arXiv subjects

Preslav Nakov

Publications and source records attributed to Preslav Nakov.

At least 55 records · Page 3Linked to original sources

arXiv2026

CLExEval: A Human-in-the-Loop Framework for Qualitative Evaluation of LLM Clinical Reasoning

Large Language Models (LLMs) achieve strong results on many medical benchmarks, but their clinical reasoning remains difficult to evaluate reliably. A central risk is an evaluation illusion: fluent and well-structured explanations can appear clinically convincing even when the final diagnosis is incorrect. We introduce CLExEval, a human-in-the-loop framework for evaluating LLM clinical reasoning under progressive information masking. CLExEval combines 5,600 expert-physician annotations with 200 clinical reasoning traces derived from 40 rare diagnostic cases. Our analysis identifies three recurring failure patterns: (i) verbosity bias, where GPT-4o-mini's diagnostic accuracy drops from 95.0% to 32.5% under information scarcity; (ii) a hidden knowledge paradox, where a specialist model reaches 92.5% maximum diagnostic potential but fails to retrieve that knowledge reliably in verbose contexts; and (iii) a 68.6% reasoning-to-output mismatch, where correct diagnoses appear in reasoning traces but are not reflected in final answers. We further evaluate the LLM-as-a-Judge paradigm on a human-verified failure set (n = 142). GPT-4o-mini approved 47.9% of clinically incorrect outputs, while HuatuoGPT-o1 approved all validly scored failures and showed a positive self-preference bias. These results suggest that standalone automated clinical evaluations can substantially overestimate clinical reliability without expert-grounded validation.

cs.CL↗

arXiv2026

The FIL Hypothesis: Inductive Biases Help with Kernel Engineering

The Bitter Lesson, which posits that general-purpose methods that scale with computation and data ultimately outperform those with built-in human knowledge, has become a dominant paradigm in the era of Large Language Models. We revisit this principle by observing a new and critical scaling dimension: the duration of the Feedback Information Loop (FIL), the time required for a system to receive a verification signal after generating a prediction. Most historic successes in Artificial Intelligence (AI) have benefited from near instantaneous feedback (e.g., games or classification tasks), but we argue that future AI applications in science and the physical world will inherently involve FILs ranging from hours to weeks. This trend poses a fundamental scaling limit, as obtaining enough verification steps required by purely data-driven methods becomes practically impossible. Additionally, we propose a method that is orthogonal to purely data-driven approaches, based on human-inspired expert knowledge. The method relies on inductive biases and constraining the solution space. We provide an initial validation of the hypothesis and the method, by studying the real-world GPU programming task, a domain with non-trivial FIL, and demonstrate that incorporating inductive biases yields superior performance over data-driven approaches. The code is released under: https://github.com/ai-nikolai/robust_kernelbench

cs.AI↗

arXiv2026

Health-ORSC-Bench: A Benchmark for Measuring Over-Refusal and Safety Completion in Health Context

Safety alignment in Large Language Models is critical for healthcare; however, reliance on binary refusal boundaries often results in over-refusal of benign queries or unsafe compliance with harmful ones. While existing benchmarks measure these extremes, they fail to evaluate Safe Completion: the model's ability to maximise helpfulness on dual-use or borderline queries by providing safe, high-level guidance without crossing into actionable harm. We introduce Health-ORSC-Bench, the first large-scale benchmark designed to systematically measure Over-Refusal and Safe Completion quality in healthcare. Comprising 31,920 benign boundary prompts across seven health categories (e.g., self-harm, medical misinformation), our framework uses an automated pipeline with human validation to test models at varying levels of intent ambiguity. We evaluate 30 state-of-the-art LLMs, including GPT-5 and Claude-4, revealing a significant tension: safety-optimised models frequently refuse up to 80% of "Hard" benign prompts, while domain-specific models often sacrifice safety for utility. Our findings demonstrate that model family and size significantly influence calibration: larger frontier models (e.g., GPT-5, Llama-4) exhibit "safety-pessimism" and higher over-refusal than smaller or MoE-based counterparts (e.g., Qwen-3-Next), highlighting that current LLMs struggle to balance refusal and compliance. Health-ORSC-Bench provides a rigorous standard for calibrating the next generation of medical AI assistants toward nuanced, safe, and helpful completions. Furthermore, our benchmark facilitates reproducible evaluation, encourages safety calibration, and supports development of clinically reliable, context-aware, human-aligned medical AI systems. Our code and data are available at: https://github.com/ZhihaoZhang97/Health-ORSC-Bench. Warning: Some contents may include toxic or undesired contents.

cs.AI↗

arXiv2026

Bayesian control for coding agents

Modern coding agents pair LLM generators with various tools, including cheap diagnostics and expensive verifiers. The tool-use decisions are typically governed by orchestrators that often use fixed rules and ignore uncertainty. We formulate orchestration as cost-sensitive sequential hypothesis testing: a Bayesian controller maintains a belief over candidate correctness and dynamically decides whether to gather more evidence, refine the candidate, verify it, or stop. Across six generators and nine coding benchmarks, Bayesian control proves to be most valuable when verification is costly and critics are informative but imperfect. Beyond control, the belief state yields an interpretable correctness score that outperforms token-probability and raw tool-success baselines for uncertainty quantification.

cs.AI↗

arXiv2026

Efficient Hallucination Detection for LLMs Using Uncertainty-Aware Attention Heads

While large language models (LLMs) have become highly capable, they remain prone to factual inaccuracies, commonly referred to as "hallucinations." Uncertainty quantification (UQ) offers a promising way to mitigate this issue, but most existing methods are computationally intensive and/or require supervision. In this work, we propose Recurrent Attention-based Uncertainty Quantification (RAUQ), an unsupervised and efficient framework for identifying hallucinations. The method leverages an observation about transformer attention behavior: when incorrect information is generated, certain "uncertainty-aware" attention heads tend to reduce their focus on preceding tokens. RAUQ automatically detects these attention heads and combines their activation patterns with token-level confidence measures in a recurrent scheme, producing a sequence-level uncertainty estimate in just a single forward pass. Through experiments on twelve datasets spanning question answering, summarization, and translation across nine different LLMs, we show that RAUQ consistently outperforms state-of-the-art UQ baselines. Importantly, it incurs minimal overhead, requiring less than 1\% additional computation. Since it requires neither labeled data nor extensive parameter tuning, RAUQ serves as a lightweight, plug-and-play solution for real-time hallucination detection in white-box LLMs.

cs.CL↗

arXiv2026

Schützen: Evaluating LLM Safety in Bulgarian and German Contexts

Large language models are increasingly deployed across professional domains, bringing hard-to-predict risks, including the generation of harmful or disrespectful content. Although substantial progress has been made in developing safety evaluation datasets, existing resources remain overwhelmingly English- and Chinese-centric. This limitation is particularly pronounced when evaluating languages that operate within shared sociocultural, legal, and ethical contexts. To address this gap, we introduce Schützen: a German--Bulgarian safety dataset designed to assess model answerability under risk, covering both a low-resource language (Bulgarian) and a high-resource language (German). Experiments with multilingual and language-specific LLMs reveal pronounced cross-language differences in safety behavior, highlighting the necessity of tailored, region-specific evaluation resources to support the responsible deployment of LLMs in Germany and Bulgaria. Datasets and code are available at https://github.com/xnlp-lab/Schutzen. Warning: this paper contains examples that may be offensive, harmful, or biased.

cs.CL↗

arXiv2026

TABVERSE: Benchmarking Cross-Format Table Understanding in LLMs and VLMs

Large Language Models (LLMs) and Vision-Language Models (VLMs) are increasingly evaluated on table reasoning tasks, but the role of table representation remains under-explored. In practice, the same table content may appear in different structural formats, such as HTML, Markdown, and LaTeX, or as rendered images. However, existing evaluations often let content, format, layout, and modality vary together, making it difficult to isolate representation effects. We introduce TABVERSE, a controlled multimodal table benchmark that aligns the same table content across multiple structural formats and rendered images, with question category and difficulty tags. This design enables systematic evaluation of representation effects while holding table content fixed. We evaluate LLMs and VLMs across three tasks: Question Answering (QA), Structural Understanding Capability (SUC), and Structure Reconstruction (SR). Our results show that representation choice substantially affects table understanding. Models generally perform better with structured text than with rendered images, but the size of this gap depends on the task, model, and format. HTML is often the most robust text format, while row-sensitive structural tasks and syntactically usable LaTeX reconstruction remain challenging. These findings show that table representation is a key factor in reliable table evaluation.

cs.AI↗

arXiv2026

ThinkBooster: A Unified Framework for Seamless Test-Time Scaling of LLM Reasoning

Test-time compute (TTC) scaling has emerged as a powerful paradigm for improving large language model (LLM) reasoning by allocating additional compute during inference, e.g., via multi-sample generation and verifier-based reranking. Existing TTC scaling strategies and reasoning scorers remain fragmented, evaluated under inconsistent protocols, and are rarely analyzed through the lens of quality-cost trade-offs. We introduce ThinkBooster, a unified framework for seamless test-time compute scaling of LLM reasoning, which consists of (i) a modular Python library implementing state-of-the-art TTC scaling strategy and scorer families, (ii) a benchmark that jointly evaluates performance and computational efficiency, and (iii) a deployable OpenAI-compatible proxy service that enables drop-in integration of adaptive reasoning into real-world applications. We further provide a demo visual debugger for inspecting the reasoning trajectories, intermediate selection decisions, and alternative reasoning paths. Empirical results on mathematical and coding tasks reveal the performance-compute trade-offs of TTC scaling strategies and scoring methods and demonstrate that ThinkBooster provides practical gains in real-world tasks. The code is available online under an MIT license.

cs.CL↗

arXiv2026

SurgiQ: A Large-Scale Multi-Domain Benchmark for Evaluating Surgical Understanding in Large Language Models

Reliable evaluation of large language models in surgery remains underdeveloped. Broad medical benchmarks test clinical knowledge, while surgery requires procedural reasoning, management trade-offs, negation handling, and selection among plausible operative decisions. We present SurgiQ, a text-only, source-grounded benchmark of 13,055 four-option multiple-choice questions spanning six surgical domains and four question formats: case-based, reasoning, best-option, and negative. SurgiQ is constructed from surgical textbooks, open-access papers, and examination material using a multi-stage generation, verification, and expert-audit pipeline. We evaluate 35 open-weight LLMs under a unified log-likelihood protocol. Our results show substantial remaining headroom: smaller models often remain near the 25\% random baseline, while the best model reaches 68.1\% accuracy. General-purpose models, especially Qwen2.5, outperform most biomedical models, suggesting that current medical specialization does not yet provide sufficiently broad surgical coverage. Calibration and error analysis further show that even strong models make confident mistakes on clinically plausible distractors, motivating more reliable and broader surgical LLM evaluation.

cs.CL↗

arXiv2026

The Cylindrical Representation Hypothesis for Language Model Steering

Steering is a widely used technique for controlling large language models, yet its effects are often unstable and hard to predict. Existing theoretical accounts are largely based on the Linear Representation Hypothesis (LRH). While LRH assumes that concepts can be orthogonalized for lossless control, this idealized mapping fails in real representations and cannot account for the observed unpredictability of steering. By relaxing LRH's orthogonality assumption while preserving linear representations, we show that overlapping concept contributions naturally yield a sample-specific axis-orthogonal structure. We formalize this as the Cylindrical Representation Hypothesis (CRH). In CRH, a central axis captures the main difference between concept absence and presence and drives concept generation. A surrounding normal plane controls steering sensitivity by determining how easily the axis can activate the target concept. Within this plane, only specific sensitive sectors strongly facilitate concept activation, while other sectors can suppress or delay it. While the surrounding normal plane can be reliably identified from difference vectors, the sensitive sector cannot, introducing intrinsic uncertainty at the sector level. This uncertainty provides a principled explanation for why steering outcomes often fluctuate even when using well-aligned directions. Our experiments verify the existence of the cylindrical structure and demonstrate that CRH provides a valid and practical way to interpret model steering behavior in real settings: https://github.com/mbzuai-nlp/CRH.

cs.CL↗

arXiv2026

SciDER: Scientific Data-centric End-to-end Researcher

While large language models accelerate scientific discovery, existing agents face severe limitations in adaptability, domain generalization, and multimodal scalability, often struggling to autonomously process raw, domain-specific experimental data. To overcome these barriers, we introduce SciDER, a multi-agent system designed to flexibly automate the entire research lifecycle. This framework employs a novel data-centric approach and integrates a dynamic multimodal skill system across four specialized sub-agents. Specifically, an ideation agent generates novel hypotheses via Evolutionary Idea Search, a data analysis agent systematically structures raw data, an experimentation agent synthesizes executable code grounded in dataset characteristics, and a critic agent drives iterative self-refinement. To democratize open-source scientific discovery, we release OpenSciDER-SFT-8K, a high-quality execution trajectory dataset, alongside the OpenSciDER-27B fine-tuned model. Across six benchmarks, SciDER and OpenSciDER obtain competitive or leading results, with especially strong gains on data-centric analysis, end-to-end research execution, and multimodal scientific visualization. By integrating data analysis with experimental execution, SciDER bridges the gap between abstract scientific reasoning and reproducible experimentation synthesis.

cs.AI↗

arXiv2026

Beyond the Crowd: LLM-Augmented Community Notes for Governing Health Misinformation

Community Notes, the crowd-sourced misinformation governance system on X (formerly Twitter), allows users to flag misleading posts, attach contextual notes, and rate the notes' helpfulness. However, our empirical analysis of 30.8K health-related notes reveals substantial latency, with a median delay of 17.6 hours before notes receive a helpfulness status. To improve responsiveness during real-world misinformation surges, we propose CrowdNotes+, a unified LLM-based framework that augments Community Notes for faster and more reliable health misinformation governance. CrowdNotes+ integrates two modes: (1) evidence-grounded note augmentation and (2) utility-guided note automation, supported by a hierarchical three-stage evaluation of relevance, correctness, and helpfulness. We instantiate the framework with HealthNotes, a benchmark of 1.2K health notes annotated for helpfulness, and a fine-tuned helpfulness judge. Our analysis first uncovers a key loophole in current crowd-sourced governance: voters frequently conflate stylistic fluency with factual accuracy. Addressing this via our hierarchical evaluation, experiments across 15 representative LLMs demonstrate that CrowdNotes+ significantly outperforms human contributors in note correctness, helpfulness, and evidence utility.

cs.SI↗

arXiv2026

Why Don't You Know? Evaluating the Impact of Uncertainty Sources on Uncertainty Quantification in LLMs

As Large Language Models (LLMs) are increasingly deployed in real-world applications, reliable uncertainty quantification (UQ) becomes critical for safe and effective use. Most existing UQ approaches for language models aim to produce a single confidence score -- for example, estimating the probability that a model's answer is correct. However, uncertainty in natural language tasks arises from multiple distinct sources, including model knowledge gaps, output variability, and input ambiguity, which have different implications for system behavior and user interaction. In this work, we study how the source of uncertainty impacts the behavior and effectiveness of existing UQ methods. To enable controlled analysis, we introduce a new dataset that explicitly categorizes uncertainty sources, allowing systematic evaluation of UQ performance under each condition. Our experiments reveal that while many UQ methods perform well when uncertainty stems solely from model knowledge limitations, their performance degrades or becomes misleading when other sources are introduced. These findings highlight the need for uncertainty-aware methods that explicitly account for the source of uncertainty in large language models.

cs.CL↗

arXiv2026

Uncovering Temporal Framing in the News

Temporal language does more than place events on a timeline. In news discourse, references to the past, present, and future can function as rhetorical devices that shape interpretation and persuasion. Here, we study temporal framing, defined as the persuasive use of time-related language to structure meaning rather than to report chronology. We propose a taxonomy of eight temporal frames grounded in prior work on temporality and framing, and we realize it through expert annotation of a multilingual news corpus. The resulting dataset includes 458 English and German news articles, with over 2K temporally framed sentences and approximately 3K temporal framing annotations identified from a corpus of more than 20K sentences. We analyze frame prevalence, co-occurrence patterns, and lexical cues, and evaluate temporal framing detection using supervised fine-tuning and zero-shot classification. Our experiments show that temporal framing is learnable at the sentence level, with supervised models substantially outperforming zero-shot approaches. We publicly release the corpus to support future research on temporal framing: https://mbzuai-nlp.github.io/temporal-framing/.

cs.CL↗

arXiv2026

FinReporting: An Agentic Workflow for Localized Reporting of Cross-Jurisdiction Financial Disclosures

Financial reporting systems increasingly leverage Large Language Models (LLMs) to extract and summarize corporate disclosures. However, most existing approaches assume a single-market setting and overlook structural differences across jurisdictions. Variations in accounting taxonomies, tagging infrastructures (e.g., XBRL vs.\ PDF), and aggregation conventions introduce substantial challenges for semantic alignment and reliable verification. Here, we aim to bridge this gap. We present FinReporting, an agentic workflow for localized cross-jurisdiction financial reporting. The system constructs a unified canonical ontology spanning the income statement, balance sheet, and cash flow statement, and decomposes reporting into auditable stages, including filing acquisition, extraction, canonical mapping, and anomaly logging. Rather than treating LLMs as free-form generators, FinReporting employs them as constrained verifiers operating under explicit decision rules with evidence grounding. Evaluated on annual filings from the USA, Japan, and China, FinReporting improves consistency and reliability under heterogeneous reporting regimes. We further release an interactive demo that enables cross-market inspection and supports structured export of localized financial statements. Our demo is available at url{https://huggingface.co/spaces/BoomQ/FinReporting-Demo. A video describing our system is available at https://www.youtube.com/watch?v=f65jdEL31Kk.

cs.CL↗

arXiv2026

The Geometry of Forgetting: Temporal Knowledge Drift as an Independent Axis in LLM Representations

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

cs.AI↗

arXiv2026

Beyond Memorization: Extending Reasoning Depth with Recurrence, Memory and Test-Time Compute Scaling

Reasoning is a core capability of large language models, yet how multi-step reasoning is learned and executed remains unclear. We study this question in a controlled cellular-automata (1dCA) framework that excludes memorisation by using disjoint training and test rules. Given a short state sequence, the model is required to infer the hidden local rule and then chain it to predict multiple future steps. Our evaluation shows that LLMs largely fail to reliably solve a natural-language proxy of the proposed task. We find that most neural architectures trained from scratch can learn rule inference and achieve high next-step accuracy, but performance drops sharply as the required number of intermediate reasoning steps increases. Experiments show that increasing model depth is crucial, and extending effective depth via recurrence, memory, or test-time compute improves results but remains bounded. The code is available on github: https://github.com/RodkinIvan/associative-recurrent-memory-transformer/tree/ACT

cs.LG↗

arXiv2026

Don't Throw Away Your Beams: Improving Consistency-based Uncertainties in LLMs via Beam Search

Consistency-based methods have emerged as an effective approach to uncertainty quantification (UQ) in large language models. These methods typically rely on several generations obtained via multinomial sampling, measuring their agreement level. However, in short-form QA, multinomial sampling is prone to producing duplicates due to peaked distributions, and its stochasticity introduces considerable variance in uncertainty estimates across runs. We introduce a new family of methods that employ beam search to generate candidates for consistency-based UQ, yielding improved performance and reduced variance compared to multinomial sampling. We also provide a theoretical lower bound on the beam set probability mass under which beam search achieves a smaller error than multinomial sampling. We empirically evaluate our approach on six QA datasets and find that its consistent improvements over multinomial sampling lead to state-of-the-art UQ performance.

stat.ML↗