Language Model Error Detection

Large language models hallucinate during multi-step reasoning, but most existing detectors operate at the trace level: they assign one confidence score to a full output, fail to localize the first error, and often require multiple sampled completions. We frame hallucination instead as a property of the hidden-state trajectory produced during a single forward pass. Correct reasoning moves through a stable manifold of locally coherent transitions; a first error appears as a localized excursion in transport cost away from this manifold. We operationalize this view with a label-conditioned teacher that builds a trace-specific contrastive PCA lens and scores each step with seven geometric transition features, and a deployable BiLSTM student distilled from the teacher that operates on raw hidden states without inference-time labels. We prove that contrastive PCA is the optimal projection for a transport-separation objective between first error and correct states, and that single-pass first error localization holds whenever the first error creates a positive transport margin over preceding correct transitions. On ProcessBench, PRM800K, HaluEval, and TruthfulQA, both models outperform entropy-based, probing-based, and attention-based baselines in-domain; the teacher transfers stably across language models and datasets, while the student collapses under shift, a gap our distillation theory predicts. These results recast step-level hallucination detection as a problem of trajectory dynamics and identify the central obstacle to deployment: preserving the contrastive transport margin under distribution shift.

Large Language Models for code generation frequently produce hallucinations in Fill-in-the-Middle (FIM) tasks -- plausible but incorrect completions such as invented API methods, invalid parameters, undefined variables, or non-existent imports. These failures pass superficial review yet introduce runtime errors. We introduce Delulu, a verified multi-lingual benchmark of 1,951 FIM samples across 7 languages and 4 hallucination types. Samples are curated through an adversarial pipeline: a frontier LLM generates plausible hallucinations, four diverse judge models evaluate them, embedding-based clustering mines progressively harder examples, self-contained Docker containers verify that golden completions compile while hallucinated variants produce the expected runtime error, and a final human-expert review removes any remaining biased or trivially decidable samples. We evaluate 11 open-weight FIM models from five families spanning 0.5B-32B parameters: a six-point Qwen2.5-Coder scaling slate, plus a cross-family slate (CodeLlama, DeepSeek-Coder-V2, StarCoder2). The strongest model reaches only 84.5% pass@1, no family exceeds 0.77 Edit Similarity, and every family produces hallucination-aligned completions on a non-trivial share of samples, confirming that the difficulty exposed by Delulu is task-intrinsic rather than family-specific. We release the benchmark, containers, and evaluation framework at https://github.com/microsoft/delulu.

We propose a lightweight and single-pass uncertainty quantification method for detecting hallucinations in Large Language Models. The method uses attention matrices to estimate uncertainty without requiring repeated sampling or external models. Specifically, we measure the Kullback-Leibler divergence between each attention head's distribution and a uniform reference distribution, and use these features in a logistic regression probe. Across multiple datasets, task types, and model families, attention divergence is highly predictive of answer correctness and performs competitively with existing uncertainty estimation methods. We find that this signal is concentrated in middle layers and on factual tokens such as named entities and numbers, suggesting that attention dynamics provides an efficient and interpretable white-box signal of model uncertainty.

Most hallucination evaluations focus on English, leaving it unclear whether findings transfer to lower-resource languages. We investigate faithfulness hallucinations, defined as model-generated content that is fluent and plausible but diverges from the provided input or is internally inconsistent. Leveraging the multilingual MultiWikiQA dataset, we utilize the LettuceDetect framework to create synthetic hallucination datasets for 306 languages, from which we train token-level hallucination classifiers for 30 European languages. In this work, we present evaluations of model hallucinations on a selection of languages: English, Danish, German, and Icelandic. Using these classifiers, we evaluate the hallucination rates for Qwen3-0.6B, Qwen3-14B, Gemma-3-12B-IT, cogito-v1-preview-qwen-32B, and cogito-v1-preview-llama-70B. Our classifiers reveal notably higher hallucination rates for Qwen3-0.6B (up to 60% of answers containing at least one hallucination, peaking in Icelandic) and generally lower rates for larger models, with cogito-v1-preview-qwen-32B and cogito-v1-preview-llama-70B performing best on most languages. Hallucination rates are consistently higher for lower-resource languages, particularly Icelandic.

Count-based F1 is widely used as a proxy for LLM error-detection quality, but this paper shows that it can rise dramatically without a corresponding improvement in span localization, a gap termed F1 Inflation. The paper introduces ErrorBench, a controlled stress-test protocol for prompt-induced count distortion. ErrorBench evaluates six contemporary LLMs under five prompt conditions over 4,290 responses from 143 CoNLL-2014 passages. Under CoNLL-2014 M2-style scoring, anchored prompts produce up to 0.79 points of F1 Inflation, and up to 0.96 under strict matching. A 100-passage replication using the official ERRANT 3.0.0 pipeline and multi-reference scoring reproduces the pattern: averaged over six models, the Blind-to-Anchored prompt shift raises Count-F1 by +0.21 while raising multi-reference ERRANT F0.5 by only +0.04. The study finds larger count responses in highly instruction-compliant GPT/Claude systems and smaller responses in the Gemini family under this stress-test protocol. The findings suggest that LLM proofreading and document-review evaluations should avoid pre-populated error counts and should report span-aware metrics alongside count-based metrics.

Autoregressive transformers make confident errors that output-confidence monitoring cannot catch. Activation monitors catch them only when training leaves a decision-quality signal beyond what the output already exposes. This signal is an architectural property of the trained model, fixed upstream of any monitor. Controlling for output confidence removes 60.3% of the raw activation-probe signal on average across 14 models. Raw probe signal is mostly output confidence, and output-side readouts cannot recover the residual. What remains depends on architecture and training. In Pythia's controlled training, both matched-width configurations form the signal early. One preserves it through convergence while another erases it as perplexity continues to improve. Capability and observability are not inherently in tension. Across independently trained families this pattern persists, even as the collapse point shifts. Where the signal survives, monitoring catches what confidence cannot. On downstream QA, a WikiText-trained probe with no task-specific tuning catches about one in eight confident errors that output-confidence monitoring misses, at a 20% flag rate. These results establish signal engineering as a training-time design axis alongside loss and capability. Architecture sets the conditions for observability, and training determines what remains readable.

Large language models can detect their own errors and sometimes correct them without external feedback, but the underlying mechanisms remain unknown. We investigate this through the lens of second-order models of confidence from decision neuroscience. In a first-order system, confidence derives from the generation signal itself and is therefore maximal for the chosen response, precluding error detection. Second-order models posit a partially independent evaluative signal that can disagree with the committed response, providing the basis for error detection. Kumaran et al. (2026) showed that LLMs cache a confidence representation at a token immediately following the answer (i.e. post-answer newline: PANL) -- that causally drives verbal confidence and dissociates from log-probabilities. Here we test whether this PANL signal extends beyond confidence to support error detection and self-correction. Here we test whether this signal supports error detection and self-correction, deriving predictions from the second-order framework. Using a verify-then-correct paradigm, we show that: (i) verbal confidence predicts error detection far beyond token log-probabilities, ruling out a first-order account; (ii) PANL activations predict error detection beyond verbal confidence itself; and (iii) PANL predicts which errors the model can correct -- where all behavioural signals fail. Causal interventions confirm that PANL signals rescue error detection behavior when answer information is corrupted. All findings replicate across models (Gemma 3 27B and Qwen 2.5 7B) and tasks (TriviaQA and MNLI). These results reveal that LLMs naturally implement a second-order confidence architecture whose internal evaluative signal encodes not only whether an answer is likely wrong but whether the model has the knowledge to fix it.

Large language models frequently commit unrecoverable reasoning errors mid-generation: once a wrong step is taken, subsequent tokens compound the mistake rather than correct it. We introduce $\textbf{Latent Phase-Shift Rollback}$ (LPSR): at each generation step, we monitor the residual stream at a critical layer lcrit, detect abrupt directional reversals (phase shifts) via a cosine-similarity $+$ entropy dual gate, and respond by rolling back the KV-cache and injecting a pre-computed steering vector. No fine-tuning, gradient computation, or additional forward passes are required. LPSR achieves $\mathbf{44.0\%}$ on MATH-500 with an 8B model versus $28.8\%$ for standard AR ($+15.2$ pp; McNemar $χ^2 = 66.96$, $p < 10^{-15}$). Critically, prompted self-correction, the most natural inference-time baseline, scores only $19.8\%$, below standard AR; LPSR exceeds it by $+24.2$ pp ($χ^2 = 89.4$, $p \approx 0$). LPSR also outperforms Best-of-16 ($+7.8$ pp) at $5.4\times$ lower token cost, and surpasses a standard 70B model ($35.2\%$) with $8.75\times$ fewer parameters at ${\sim}3\times$ the token budget. A 32-layer sweep reveals a novel detection-correction dissociation: error-detection AUC peaks at layer14 ($0.718$) but task accuracy peaks at layer16 ($44.0\%$ vs.\ $29.2\%$), demonstrating that optimal monitoring depth differs for detection and correction.

This report summarizes the results of the first edition of the Large Language Model (LLM) Testing competition, held as part of the DeepTest workshop at ICSE 2026. Four tools competed in benchmarking an LLM-based car manual information retrieval application, with the objective of identifying user inputs for which the system fails to appropriately mention warnings contained in the manual. The testing solutions were evaluated based on their effectiveness in exposing failures and the diversity of the discovered failure-revealing tests. We report on the experimental methodology, the competitors, and the results.

While Large Language Models (LLMs) have demonstrated remarkable capabilities, their reliability is significantly compromised by hallucinations. Existing intrinsic self-correction methods attempt to address this, but often fail due to self-bias, where models struggle to identify errors in their own outputs without external verification. To overcome these limitations, we propose the LDPC-inspired semantic error correction for retrieval-augmented generation (SERC), providing a theoretical framework to interpret and mitigate LLM hallucinations. We reformulate the text generation process as a semantic noisy channel, treating generated responses as noise-corrupted codewords. Inspired by low-density parity-check (LDPC) codes, SERC employs a sparse verification strategy: instead of exhaustively checking all facts, it generates low-density verification queries and validates them against external evidence to efficiently detect and correct errors. We evaluate SERC on LongForm Bio and TruthfulQA benchmarks using Llama-3-8B and Qwen2.5-14B. Experimental results demonstrate that SERC outperforms both intrinsic self-correction methods and strong retrieval-augmented baselines, demonstrating significant gains especially in factual precision (FactScore). Notably, SERC enables small language models (SLMs) to surpass the performance of larger baselines in hallucination reduction and information preservation. Our findings demonstrate that SERC provides a training-free, model-agnostic solution that significantly reduces verification overhead compared to dense methods, achieving an optimal trade-off between cost and fidelity in resource-constrained environments.

Terminology substitution errors in clinical notes, where one medical term is replaced by a linguistically valid but clinically different term, pose a persistent challenge for automated error detection in healthcare. We introduce BLUEmed, a multi-agent debate framework augmented with hybrid Retrieval-Augmented Generation (RAG) that combines evidence-grounded reasoning with multi-perspective verification for clinical error detection. BLUEmed decomposes each clinical note into focused sub-queries, retrieves source-partitioned evidence through dense, sparse, and online retrieval, and assigns two domain expert agents distinct knowledge bases to produce independent analyses; when the experts disagree, a structured counter-argumentation round and cross-source adjudication resolve the conflict, followed by a cascading safety layer that filters common false-positive patterns. We evaluate BLUEmed on a clinical terminology substitution detection benchmark under both zero-shot and few-shot prompting with multiple backbone models spanning proprietary and open-source families. Experimental results show that BLUEmed achieves the best accuracy (69.13%), ROC-AUC (74.45%), and PR-AUC (72.44%) under few-shot prompting, outperforming both single-agent RAG and debate-only baselines. Further analyses across six backbone models and two prompting strategies confirm that retrieval augmentation and structured debate are complementary, and that the framework benefits most from models with sufficient instruction-following and clinical language understanding.

Discrete masked diffusion language models such as LLaDA generate text through iterative denoising, where mask tokens are progressively replaced with predicted tokens. LLaDA2.1 introduced a Token-to-Token (T2T) editing mechanism that accelerates generation by directly replacing committed tokens suspected of being incorrect. However, we identify fundamental limitations of T2T editing: it couples error detection with replacement, pollutes the generation context with potentially incorrect tokens, and introduces a train-inference noise mismatch where systematic model-generated errors differ from the random perturbations seen during training. We propose Token-to-Mask (T2M) remasking, a training-free, drop-in replacement for T2T editing that resets suspected erroneous tokens back to the mask state, allowing the diffusion process to re-predict them under cleaner context. We design and empirically validate three complementary error detection strategies -- probability-based, trigger-mirrored, and temporal-difference-based -- and provide a unified theoretical analysis showing that T2M remasking purifies the generation context, converts systematic inference errors back to the model's native mask noise type, and enables delayed commitment for joint multi-position optimization. Comprehensive experiments across 12 benchmarks spanning knowledge, reasoning, mathematics, coding, and instruction following show that T2M generally improves performance on tasks requiring precise token-level output, with the largest gain on mathematics (+5.92% on CMATH). Error analysis on CMATH reveals that the dominant failure mode is last-mile token corruption -- where correct reasoning produces a corrupted final answer -- and that T2M repairs 59.4% of such cases.

Large language models with web search are increasingly used in scientific publishing agents, yet they produce BibTeX entries with pervasive field-level errors stemming from omission, partial corruption, substitution, and hallucination. We construct a benchmark of 931 papers across four domains and three citation tiers---popular, low-citation, and recent post-cutoff---with version-aware ground truth. Three search-enabled frontier models (GPT-5, Claude Sonnet-4.6, Gemini-3 Flash) generate approximately 23,000 field-level observations. Overall accuracy is 83.6%, but only 50.9% of entries are fully correct; accuracy drops 27.7 pp from popular to recent papers, revealing heavy reliance on parametric memory even when search is available. Co-occurrence analysis identifies two failure modes: wholesale entry substitution and isolated field error. We present clibib, an open-source tool for deterministic BibTeX retrieval, as a mitigation mechanism. Two-stage integration raises accuracy to 91.5% (+8.0 pp) and fully correct entries to 78.3%, with a 0.8% regression rate. Separating search from revision yields larger gains and lower regression than single-stage tool loops (0.8% vs. 4.8%), demonstrating that integration architecture matters independently of model/tool capability. We release clibib and an accompanying agent skill under the MIT License to improve citation accuracy in increasingly automated scientific workflows.

Message Passing Interface (MPI) is a foundational technology in high-performance computing (HPC), widely used for large-scale simulations and distributed training (e.g., in machine learning frameworks such as PyTorch and TensorFlow). However, maintaining MPI programs remains challenging due to their complex interplay among processes and the intricacies of message passing and synchronization. With the advancement of large language models like ChatGPT, it is tempting to adopt such technology for automated error detection and repair. Yet, our studies reveal that directly applying large language models (LLMs) yields suboptimal results, largely because these models lack essential knowledge about correct and incorrect usage, particularly the bugs found in MPI programs. In this paper, we design a bug detection and repair technique alongside Few-Shot Learning (FSL), Chain-of-Thought (CoT) reasoning, and Retrieval Augmented Generation (RAG) techniques in LLMs to enhance the large language model's ability to detect and repair errors. Surprisingly, such enhancements lead to a significant improvement, from 44% to 77%, in error detection accuracy compared to baseline methods that use ChatGPT directly. Additionally, our experiments demonstrate our bug referencing technique generalizes well to other large language models.

Large Language Models (LLMs) are increasingly used in math education not only as problem solvers but also as assessors of learners' reasoning. However, it remains unclear whether stronger math problem-solving ability is associated with stronger step-level assessment performance. This study examines that relationship using the GSM8K and MATH subsets of PROCESSBENCH, a human-annotated benchmark for identifying the earliest erroneous step in mathematical reasoning. We evaluate two LLM-based math tutor agent settings, instantiated with GPT-4 and GPT-5, in two independent tasks on the same math problems: solving the original problem and assessing a benchmark-provided solution by predicting the earliest erroneous step. Results show a consistent within-model pattern: assessment accuracy is substantially higher on math problem items the same model solved correctly than on items it solved incorrectly, with statistically significant associations across both models and datasets. At the same time, assessment remains more difficult than direct problem solving, especially on error-present solutions. These findings suggest that math problem-solving expertise supports stronger assessment performance, but reliable step-level diagnosis also requires additional capabilities such as step tracking, monitoring, and precise error localization. The results have implications for the design and evaluation of AI-supported Adaptive Instructional Systems (AISs) for formative assessment in math education.

Detecting when a language model is wrong without ground truth labels is a fundamental challenge for safe deployment. Existing approaches rely on a model's own uncertainty -- such as token entropy or confidence scores -- but these signals fail critically on the most dangerous failure mode: confident errors, where a model is wrong but certain. In this work we introduce cross-model disagreement as a correctness indicator -- a simple, training-free signal that can be dropped into existing production systems, routing pipelines, and deployment monitoring infrastructure without modification. Given a model's generated answer, cross-model disagreement computes how surprised or uncertain a second verifier model is when reading that answer via a single forward pass. No generation from the verifying model is required, and no correctness labels are needed. We instantiate this principle as Cross-Model Perplexity (CMP), which measures the verifying model's surprise at the generating model's answer tokens, and Cross-Model Entropy (CME), which measures the verifying model's uncertainty at those positions. Both CMP and CME outperform within-model uncertainty baselines across benchmarks spanning reasoning, retrieval, and mathematical problem solving (MMLU, TriviaQA, and GSM8K). On MMLU, CMP achieves a mean AUROC of 0.75 against a within-model entropy baseline of 0.59. These results establish cross-model disagreement as a practical, training-free approach to label-free correctness estimation, with direct applications in deployment monitoring, model routing, selective prediction, data filtering, and scalable oversight of production language model systems.

As large language models are deployed as autonomous agents with tool execution privileges, a critical assumption underpins their security architecture: that model errors are detectable at runtime. We present empirical evidence that this assumption fails for two of three instruction-following models evaluable for conflict detection. We introduce governability -- the degree to which a model's errors are detectable before output commitment and correctable once detected -- and demonstrate it varies dramatically across models. In six models across twelve reasoning domains, two of three instruction-following models exhibited silent commitment failure: confident, fluent, incorrect output with zero warning signal. The remaining model produced a detectable conflict signal 57 tokens before commitment under greedy decoding. We show benchmark accuracy does not predict governability, correction capacity varies independently of detection, and identical governance scaffolds produce opposite effects across models. A 2x2 experiment shows a 52x difference in spike ratio between architectures but only +/-0.32x variation from fine-tuning, suggesting governability is fixed at pretraining. We propose a Detection and Correction Matrix classifying model-task combinations into four regimes: Governable, Monitor Only, Steer Blind, and Ungovernable.

Evaluating the grammatical competence of second language (L2) learners is essential both for providing targeted feedback and for assessing proficiency. To achieve this, we propose a novel framework leveraging the English Grammar Profile (EGP), a taxonomy of grammatical constructs mapped to the proficiency levels of the Common European Framework of Reference (CEFR), to detect learners' attempts at grammatical constructs and classify them as successful or unsuccessful. This detection can then be used to provide fine-grained feedback. Moreover, the grammatical constructs are used as predictors of proficiency assessment by using automatically detected attempts as predictors of holistic CEFR proficiency. For the selection of grammatical constructs derived from the EGP, rule-based and LLM-based classifiers are compared. We show that LLMs outperform rule-based methods on semantically and pragmatically nuanced constructs, while rule-based approaches remain competitive for constructs that rely purely on morphological or syntactic features and do not require semantic interpretation. For proficiency assessment, we evaluate both rule-based and hybrid pipelines and show that a hybrid approach combining a rule-based pre-filter with an LLM consistently yields the strongest performance. Since our framework operates on pairs of original learner sentences and their corrected counterparts, we also evaluate a fully automated pipeline using automatic grammatical error correction. This pipeline closely approaches the performance of semi-automated systems based on manual corrections, particularly for the detection of successful attempts at grammatical constructs. Overall, our framework emphasises learners' successful attempts in addition to unsuccessful ones, enabling positive, formative feedback and providing actionable insights into grammatical development.

As LLM-powered chatbots are increasingly deployed in mental health services, detecting hallucinations and omissions has become critical for user safety. However, state-of-the-art LLM-as-a-judge methods often fail in high-risk healthcare contexts, where subtle errors can have serious consequences. We show that leading LLM judges achieve only 52% accuracy on mental health counseling data, with some hallucination detection approaches exhibiting near-zero recall. We identify the root cause as LLMs' inability to capture nuanced linguistic and therapeutic patterns recognized by domain experts. To address this, we propose a framework that integrates human expertise with LLMs to extract interpretable, domain-informed features across five analytical dimensions: logical consistency, entity verification, factual accuracy, linguistic uncertainty, and professional appropriateness. Experiments on a public mental health dataset and a new human-annotated dataset show that traditional machine learning models trained on these features achieve 0.717 F1 on our custom dataset and 0.849 F1 on a public benchmark for hallucination detection, with 0.59-0.64 F1 for omission detection across both datasets. Our results demonstrate that combining domain expertise with automated methods yields more reliable and transparent evaluation than black-box LLM judging in high-stakes mental health applications.

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

As Large Language Models shift the programming toward human-guided ''vibe coding'', agentic coding tools increasingly rely on models to self-diagnose and repair their own subtle faults -- a capability central to autonomous software engineering yet never systematically evaluated. We present \name{}, the first empirical decomposition that jointly evaluates two coupled tasks: Fault-Triggering Test Generation (FT-Test) constructing a discriminative witness that exposes a latent bug, and Fault-targeted Program Repair (FPR), repairing it under varying diagnostic conditions. \name{} pairs competitive programming problems with LLM-generated solutions that pass partial test suites but fail on semantic edge cases, enabling controlled identification of where the diagnostic chain breaks down. Evaluating 12 frontier LLMs, we find that fault-targeted reasoning does not scale with general coding ability. Models produce syntactically valid test inputs at near-ceiling rates yet collapse on discriminative generation, with fault hypothesis generation -- not output validation -- as the dominant bottleneck. Test-guided repair reveals a complementary insight: when self-generated tests successfully witness a fault, the resulting repair matches or outperforms repair guided by externally provided tests, but tests that fail to witness the fault actively degrade repair below unguided baselines. Together, these results reframe the challenge of autonomous debugging: the binding bottleneck is not code synthesis or test validity but fault-target reasoning, a capability that remains deficient across all frontier models. As Large Language Models shift the programming toward human-guided ''vibe coding'', agentic coding tools increasingly rely on models to self-diagnose and repair their own subtle faults -- a capability central to autonomous software engineering yet never systematically evaluated.

The integration of Large Language Models into Multi-Agent Systems (MAS) has enabled the so-lution of complex, long-horizon tasks through collaborative reasoning. However, this collec-tive intelligence is inherently fragile, as a single logical fallacy can rapidly propagate and lead to system-wide failure. Most current research re-lies on post-hoc failure analysis, thereby hinder-ing real-time intervention. To address this, we propose PROMAS, a proactive framework utiliz-ing Markov transitions for predictive error anal-ysis. PROMAS extracts Causal Delta Features to capture semantic displacement, mapping them to a quantized Vector Markov Space to model reasoning as probabilistic transitions. By inte-grating a Proactive Prediction Head with Jump Detection, the method localizes errors via risk acceleration rather than static thresholds. On the Who&When benchmark, PROMAS achieves 22.97% step-level accuracy while processing only 27% of reasoning logs. This performance rivals reactive monitors like MASC while reducing data overhead by 73%. Although this strategy entails an accuracy trade-off compared to post-hoc meth-ods, it significantly improves intervention latency, balancing diagnostic precision with the real-time demands of autonomous reasoning.

Many errors in student essays can be explained by influence from the native language (L1). L1 interference refers to errors influenced by a speaker's first language, such as using stadion instead of stadium, reflecting lexical transliteration from Russian. In this work, we address the task of detecting such errors in English essays written by Russian-speaking learners. We introduce RILEC, a large-scale dataset of over 18,000 sentences, combining expert-annotated data from REALEC with synthetic examples generated through rule-based and neural augmentation. We propose a framework for generating L1-motivated errors using generative language models optimized with PPO, prompt-based control, and rule-based patterns. Models fine-tuned on RILEC achieve strong performance, particularly on word-level interference types such as transliteration and tense semantics. We find that the proposed augmentation pipeline leads to a significant performance improvement, making it a potentially valuable tool for learners and teachers to more effectively identify and address such errors.

What happens when a storyteller forgets its own story? Large Language Models (LLMs) can now generate narratives spanning tens of thousands of words, but they often fail to maintain consistency throughout. When generating long-form narratives, these models can contradict their own established facts, character traits, and world rules. Existing story generation benchmarks focus mainly on plot quality and fluency, leaving consistency errors largely unexplored. To address this gap, we present ConStory-Bench, a benchmark designed to evaluate narrative consistency in long-form story generation. It contains 2,000 prompts across four task scenarios and defines a taxonomy of five error categories with 19 fine-grained subtypes. We also develop ConStory-Checker, an automated pipeline that detects contradictions and grounds each judgment in explicit textual evidence. Evaluating a range of LLMs through five research questions, we find that consistency errors show clear tendencies: they are most common in factual and temporal dimensions, tend to appear around the middle of narratives, occur in text segments with higher token-level entropy, and certain error types tend to co-occur. These findings can inform future efforts to improve consistency in long-form narrative generation. Our project page is available at https://picrew.github.io/constory-bench.github.io/.

Errors in medical text can cause delays or even result in incorrect treatment for patients. Recently, language models have shown promise in their ability to automatically detect errors in medical text, an ability that has the opportunity to significantly benefit healthcare systems. In this paper, we explore the importance of prompt optimisation for small and large language models when applied to the task of error detection. We perform rigorous experiments and analysis across frontier language models and open-source language models. We show that automatic prompt optimisation with Genetic-Pareto (GEPA) improves error detection over the baseline accuracy performance from 0.669 to 0.785 with GPT-5 and 0.578 to 0.690 with Qwen3-32B, approaching the performance of medical doctors and achieving state-of-the-art performance on the MEDEC benchmark dataset. Code available on GitHub: https://github.com/CraigMyles/clinical-note-error-detection

Large Language Models (LLMs) are transforming scholarly tasks like search and summarization, but their reliability remains uncertain. Current evaluation metrics for testing LLM reliability are primarily automated approaches that prioritize efficiency and scalability, but lack contextual nuance and fail to reflect how scientific domain experts assess LLM outputs in practice. We developed and validated a schema for evaluating LLM errors in scholarly question-answering systems that reflects the assessment strategies of practicing scientists. In collaboration with domain experts, we identified 20 error patterns across seven categories through thematic analysis of 68 question-answer pairs. We validated this schema through contextual inquiries with 10 additional scientists, which showed not only which errors experts naturally identify but also how structured evaluation schemas can help them detect previously overlooked issues. Domain experts use systematic assessment strategies, including technical precision testing, value-based evaluation, and meta-evaluation of their own practices. We discuss implications for supporting expert evaluation of LLM outputs, including opportunities for personalized, schema-driven tools that adapt to individual evaluation patterns and expertise levels.

As large language models (LLMs) are increasingly deployed in critical decision-making systems, the lack of reliable methods to measure their uncertainty presents a fundamental trustworthiness risk. We introduce a normalized confidence score based on output anchor token probabilities: classification labels for structured tasks and self-evaluation responses (Yes/No) for open-ended generation. This enables direct detection of errors and hallucinations with minimal overhead and without external validation. We make three key contributions. First, we propose a normalized confidence score and self-evaluation framework that exposes reliable confidence estimates for error detection across seven diverse benchmark tasks and five LLMs of varying architectures and sizes. Second, our theoretical analysis reveals that supervised fine-tuning (SFT) yields well-calibrated confidence through maximum-likelihood estimation, whereas reinforcement learning methods (PPO, GRPO) and DPO induce overconfidence via reward exploitation. Third, we propose post-RL SFT with self-distillation to restore confidence reliability in RL-trained models. Empirical results demonstrated that SFT improved average confidence-correctness AUROC from 0.806 to 0.879 and reduced calibration error from 0.163 to 0.034 on Qwen3-4B, while GRPO and DPO degraded confidence reliability. We demonstrated practical value through adaptive retrieval-augmented generation (RAG) that selectively retrieves context when the model lacks confidence, using only 58% of retrieval operations to recover 95% of the maximum achievable accuracy gain on TriviaQA

Large Language Models (LLMs) are increasingly used for content creation in regulated domains such as pharmaceuticals, where content must be medical accurate and legally compliant. Traditional manual quality control (QC) is slow, error-prone, and create publication bottlenecks. To address this, we introduce a modular, LLM/VLMs-driven QC architecture, LRBTC (Language, Regulatory, Brand, Technical, Content Structure), implemented through a Student Teacher HITL architecture and Waterfall rule filtering logic for scalable, verifiable content validation. This architecture is designed to provide verifiable, traceable, and scalable content optimization. Our methods achieve 83.0% F1 and 97.5% recall on AIReg Bench, reducing missed violations five fold compared with Gemini 2.5 Pro. On CSpelling, it improves mean accuracy by +26.7 %. Our error analysis further reveals that while current models are strong at detecting misspellings (92.5% recall), they fail to identify complex grammatical (25.0% recall) and punctuation (41.7% recall) errors, highlighting a key area for future work. This work provides a practical, plug-and-play solution for reliable, transparent quality control of content in high-stakes, compliance-critical industries. We also provide demo access.

Hallucinations remain a major obstacle for large language models (LLMs), especially in safety-critical domains. We present HALT (Hallucination Assessment via Log-probs as Time series), a lightweight hallucination detector that leverages only the top-20 token log-probabilities from LLM generations as a time series. HALT uses a gated recurrent unit model combined with entropy-based features to learn model calibration bias, providing an extremely efficient alternative to large encoders. Unlike white-box approaches, HALT does not require access to hidden states or attention maps, relying only on output log-probabilities. Unlike black-box approaches, it operates on log-probs rather than surface-form text, which enables stronger domain generalization and compatibility with proprietary LLMs without requiring access to internal weights. To benchmark performance, we introduce HUB (Hallucination detection Unified Benchmark), which consolidates prior datasets into ten capabilities covering both reasoning tasks (Algorithmic, Commonsense, Mathematical, Symbolic, Code Generation) and general purpose skills (Chat, Data-to-Text, Question Answering, Summarization, World Knowledge). While being 30x smaller, HALT outperforms Lettuce, a fine-tuned modernBERT-base encoder, achieving a 60x speedup gain on HUB. HALT and HUB together establish an effective framework for hallucination detection across diverse LLM capabilities.

Large language models (LLMs) can generate fluent peer reviews but often miss sparse, subtle, and document-level scientific errors. We introduce PaperAudit-Bench, a benchmark for evaluating paper-level error detection and error-aware automated review. It comprises (1) PaperAudit-Dataset, 220 high-quality AI/ML papers containing more than 2,200 controlled synthetic errors across eight categories, and (2) PaperAudit-Review, an evidence-aware framework that links structured error detection to review generation. Experiments across models, error categories, paper sections, and analysis depths show that sparse long-context auditing remains challenging for current LLMs. Conditioning review generation on explicit, manuscript-grounded error evidence produces more conservative and discriminative assessments that align more closely with human judgments than representative automated-review baselines. The dataset also supports post-training compact detectors through SFT and RL; Qwen3-14B achieves 55.8% F1 at Detection@1 under a fixed held-out protocol. A PubPeer transfer study and deployment in ScienceOne, a scientific foundation-model platform led by the Chinese Academy of Sciences, further illustrate practical applicability. Code and Dataset Page: https://github.com/TU2021/PaperAudit ScienceOne Manuscript Review: https://www.scienceone.cn/lit/#/assistant