded, unified framework for interpreting pragmatic functions of discourse particles. Applying these two, we prompt ten off-the-shelf LLMs to perform three prediction tasks. The experimental results reveal substantial challenges for current LLMs to accurately connect discourse particles and their pragmatic functions in Malay. The provision of the five attributes designed in this study is found to significantly improve the connections, highlighting the need for structured scaffolding for models' pragmatic competence.

Memory is essential for enabling large language models to support long-horizon reasoning, yet existing memory systems remain unreliable and difficult to debug. Tracing memory's dynamic evolution is crucial to understand how information is synthesized, propagated, or corrupted over time. In this work, we study the new problem of error tracing and attribution in LLM memory systems. We propose a novel framework that transforms memory pipelines into executable memory evolution graphs, enabling fine-grained tracing of operational information flow. We then construct MemTraceBench, a benchmark collected from representative memory systems such as Long-Context, RAG, Mem0, and EverMemOS, to systematically study memory failure modes. We further introduce an automatic attribution method that iteratively traces operation subgraphs to pinpoint the root cause of any failed case. Our analysis reveals that memory failures are systematic, stemming from operation-level issues like information loss and retrieval misalignment. Crucially, we leverage these fine-grained attribution signals to guide downstream prompt optimization, establishing a closed-loop system that automatically corrects faults and boosts end-task performance by up to 7.62%. Code will be released at https://github.com/zjunlp/MemTrace.

Large Language Models (LLMs) often produce explicit reflective traces during complex reasoning, accompanied by anthropomorphic markers such as wait, hmm, and alternatively. Although these markers are commonly used as visible indicators of reflection, their mechanisms remain unclear, which leaves the risk of overthinking associated with redundant and repetitive reflection markers. In this work, we revisit anthropomorphic reflection markers, examining their necessity for reasoning and role in the reflection. We suppress these markers through prompt-level and token-level interventions, and analyze their effects on task performance across four benchmarks and two model scales. Our results show that anthropomorphic markers are not uniformly necessary for reasoning performance: suppressing them can preserve or improve performance in several settings, especially under larger sampling budgets. Meanwhile, marker suppression does not necessarily remove reflection behavior, as models can still perform marker-free verification. These suggest that anthropomorphic markers tend to be surface cues rather than reliable proxies for reflection itself, and motivate future research on reasoning mechanisms beyond explicit marker patterns.

Whether large language models (LLMs) construct internal spatial world models from pure-text descriptions remains contested, and whether such capabilities transfer across languages has not been systematically studied. We introduce MentalMap, a multilingual diagnostic benchmark with a six-level capability hierarchy (L0-L5) spanning atomic spatial facts to generative world-graph construction, together with four diagnostic axes probing frame of reference, reading-direction bias, reasoning-effort allocation, and hallucination. MentalMap is built from 100 ProcTHOR household scenes, covers eight typologically diverse languages plus a structured-text control, and contains 39 task families across 1,950 evaluation cells. Evaluating thirteen LLMs across scales and model families, we identify a universal L3 reasoning cliff: no model retains even half of its L0 performance on viewpoint reasoning once baseline atomic accuracy exceeds 40%. The cliff persists across languages, scales, and prompting strategies, while structured-output failures and reasoning patterns vary substantially across models. Human evaluation under the identical pure-text protocol reproduces the same failure pattern, suggesting that the bottleneck arises from text-only working memory constraints rather than being specific to current LLM architectures. Our findings reframe pure-text spatial reasoning as a multi-axis world-modeling problem and motivate multimodal and scratchpad-augmented reasoning as future directions.

In-context learning (ICL) enables multimodal large language models (MLLMs) to classify images from a few labelled examples. Yet, how these models use the provided context remains opaque. While Chain-of-Thought prompting is widely used, recent work argues that it may not reflect true internal computation. In this paper, we systematically evaluate the concept-based explainability of frozen MLLMs under few-shot ICL using five conditions of increasing formal rigour, ranging from baseline classification to Description Logics (DL) axiom generation. Evaluating four state-of-the-art MLLMs via an independent LLM-as-a-judge pipeline, we demonstrate that explaining is genuinely harder than predicting alone. Surprisingly, forcing models to generate formally structured, concept-based explanations degrades predictive accuracy monotonically (from 93.8% to 90.1%), contradicting the assumption that explicit reasoning universally aids performance. However, when models successfully articulate class-discriminative visual features, explanation quality strongly correlates with correct predictions. Our findings suggest that while MLLMs excel at visual classification, they lack the specific instruction-tuning required for formal, machine-verifiable explainability.

Large Language Models (LLMs) are being rapidly adopted in agricultural imaging applications, ranging from crop interpretation to synthetic field image generation. However, these models frequently exhibit hallucinations outputs that appear confident yet deviate from biological or environmental reality potentially leading to misinformed agronomic insights. This study investigates such hallucinations in two complementary directions: image-to-text, where LLMs interpret crop or field imagery to describe conditions such as biotic and abiotic stresses, and text-to-image, where models generate synthetic agricultural scenes based on descriptive prompts. We examine errors involving biological inconsistency, contextual inaccuracy, and agronomic implausibility, evaluating the outputs under domain-informed criteria across multiple imaging modalities. Our analysis identifies recurring hallucination patterns within both interpretive and generative tasks. In image interpretation, LLMs (e.g., Gemma, LLAVA, Qwen, and MiniCPM) achieved modest zero-shot accuracy (63 to 75 percent), whereas few-shot prompting improved performance up to 86.8 percent, exhibiting false detections and missed infections, indicating residual hallucination effects. In text-to-image tasks, advanced models such as GPT-5 and Gemini 2.5 Flash generate up to 91 percent biologically inconsistent scenes under relaxed prompt constraints, revealing fundamental weaknesses in current LLMs. This systematic assessment of visual reasoning and generation offers critical insights toward enhancing the reliability and trustworthiness of LLM-based agricultural imaging platforms.

Causal discovery is a cornerstone of scientific reasoning, yet whether large language models can perform it reliably remains an open question. Recent benchmarks show that even fine-tuned models plateau on simple causal graphs and degrade as complexity grows, but why they fail has not been established. We prove the failure is fundamental: supervised fine-tuning, direct preference optimization, and in-context learning all produce predictors that cannot distinguish between causal graphs generating similar observational data, and any attempt to do so requires the model's internal representations to grow unboundedly, violating the very conditions under which these methods work. We formalize this as a kernel obstruction theorem, establishing that the limitation is intrinsic to the learning paradigm, \emph{not any particular model or dataset}. We propose Agentic Causal Bayesian Optimization (A-CBO), wherein a frozen language model serves as an interventional oracle answering targeted queries about intervention effects, while an external Bayesian loop concentrates beliefs over candidate graphs in logarithmically many rounds. Because the decision operates outside the space where the obstruction applies, A-CBO provably converges while the underlying model remains unchanged. On Corr2Cause, A-CBO matches fine-tuned baselines without any training. On Extended Corr2Cause, a new benchmark scaling to 24 variables with 18K test samples, A-CBO significantly outperforms both fine-tuning and preference optimization, with the advantage growing

Counterfactual tuning (CFT) has emerged as a promising paradigm for Large Language Model (LLM) unlearning by training models to generate alternative fictitious knowledge in place of undesired content. However, in this work, we find that this paradigm still underperforms other paradigms in some aspects, and identify two previously overlooked pitfalls underlying this gap: (1) knowledge conflict, where mutual inconsistencies within counterfactual corpora induce conflicting gradients that disrupt parameter optimization, and (2) hallucination spillover, where fitting false targets instills a persistent fabrication bias, inflating hallucination rates on unrelated domains. To systematically diagnose these issues, we introduce RWKU+, an extended benchmark equipped with novel trade-off metrics and gradient-level diagnostic tools. Our work further discusses the limitations and overhead of the paradigm, aiming to provide insights and actionable guidance for more rigorous LLM unlearning research.

Large language model (LLM) agents increasingly rely on external memory systems to remain consistent across long-horizon interactions, but little empirical work has been done to understand the specific failure modes and design choices that these systems present. Existing benchmarks report aggregate question-answering accuracy and treat memory systems as black boxes, making it impossible to attribute an incorrect answer to a particular failure mode of the system. We introduce MemFail, a diagnostic benchmark that isolates the failure modes of modern LLM memory systems. We begin by formalizing memory systems as the composition of three canonical operations -- summarization, storage, and retrieval -- and identify the potential failure modes induced by each. Based on these hypothesized failure modes, we construct five datasets spanning four tasks, each adversarially designed to test a specific operation of a memory system. Using these datasets, we evaluate four state-of-the-art memory systems on MemFail and demonstrate how MemFail can be used to empirically understand the tradeoffs induced by differences in memory system architectures.

In many reasoning tasks, large language models (LLMs) rely on structured external knowledge, such as graphs and tables, which is typically linearized into sequential token representations. However, even when sufficient knowledge is available, LLMs can still produce hallucinated outputs, and the underlying mechanisms behind such failures remain poorly understood. We investigate these mechanisms and find that hallucinations arise from systematic internal dynamics rather than random noise. First, attention disproportionately concentrates toward shortcut-like structural cues rather than distributing across the full context. Second, feed-forward representations fail to ground the provided knowledge, causing the model to revert to parametric memory. Moreover, our results indicate that hallucination is consistently associated with failures in semantic grounding within feed-forward layers, while attention allocation exhibits greater task-dependent variability. Finally, we show that these mechanistic patterns generalize beyond single-hop graphs to multi-hop and tabular settings, enabling effective hallucination detection across structured knowledge formats.

Multimodal Large Language Models (MLLMs) excel at structural reasoning yet suffer from a sharp logical brittleness in structural consistency. We term this phenomenon Structural Cognitive Overload (SCO), a byproduct of the contention between deep reasoning and safety alignment. However, prior work has predominantly targeted typographic and pixel-level perturbations, leaving the study of SCO largely unexplored. To this end, we propose StructBreak, an automated end-to-end framework designed to quantify SCO. By leveraging StructBreak, we uncover a novel higher-order cognitive overload attack paradigm; notably, this attack operates under a practical black-box setting, requiring no internal model access. Consequently, we utilize this framework to establish a comprehensive benchmark spanning ten diverse threat scenarios. Empirical evaluations on six leading MLLMs reveal that SCO readily triggers toxic generation, yielding a 92% average ASR (up to 97% on Gemini 2.5). To elucidate the mechanism of SCO, we further conduct model-level interpretations spanning attention dynamics, latent space topology, and geometric analysis. Our findings reveal that StructBreak acts as a novel structural channel to circumvent safety filters. Furthermore, the limited efficacy of inherent safety mechanisms underscores that current alignment paradigms are insufficient for the era of complex multimodal reasoning.

While LLMs excel at reasoning over prompts using static pretrained knowledge, they struggle significantly with context learning-the ability to dynamically extract, internalize, and apply new knowledge from complex, task-specific contexts. Recent evaluations on the CL-Bench reveal a critical capability gap: frontier models solve only 17.2% of context-dependent tasks on average.

Evaluating whether large language models (LLMs) capture the structure of natural language beyond local fluency remains an open challenge. Existing evaluation methods, largely based on task performance or short-context behavior, provide limited insight into the long-range statistical organization of generated text. We propose a complementary evaluation framework based on repeated subsequences. By analyzing their distribution across scales and relating it to higher-order Rényi entropies, we probe how texts reuse previously established structure under finite-length conditions. Experiments on human-written texts and length-matched GPT-generated texts show that, while power-law models can describe restricted ranges of block length, the observed entropy growth is often equally or better characterized by logarithmic--power forms. Across datasets, natural language exhibits stable entropy-growth patterns over accessible ranges, with consistent average behavior despite variability across individual texts. In contrast, GPT-generated texts show systematic and statistically significant shifts in estimated exponents with model size. These results demonstrate that repeated-subsequence entropy provides a quantitative structural diagnostic that reveals systematic differences in long-range organization, distinguishing natural language from state-of-the-art LLM outputs beyond surface-level fluency.

Backtesting large language models (LLMs) on historical financial data is unreliable because pre-training cuts off after the events happened. An LLM trained in 2024 already "knows" which way 2018-2020 stocks moved. We name this failure parametric look-ahead bias and propose FinCAD, an inference-time adaptation of Context-Aware Decoding that suppresses an LLM's memory of historical outcomes without retraining. FinCAD pairs an adversarial bias-discovery pipeline that learns a model-specific memory-activating prior prompt with an entity- and date-adaptive rule that scales the CAD strength to per-(entity, date) memorisation, so the penalty fires on memorised in-sample dates and decays to zero out-of-sample. Across five 7-14B LLMs and five mega-cap equities, FinCAD cuts in-sample backtest returns by up to -67.1% on memorised dates while leaving 2025 out-of-sample returns within $8K and Sharpe within 0.10 of baseline, and preserves general-purpose reasoning within 1.7 pts. On an eleven-model leaderboard, it raises the in-sample / out-of-sample Spearman correlation from +0.779 to +0.846, recovering rankings that genuinely predict out-of-sample performance.

How do learners acquire knowledge of what is unacceptable without negative evidence? Construction Grammar proposes statistical preemption: exposure to a conventional form (e.g., "donated the books to the library") preempts structurally possible but unattested alternatives ("*donated the library the books"). We present a computational study that, for the first time, directly dissociates statistical preemption from the competing entrenchment hypothesis in large language models within a single converging design. Across four experiments spanning 120 English verb-construction pairings (dative, causative, locative), we show that (1) LLM surprisal patterns correlate strongly with human acceptability judgments ($r = 0.79$), validated against three independent behavioral datasets; (2) these patterns are driven by competing-form frequency rather than overall verb frequency, confirmed by non-circular partial correlations; (3) preemption sensitivity scales as a power law with model size; and (4) a controlled fine-tuning intervention causally demonstrates that manipulating competing-form frequencies shifts preemption behavior in the predicted direction, with reverse-direction controls ruling out frequency-sensitivity confounds. These results provide converging evidence that neural language models acquire negative linguistic knowledge through distributional competition, the core mechanism posited by Construction Grammar.

Although Large Language Models (LLMs) demonstrate strong capabilities across various tasks, they exhibit significant performance discrepancies across languages. While prompting LLMs in English typically yields the highest general performance, it often induces a Western-centric bias, hindering the model's ability to accurately reflect diverse cultural knowledge. We hypothesize that LLMs already possess rich cultural knowledge embedded within local-language representations, but fail to retrieve it when prompted in English. To bridge this cross-lingual knowledge gap, we propose a novel self-supervised framework. Our method leverages multilingual self-consistency to identify the most reliable cultural responses across languages, combined with a self-critique mechanism to transfer this knowledge to the weaker language. Evaluations on the BLEnD benchmark demonstrate that our approach significantly improves cultural alignment-boosting performance on English queries by an average of 5.03%-relying entirely on self-generated data. Ultimately, our work demonstrates that latent cultural knowledge can be successfully surfaced and propagated across languages, enabling more culturally equitable and consistent LLMs.

Hallucination is often viewed as a direct consequence of missing knowledge: a model answers incorrectly when the correct answer is absent from its generation-time distribution, and correctly when it is present. We test this assumption by introducing a semantic notion of answer availability that aggregates token-level variants expressing the same answer concept, and asks whether the correct concept is already available at the moment the model commits to an answer. Across Qwen and Llama models from 0.8B to 72B in both Instruct and Base variants, 16-47% of Instruct hallucinations occur with substantial probability mass already on the correct concept, and the rate rises monotonically with scale. Comparing such failures against correct generations with matched semantic support, the distinguishing factor is not whether the correct concept is represented, but how its probability is distributed: correct generations concentrate mass on a single surface form, hallucinations disperse it across alternatives. The same sharpening asymmetry extends across multi-token generation and is detectable in pre-generation hidden states. Together, these results identify a single mechanism: instruction tuning sharpens answer commitment with scale, making helpfulness and confident hallucination two consequences of the same underlying disposition.

When a multi-module LLM agent fails, the module most responsible for the failure is not necessarily the best place to intervene. We demonstrate this Diagnostic Paradox empirically: causal analysis consistently identifies the routing module -- which selects which tool to call next -- as the primary bottleneck across three independent agent families. Yet injecting prompt-level correction examples into this module consistently degrades performance, sometimes severely. Patching an upstream query-rewriting module instead reliably improves outcomes. The effect holds with statistical significance on two agent families and directional consistency on a third; alternative repair strategies at the routing module (instruction rewriting, model upgrade) are neutral, confirming that the harm is specific to correction-injection patching. We explain this asymmetry through the Linguistic Contract hypothesis: each downstream module implicitly adapts to its upstream's characteristic error distribution, so correcting the bottleneck breaks this implicit alignment in a way that upstream corrections do not. We operationalize this via a per-agent co-adaptation measure, derived from diagnosis alone, and show it is consistently associated with patching harm across agent families: higher co-adaptation co-occurs with harm, lower with safety. This trend holds across all three agent families, providing preliminary support for the hypothesis beyond a single-agent observation.

Large language models (LLMs) have demonstrated impressive reasoning abilities across a wide range of tasks, but data contamination undermines the objective evaluation of these capabilities. This problem is further exacerbated by malicious model publishers who use evasive, or indirect, contamination strategies, such as paraphrasing benchmark data to evade existing detection methods and artificially boost leaderboard performance. Current approaches struggle to reliably detect such stealthy contamination. In this work, we uncover a critical phenomenon: a model's generated reasoning steps actively mask its underlying memorization. Inspired by this, we propose the Zero-CoT Probe (ZCP), a novel black-box detection method that deliberately truncates the entire Chain-of-Thought (CoT) process to expose latent shortcut mappings. To further isolate memorization from the model's intrinsic problem-solving capabilities, ZCP compares the model's zero-CoT performance on the original benchmark against an isomorphically perturbed reference dataset. Furthermore, we introduce Contamination Confidence, a metric that quantifies both the likelihood and severity of contamination, moving beyond simple binary classifications. Extensive experiments on both previously identified contaminated models and specially fine-tuned contaminated models demonstrate that ZCP robustly detects both direct and evasive data contamination. The code for ZCP is accessible at https://github.com/Yifan-Lan/zero-cot-probe.

Many safety and alignment failures of large language models (LLMs) occur due to out-of-distribution (OOD) situations: unusual prompt or response patterns that are unforeseen by model developers. We systematically study whether LLM monitoring pipelines can detect these OOD alignment failures by introducing a benchmark called Misalignment Out Of Distribution (MOOD). It is difficult to find failures that are truly OOD for off-the-shelf models trained on vast safety datasets. We sidestep this by including a restricted training set in MOOD that we use to train our own monitors, as well as seven test sets with diverse alignment failures that are outside the training distribution. Using MOOD, we find that guard models (safety classifiers) often fail to generalize OOD. To fix this, we propose combining guard models with OOD detectors. We test four types of OOD detectors and find that a combination of a guard model with Mahalanobis distance and perplexity-based OOD detectors can improve recall from 39% to 45%. We also establish positive scaling trends across model scales for monitors that combine a guard model and OOD detector; we find that incorporating OOD detection into monitoring achieves a higher recall gain than using a guard model with 20 times more parameters. Our work suggests that OOD detection should be a crucial component of LLM monitoring and provides a foundation for further work on this important problem. We release the code and data for our experiments publicly, and you can find the relevant links here: https://github.com/Dylan102938/mood-bench.

Humans effortlessly go beyond literal meanings: If you mow the lawn, I will give you fifty dollars, is typically understood as implying that the speaker will pay only if the lawn is mowed, whereas If you are hungry, there is pizza in the oven implies that pizza is available regardless of the hearers hunger. Large Language Models - LLMs - show human-like performance on many tasks, yet it remains unclear whether they reason like humans. To address this, we conducted a population-matching experiment assessing how twentyfive LLMs compute conditional inferences across four languages, compared to an equal number of humans per language. We find that humans enrich logical reasoning through pragmatic inferences across languages. Model behavior is more variable. Some LLMs perfectly follow the truth-table of conditionals but they ignore pragmatic inferences, while others deviate from the truth-table, adhering to a single interpretation across the board, thus reflecting accurate rule-based processing but not human-like reasoning. Overall, LLMs are accurate semantic operators, but fail to capture the pragmatic enrichments characteristic of human reasoning. Crucially, LLM accuracy is neither predicted nor boosted by open vs. closed status, training orientation, or architecture type, suggesting that pragmatic reasoning is still an emerging ability in the cognitive toolkit of artificial systems.

Large language models (LLMs) are increasingly deployed in socially sensitive settings despite substantial documentation that they encode gender biases. Chain-of-Thought (CoT) prompting has been proposed as a bias-mitigation approach. However, existing evaluations primarily focus on changes in LLM benchmark performance, providing limited insight into whether apparent bias reductions reflect meaningful changes in a model's internal mechanisms. In this work, we investigate how CoT prompting affects gender bias in LLMs, combining benchmark-based evaluation with mechanistic interpretability techniques and reasoning chain failure analysis. Our results confirm a stereotypical bias present in LLM outputs across benchmarks, showing that CoT prompting does not consistently reduce the bias gap. Mechanistic analyses reveal that although CoT balances biased behavior in certain attention head clusters, gender bias remains embedded in hidden representations, indicating only superficial mitigation. Inspection of reasoning chains further suggests that these improvements stem from memorization and familiarity with the dataset rather than genuine understanding of bias.

Large Language Models are increasingly proposed as cognitive components for robotic systems, yet their opaque decision processes make it difficult to explain success or failure in closed-loop embodied tasks. Following an empirical AI methodology, we study embodied LLM agents behaviorally by varying the information available to the agent and measuring the resulting changes in behavior. Using the Lockbox, a sequential mechanical puzzle with hidden interdependencies, we evaluate LLMs across RGB, RGB-D, and ground-truth symbolic observations in a physical robotic setup and use controlled simulation to probe the resulting behavior. Counterintuitively, agents perform best under raw RGB input and worst under perfect ground-truth observations. In simulation, we probe this effect by randomly flipping perceived action outcomes and find that moderate noise improves performance, peaking at a 40% flip probability with a 2.85-fold success rate increase over the noise-free baseline. Further analysis links this gain to a reduction in repetitive action loops. These findings suggest that success rates alone are insufficient for evaluating LLMs, as measured performance may reflect the interaction between perceptual errors and reasoning failures rather than robust problem solving.

Automatic speech recognition (ASR) has improved substantially in recent years, yet performance remains limited for low-resource languages. Large language models (LLMs) have shown promise for improving ASR through generative error correction (GER), but their effectiveness in low-resource settings remains underexplored. In addition, it remains unclear to what extent data contamination influences the reported improvements in LLM-based GER. This study investigates LLM-based GER for low-resource Frisian. In addition to a public corpus, we construct and use a Frisian offline dataset with non-public texts for evaluation to control for potential data contamination. Results show that GER improves ASR performance in most settings, with the best GPT-5.1 results surpassing oracle WERs. Comparable gains on the offline dataset indicate that improvements reflect true correction ability. We further provide a detailed error analysis revealing model correction patterns.

Large language models (LLMs) have achieved remarkable success in complex reasoning tasks via long chain-of-thought (CoT), yet their immense computational overhead hinders real-world deployment. LLM reasoning distillation addresses this by transferring reasoning capabilities from formidable teacher models to compact student models. However, existing distillation paradigms face a fundamental dilemma. Typical off-policy distillation strictly utilizes teacher-generated golden trajectories, suffering from an exposure bias due to the mismatch between training distributions and student-generated inference contexts, which leads to error cascades in long CoT reasoning. To address this, on-policy distillation allows students to explore their own trajectories, but we demonstrate that it inherently introduces a reciprocal reversed exposure bias: the teacher model also struggles to provide positive guidance when conditioned on student-generated sub-optimal contexts. To resolve this dual exposure biases problem, we propose Monitoring Trajectories and Backtracking when it strays (MOTAB), a new LLM reasoning distillation pipeline. Specifically, MOTAB dynamically monitors the student's on-policy generation against an adaptive safety boundary. When the generation strays and exceeds this threshold, MOTAB backtracks to the last safe state and leverages teacher intervention to correct the course. This approach inherently tolerates minor student errors to mitigate exposure bias, while preventing sub-optimal contexts to circumvent reversed exposure bias. Extensive experiments on the LIMO-v2 and AceReason datasets demonstrate that MOTAB effectively alleviates the dual exposure biases, yielding a roughly 3% average performance improvement in reasoning tasks.

Large language models (LLMs) are increasingly explored for clinical decision support, yet most evaluations are conducted in English, leaving their reliability in other languages uncertain. Here we evaluate the impact of prompting language on diagnostic reasoning and final diagnosis accuracy by comparing English and French performance across five LLMs (o3, DeepSeek-R1, GPT-4-Turbo, Llama-3.1-405B-Instruct, and BioMistral-7B). A total of 180 clinical vignettes covering 16 medical specialties were assessed by two physicians using an 18-point scale evaluating both diagnosis accuracy and reasoning quality. Four of the five models performed better in English (mean difference 0.37-0.91, adjusted p < 0.05), with the gap spanning multiple aspects of reasoning, including differential diagnosis, logical structure, and internal validity. o3 was the only model showing no overall language effect. These findings demonstrate that prompting language remains a critical determinant of LLM clinical performance, with implications for equitable linguistico-cultural deployment worldwide.

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.

Large Language Models (LLMs) are increasingly used in educational settings as interactive tools for collaboration. However, their tendency toward sycophancy, aligning with user beliefs even when incorrect, raises concerns for learning and decision-making, especially for less knowledgeable users. This study investigates how sycophantic alignment emerges in authentic multi-turn human-AI interactions and whether interventions targeting increasing AI literacy and prompting competencies can mitigate its effects. In a controlled mixed-design experiment, 60 participants completed analytical survival ranking tasks by first generating individual rankings and then making final decisions after collaborating with an AI assistant, both before and after receiving either general or sycophancy-focused prompting training. Preliminary results show that LLMs are highly sensitive to user input: lower-quality initial responses lead to poorer AI advice, suggesting that the model mirrors or incorporates user reasoning rather than correcting it or offering better alternatives that are missing or less frequent in the conversation. Critically, the propagation of user errors into AI responses significantly reduced both the quality of AI feedback and final user task performance, revealing a form of contextual sycophantic dependence. While the intervention did not eliminate the propagation of contextual errors, it significantly improved AI advice by reducing the direct mirroring of incorrect user rankings. These findings suggest that prompting and AI literacy alone may be insufficient to ensure epistemically independent AI support, highlighting the need for system-level approaches that better promote critical engagement in human-AI collaboration.

Large Language Models (LLMs) remain vulnerable to jailbreak attempts that circumvent safety guardrails. We investigate whether multi-turn conversations using low-resource African languages (Afrikaans, Kiswahili, isiXhosa, and isiZulu) can bypass safety mechanisms across commercial LLMs. We translated prompts from existing datasets and evaluated ChatGPT, Claude, DeepSeek, Gemini, and Grok through automated testing and human red-teaming with native speakers. Single-turn translation attacks proved ineffective, while multi-turn conversations achieved English harmful response rates from 52.7% (Claude 3.5 Haiku) to 83.6% (GPT-4o-mini), Afrikaans from 60.0% (Claude 3.5 Haiku) to 78.2% (GPT-4o-mini), and Kiswahili from 41.8% (Claude 3.5 Haiku) to 70.9% (DeepSeek). Human red-teaming increased jailbreak rates compared to automated methods. Over all evaluated languages, the average jailbreak rate increased from 59.8% to 75.8%, with improvements of +20.0% (Afrikaans), +12.7% (isiZulu), +12.3% (isiXhosa), and +1% (Kiswahili), demonstrating that poor translation quality limits jailbreak success. These findings suggest that vulnerabilities in LLMs persist in multilingual contexts and that translation quality is the critical factor determining jailbreak success in low-resource languages.

Large Language Model (LLM) cascade systems are designed to balance efficiency and performance by processing queries with lightweight models while selectively escalating complex cases to more powerful ones. Such systems seek to reduces computational cost and latency while maintaining task performance, making it an appealing choice for large-scale deployment. However, the cascade design introduces new vulnerabilities through an expanded attack surface: the inclusion of lightweight front-end models and internal decision mechanisms introduces new weaknesses. In this work, we present the first study demonstrating that LLM cascade systems are susceptible to targeted adversarial manipulation, which disrupts both performance objectives and the intended cost advantages of the cascade design. We propose a novel attack framework that employs constrained sequential collaborative optimization of adversarial suffix under cascade dependencies, enabling simultaneous exploitation of lightweight models and decision mechanisms. This framework adapts to adversaries with varying capabilities, inducing controllable degradation in both cost-efficiency and accuracy. Unlike prior attacks targeting standalone models, our approach strategically leverages the cascade structure to achieve significantly stronger impact. Extensive experiments across diverse datasets and representative LLM cascade systems validate the practicality and severity of this attack. Our findings highlight the urgent need to rigorously scrutinize the security of LLM cascade systems and call for broader attention to the systemic risks inherent in such designs.

Large Language Models (LLMs) are increasingly deployed in high-stakes domains such as clinical decision support and medical documentation. However, the robustness of these models against subtle linguistic variations, specifically stigmatizing language (SL) commonly found in human-authored clinical notes, remains critically under-explored. In this work, we investigate whether frontier LLMs inherit and propagate this human bias when processing clinical text. We systematically evaluate nine frontier LLMs across four stigmatized medical conditions, utilizing clinical vignettes injected with varying intensities and phenotypes of SL (doubt, blame, and maligning). Our results demonstrate that all evaluated models exhibit substantial bias, with clinical decision-making significantly skewed towards less aggressive patient management. Notably, we observe a high sensitivity to linguistic framing, where a single SL sentence is sufficient to alter model outputs, revealing a clear dose-response relationship. Furthermore, we evaluate standard prompt-based mitigation strategies, including Chain-of-Thought (CoT) reasoning and model self-debiasing. These approaches show limited efficacy; models struggle to explicitly identify SL while remaining implicitly influenced by it. Our findings expose a critical vulnerability in current LLMs regarding fairness and robustness in clinical NLP, underscoring the need for rigorous algorithmic guardrails to prevent the automation of health disparities.

Deploying large-scale LLM training and inference with optimal performance is exceptionally challenging due to a complex design space of parallelism strategies, system optimizations, and hardware configurations. Accurate and rapid performance simulation is critical for guiding optimization efforts and system studies by validating "what-if" Hooker Figure hypotheses. To address this, we introduce Charon, a unified, modular, and fine-grained simulator for accurately predicting LLM performance. Experiments show Charon achieves high accuracy across different models and configurations, with an overall prediction error consistently under 5.35%, and even under 3.74% for training with a large-scale GPU cluster. In a practical inference deployment case, Charon discovered a configuration that improved system throughput over an engineering-tuned baseline, demonstrating its significant real-world value.

Improving LLM performance on downstream tasks sometimes requires leveraging auxiliary datasets during post-training. In practice, however, developers face constraints on compute, labeling, and licensing costs that preclude using all available data, necessitating principled dataset-level selection. These constraints are increasingly shaped by dataset marketplaces, where data acquisition is governed by budgets and negotiation. We study dataset valuation as a subset selection problem during LLM post-training. Our goal is to identify and weight auxiliary datasets so as to maximize target task performance given constrained budgets. We first show that commonly used gradient alignment scores provide a reasonable yet incomplete valuation signal, as they ignore redundancy among datasets. To address this, we propose a scalable convex dataset-level valuation method based on kernel mean matching (KMM) in gradient space, which jointly accounts for alignment with the target task and redundancy across auxiliary datasets. Through extensive experiments across diverse post-training settings and tasks, we show that our approach consistently outperforms existing valuation baselines, achieving stronger performance with low computational overhead. Our results position dataset valuation as a practical decision tool for post-training data selection in market-constrained large language model settings. The code is available at https://github.com/uiuctml/convex_data_valuation.

Effective tutoring requires distinguishing optimal, valid but suboptimal, and incorrect student solutions, a distinction central to intelligent tutoring systems (ITS) but untested for LLM-based tutors. As LLMs are increasingly explored as conversational complements to ITS, evaluating their diagnostic precision is essential. We present a benchmark of seven LLM feedback agents in propositional logic using knowledge-graph-derived ground truth across 10,836 solution--feedback pairs and three feedback conditions. Models achieved near-ceiling performance on optimal steps but systematically over-rejected valid but suboptimal reasoning and over-validated incorrect solutions, precisely where adaptive tutoring matters most. These failures persisted across models regardless of solution context, suggesting architectural rather than informational limits. Moreover, accurate diagnosis did not reliably produce pedagogically actionable feedback, revealing a gap between diagnostic judgment and instructional effectiveness. Our findings suggest that LLMs are better suited for hybrid architectures where KG-grounded models handle diagnosis while LLMs support open-ended scaffolding and dialogue.

Linguistic ambiguity is critical to the robustness of Large Language Models (LLMs), yet existing research focuses mostly on English, with limited attention devoted to Chinese. Existing Chinese ambiguity datasets (e.g., CHAmbi) suffer from poor scalability. Guided by Potential Ambiguity (PA) Theory, we design a semi-automatic pipeline to construct CHA-Gen. It is the first PA Theory-grounded Chinese ambiguity dataset, which comprises 5,712 sentences (2,414 ambiguous, 3,298 unambiguous) across 18 potential ambiguous structures. Evaluating LLMs (e.g. Gemma 3, Qwen 2.5/3 series) via direct querying and machine translation, we find that LLMs struggle with ambiguity detection (improved by CoT prompting). Analysis of Qwen3-32B's CoT rationales reveals three common failure modes: ambiguity blindness, misattribution, and premature resolution. Uncertainty quantification with semantic entropy metric shows higher uncertainty for ambiguous sentences. Moreover, instruction tuning induces overconfidence, whereas Base models better capture semantic diversity. We further observe that models exhibit a bias toward dominant interpretations. Our work provides a scalable approach for Chinese ambiguity corpus and insights into LLMs' ambiguity handling, laying a foundation for enhancing Chinese ambiguity research in LLMs.

Premature closure, or committing to a conclusion before sufficient information is available, is a recognized contributor to diagnostic error but remains underexamined in large language models (LLMs). We define LLM premature closure as inappropriate commitment under uncertainty: providing an answer, recommendation, or clinical guidance when the safer response would be clarification, abstention, escalation, or refusal. We evaluated five frontier LLMs across structured and open-ended medical tasks. In MedQA (n = 500) and AfriMed-QA (n = 490) questions where the correct choice had been removed, models still selected an answer at high rates, with baseline false-action rates of 55-81% and 53-82%, respectively. In open-ended evaluation, models gave inappropriate answers on an average of 30% of 861 HealthBench questions and 78% of 191 physician-authored adversarial queries. Safety-oriented prompting reduced premature closure across models, but residual failure persisted, highlighting the need to evaluate whether medical LLMs know when not to answer.

Modern LLM RL systems separate rollout generation from policy optimization. These two stages are expected to produce token probabilities that match exactly. However, implementation differences can make them assign different values to the same sequence under the same model weights, inducing Training-Inference Mismatch (TIM). TIM is difficult to inspect because it is entangled with off-policy drift and common stabilization mechanisms. In this work, we isolate TIM in a zero-mismatch diagnostic setting (VeXact), and show that small token-level numerical disagreements can independently cause training collapse. We further show that TIM changes the effective optimization problem, and identify a set of remedies that could mitigate TIM. Our results suggest that TIM is not benign numerical noise, but a systems-level perturbation that should be treated as a first-order factor in analyzing LLM RL stability.

Large language models (LLMs) often expose useful signals of self-monitoring: before solving a problem, they can estimate whether they are likely to succeed, and after solving it, they can judge whether their answer is likely to be correct. However, these signals are typically measured or elicited in isolation, rather than used to control inference. In this work, we ask whether LLMs possess latent metacognitive ability that can be turned into effective test-time control. Inspired by the Nelson--Narens theory from cognitive psychology, we propose a metacognitive harness that separates monitoring from reasoning. For each problem, the model first reports a pre-solve feeling-of-knowing (FOK) signal; after each solve attempt, it reports a post-solve judgment-of-learning (JOL) signal. Rather than treating these signals as passive confidence estimates, the harness turns them into an explicit control interface for reasoning: it decides when to trust the current solution, when to retry with compact metacognitive feedback, and when to pass multiple attempts to a final aggregator. Across text, code, and multimodal reasoning benchmarks, our harness substantially improves a fixed Claude Sonnet-4.6 base model without parameter updates or benchmark-specific fine-tuning. On the evaluated public benchmark snapshots, it raises pooled accuracy from 48.3 to 56.9 and exceeds the strongest listed leaderboard entries on the three primary evaluation settings: HLE-Verified, LiveCodeBench v6, and R-Bench-V. These results suggest that strong LLMs may already possess useful metacognitive ability, but require an explicit control harness to act on it during reasoning.

Large language models (LLMs) have revolutionized natural language processing. Understanding their internal mechanisms is crucial for developing more interpretable and optimized architectures. Mechanistic interpretability has led to the development of various methods for assessing layer relevance, with cosine similarity being a widely used tool in the field. On this work, we demonstrate that cosine similarity is a poor proxy for the actual performance degradation caused by layer removal. Our theoretical analysis shows that a layer can exhibit an arbitrarily low cosine similarity score while still being crucial to the model's performance. On the other hand, empirical evidence from a range of LLMs confirms that the correlation between cosine similarity and actual performance degradation is often weak or moderate, leading to misleading interpretations of a transformer's internal mechanisms. We propose a more robust metric for assessing layer relevance: the actual drop in model accuracy resulting from the removal of a layer. Even though it is a computationally costly metric, this approach offers a more accurate picture of layer importance, allowing for more informed pruning strategies and lightweight models. Our findings have significant implications for the development of interpretable LLMs and highlight the need to move beyond cosine similarity in assessing layer relevance.

Large language models achieve strong performance in language generation and knowledge-intensive tasks, yet remain limited in settings requiring causal reasoning, persistent state tracking, and long-horizon planning. We argue that these limitations may arise from an objective-level mismatch between sequence prediction and reasoning over latent environment dynamics. To formalize this distinction, we introduce Latent Dynamics Inference (LDI), a conceptual perspective that interprets language and multimodal observations as partial evidence of underlying transition dynamics. To empirically investigate this perspective, we introduce Flux, a sequential reasoning environment specified entirely through natural-language rules. As a proof-of-concept case study, the rules are first compiled into an explicit state-transition simulator, illustrating that structured latent transition dynamics can, in some cases, be operationally extracted from textual rule descriptions. This enables a controlled comparison between the LLMs operating purely over textual observations and reinforcement-learning agents trained directly within the extracted latent state space. Within this case study, agents operating with explicit access to the latent state space exhibit substantially more stable behavior in long-horizon gameplay, achieving an aggregate win rate of approximately 79% versus 11% for LLMs. Qualitative analysis further reveals failure modes consistent with unstable persistent state tracking, including invalid actions, state-tracking errors, and short-horizon reasoning failures. The complete implementation of the Flux environment available at https://github.com/FeisalAlaswad/FLUX-RL-Agent Within the evaluated setting, these results suggest that strong sequence prediction alone may struggle to support robust long-horizon dynamic reasoning without mechanisms for persistent state tracking and transition modeling

Large language models can follow complex instructions in a single turn, yet over long multi-turn interactions they often lose the thread of instructions, persona, and rules. This degradation has been measured behaviorally but not mechanistically explained. We propose a channel-transition account: goal-defining tokens become less accessible through attention, while goal-related information may persist in residual representations. We introduce the Goal Accessibility Ratio (GAR), measuring attention from generated tokens to task-defining goal tokens, and combine it with sliding-window ablations and residual-stream probes. When attention to instructions closes, what survives reveals architecture. Across architectures, the transition yields qualitatively distinct failure modes: some models preserve goal-conditioned behavior at vanishing attention, others fail despite decodable residual goal information, and the layer at which this encoding emerges varies from 2 to 27. A within-model causal ablation that force-closes the attention channel in Mistral collapses recall from near-perfect to 11% on a 20-fact retention task and raises persona-constraint violations above an adversarial-pressure baseline without user pressure, with both effects emerging at the predictable crossover turn. Linear probes recover per-episode recall outcomes from residual representations with AUC up to 0.99 across all four primary architectures, while input embeddings remain at chance. Across architectures and model scales, the gap between attention loss and residual decodability predicts whether goal-conditioned behavior survives channel closure. We contribute GAR as a diagnostic, the channel-transition framework as a controlled mechanistic account, and a parametric prediction of failure timing under windowed attention closure.

While Large Language Models (LLMs) are widely used in open-domain Question Answering (QA), their ability to handle inferential questions-where answers must be derived rather than directly retrieved-remains still underexplored. This study investigates how the structure and quality of passages influence LLM performance on such questions. We focus on convergence, a measure of how effectively sentences (hints) eliminate incorrect answers, as a criterion for constructing passages. Using subsets of the TriviaHG dataset, we form passages by combining sentences with varying convergence levels and evaluate six LLMs of different sizes and architectures. Our results show that passages built from higher convergence sentences lead to substantially better answer accuracy than those selected by cosine similarity, indicating that convergence captures meaningful relevance for inferential reasoning. Additionally, ordering sentences by descending convergence slightly improves performance, suggesting that LLMs tend to prioritize earlier, information-rich cues. These findings highlight convergence as a practical signal for guiding passage construction and analyzing inferential reasoning behavior in LLMs.

This paper experimentally analyzes how the level of harness engineering affects the operational performance of small language models (SLMs, 2-3B parameters). Three harness conditions - model-only (raw prompt), minimal-shell (wrapper tags), and a 4-stage pipeline (plan->execute->verify->recover) - are applied to three models (Gemma4 E2B, Qwen3.5:2B, LLaMA 3.2 3B) across 24 tasks, comparing Task Success Rate (TSR) and Valid TSR (VTSR). The pipeline harness achieves TSR=0.952 and VTSR=1.000 on Gemma4 E2B (T1-T5, 21 tasks). A non-monotonic phenomenon - minimal-shell TSR < model-only TSR - is observed in two models. In LLaMA 3.2 3B model-only, seven format violations yield TSR=0.429, revealing scaffold collapse: the model abandons JSON structure under complex format requirements without harness support. Ablation shows planning and recovery each contribute approximately 24.7% of total gain. VCR (Verification Catch Rate)=0.625 across all pipeline runs.

Large Language Models (LLMs) achieve strong linguistic performance, yet their internal mechanisms for producing these predictions remain unclear. We investigate the hypothesis that LLMs encode representations of linguistic constraint violations within their parameters, which are selectively activated when processing ungrammatical sentences. To test this, we use sparse autoencoders to decompose polysemantic activations into sparse, monosemantic features and recover candidates for violation-related features. We introduce a sensitivity score for identifying features that are preferentially activated on constraint-violated versus well-formed inputs, enabling unsupervised detection of potential violation-specific features. We further propose a conjunctive falsification framework with three criteria evaluated jointly. Overall, the results are negative in two respects: (1) the falsification criteria are not jointly satisfied across linguistic phenomena, and (2) no features are consistently shared across all categories. While some phenomena show partial evidence of selective causal structure, the overall pattern provides limited support for a unified set of grammatical violation detectors in current LMs.

The CAP theorem states that a distributed system cannot simultaneously guarantee consistency, availability, and partition tolerance under network partition. Inspired by this result, this paper formulates a CAP-like conjecture for Large Language Models (LLMs). The proposed trilemma states that, under semantic underdetermination, an LLM cannot always simultaneously guarantee strong correctness, strict non-bias, and high utility. A prompt is semantically underdetermined when the given premises do not determine a unique answer. In such cases, a useful and decisive response requires the model to introduce a selection criterion, preference, prior, or value ordering. If this criterion is not supplied by the user or justified by the available premises, the response becomes biased in a broad selection-theoretic sense. Conversely, if the model avoids unsupported preferences, it may preserve correctness and non-bias but may reduce utility through refusal, hedging, or clarification. The paper formalizes this correctness--non-bias--utility trilemma, develops examples, and argues that certain LLM failures arise not merely from model limitations but from the structure of underdetermined decision requests.

We present Seirênes, a self-play RL framework that transforms contextual interference from a failure mode of LLM reasoning into an internal training signal for co-evolving more resilient reasoners. While RL with verifiable rewards has significantly advanced reasoning capabilities, models can still exhibit fragility when encountering non-idealized contexts: scenarios characterized by superfluous information, tangential instructions, or incidental correlations that differ from the clean distributions typical of standard benchmarks. Seirênes harnesses this vulnerability through a parameter-shared and adversarial self-play loop. Within this framework, a single model is trained to both construct plausible yet distracting contexts that expose its own reasoning blind spots, and solve problems by discerning the essential task from these perturbations to recover the core underlying logic. By pitting these competing objectives against each other, Seirênes compels the model to move beyond superficial pattern matching and anchors its capabilities in robust underlying reasoning. This continuous interaction sustains an informative co-evolutionary curriculum as the model improves. Across seven mathematical reasoning benchmarks and model scales from 4B to 30B, Seirênes achieves average gains of +10.2, +9.1, and +7.2 points. Besides, distracting contexts produced by the 4B Seirênes model reduce the accuracy of top-tier closed-source models (GPT and Gemini) by roughly 4--5 points, revealing Seirênes' general ability to uncover reasoning models' blind spots.

Fraud detection and anti-money-laundering (AML) compliance are high-value domains for large language models (LLMs), but their serving requirements differ sharply from generic chat workloads. Compliance prompts are often prefix-heavy, schema-constrained, and evidence-rich, combining reusable policy instructions, risk taxonomies, transaction or document context, and short structured outputs such as JSON labels or risk factors. These properties make prefix reuse, KV-cache efficiency, runtime tuning, model orchestration, and output validation first-order systems concerns. This paper introduces a workload-aware LLMOps stack for fraud and AML workloads using self-hosted open-weight models such as Meta Llama and Alibaba Qwen. The stack combines vLLM-style runtime tuning, PagedAttention, Automatic Prefix Caching, multi-adapter serving, adapter and prompt-length-aware batching, sleep/wake lifecycle management, speculative decoding, and optional prefill/decode disaggregation. To avoid exposing institution-specific data, the reproducibility track converts public synthetic AML datasets, including IBM AML and SAML-D, into prefix-heavy compliance prompts with reusable policy text, transaction evidence, typology definitions, and schema-constrained outputs. We also incorporate an LLM-as-judge quality gate using deterministic compliance checks, reference metrics, expert-adjudicated calibration data where available, and multi-judge rubric scoring. Across public-synthetic AML workloads and controlled serving benchmarks, workload-aware tuning improved throughput from 612-650 to 3,600 requests/hour, reduced P99 latency from 31-38 seconds to 6.4-8.7 seconds, and increased GPU utilization from 12% to 78%. These results show that regulated LLM performance is a workload-design, serving-optimization, and quality-gating problem, not only a model-selection problem.

As large language models are increasingly deployed in retrieval-augmented generation and agentic systems that accumulate extensive context, understanding how distracting information affects long-context performance becomes critical. Prior work shows that semantically relevant yet misleading documents degrade performance, but the quantitative relationship between the proportion of distractors and performance remains unstudied. In this work, we systematically vary the hard-distractor proportion in fixed-length contexts, revealing a striking nonlinear pattern: as the proportion of hard distractors increases, performance drops sharply within the first small fraction, while the remainder of the range yields only marginal additional decline. We term this ''The First Drop of Ink'' effect, analogous to how a single drop of ink contaminates water. Our theoretical and empirical analyses grounded in attention mechanics show that hard distractors capture disproportionate attention even at small proportions, with diminishing marginal impact as their proportion grows. Controlled experiments further show that filtering gains mainly come from context-length reduction rather than distractor removal; substantial recovery requires reducing the hard-distractor proportion to near zero, highlighting the importance of upstream retrieval precision.

Fine-tuning Large Language Models (LLMs) on benign narrow data can sometimes induce broad harmful behaviors, a vulnerability termed emergent misalignment (EM). While prior work links these failures to specific directions in the activation space, their relationship to the model's broader persona remains unexplored. We map the latent personality space of LLMs through established psychometric profiles like the Big Five, Dark Triad, and LLM-specific behaviors (e.g. evil, sycophancy), and show that the semantic geometry is highly stable across aligned models and their corrupted fine-tunes. Through causal interventions, we find that directions isolating social valence, such as the 'Evil' persona vector, and a Semantic Valence Vector (SVV) that we introduce, function as intrinsic guardrails: ablating them drives the misalignment rates above $40$%, while amplifying them suppresses the failure mode to less than $3$%. Leveraging the structural stability of the personality space, we show that vectors extracted $\textit{a priori}$ from an instruct-tuned model transfer zero-shot to successfully regulate EM in corrupted fine-tunes. Overall, our findings suggest that harmful fine-tuning does not overwrite a model's internal representation of personality, allowing conserved representations to serve as robust, cross-distribution guardrails.

Enterprise workloads are dominated by deterministic, structured, and knowledge-dependent tasks operating under strict cost, latency, and reliability constraints. While these are often addressed through large language model (LLM) deployment or distillation into smaller models, we argue this is inefficient, unreliable, and misaligned with enterprise task structures. Instead, AI systems should treat language models as interfaces rather than monolithic engines, externalizing knowledge and computation into dedicated components for greater reliability, scalability, and transparency. Our theoretical evidences show that finite-capacity models cannot fully capture the breadth of knowledge required for enterprise tasks, creating inherent limits to efficiency and interpretability. Building on this, we take the position that language models should primarily be used for structured extraction in deterministic enterprise workflows, while computation and storage are delegated to knowledge bases and symbolic procedures. We formally demonstrate that such modular architectures are more reliable and maintainable than monolithic frameworks, offering a sustainable foundation for enterprise tasks.

Users interacting with Large Language Models (LLMs) in a multi-turn conversation routinely refine their requests or pivot to new topics. LLMs, however, often miss these topic shifts and carry over irrelevant context from previous turns, leading to inaccurate responses. In this paper, we stress-test the multi-turn understanding of LLMs and study the following two sub-tasks: (1) detecting whether the user pivots or refines in the current turn, and (2) shortlisting relevant context from previous turns. To this end, we construct synthetic benchmarks based on real-world datasets from varied domains, as to simulate context shifts of different levels of difficulty. We then evaluate the zero-shot performance of ten LLMs (open-weight, closed-source and reasoning), and demonstrate that only some reasoning and strongly instructed LLMs are accurate in detecting pivots; open-weight LLMs struggle with the task and frequently carry stale context even with explicit cues; and all models suffer from a position bias. Based on the results, we discuss key takeaways for improving long-term robustness in multi-turn capabilities for LLMs.

Large language models fail at counting repeated tokens despite strong performance on broader reasoning benchmarks. These failures are commonly attributed to limitations in internal count tracking. We show this attribution is wrong. Linear probes on the residual stream decode the correct count with near-perfect accuracy at every post-embedding layer, across all model depths. This holds even at the exact layers where the wrong answer crystallizes while the model simultaneously outputs an incorrect count. Attention patterns show no evidence of collapse over repeated tokens and tokenization artifacts account for none of the failure. Instead, a format-triggered multi-layer perceptron (MLP) block overwrites the correctly-encoded count with a fixed wrong answer at roughly 88--93,% network depth. This prior fires for repeated word-tokens in space-separated list format and is absent for repeated digit-tokens. It is suppressed by comma-separated delimiters in larger models but persists in smaller ones. The finding holds across Llama-3.2 (1B and 3B) and Qwen2.5 (1.5B, 3B and 7B) at consistent relative depth. Counting failure is a failure of routing not of representation and the two require different interventions.

Abstract concepts - justice, theory, availability - have no single perceivable referent; in the human brain, their meaning emerges from a web of experiences, affect, and social context. Do large language models (LLMs) ground abstract concepts in a similar way? We study this by replicating property-generation experiments from cognitive science on 21 frontier and open-weight LLMs. Across models and experiments, we find a consistent pattern: when compared to humans, models rely too heavily on word associations, and underproduce properties tied to emotion and internal states. This yields a large and consistent grounding gap: no model exceeds a Pearson correlation r=0.37 with human responses, compared to a human-to-human ceiling above r=0.9. To better interpret this gap, we also replicate a rating experiment on grounding categories and find that here LLMs align more closely with human judgment, and alignment improves as models get larger. We then use sparse autoencoders (SAEs) to inspect whether this information is also reflected in the models' internal features, and we do identify features connected to grounding dimensions such as "sensorimotor" and "social". These findings suggest that current LLMs can recover grounding dimensions when explicitly queried, but do not recruit them in a human-like way when words are generated freely.