arXiv subjects

Preslav Nakov

Publications and source records attributed to Preslav Nakov.

At least 91 records · Page 5Linked to original sources

arXiv2026

AICD Bench: A Challenging Benchmark for AI-Generated Code Detection

Large language models (LLMs) are increasingly capable of generating functional source code, raising concerns about authorship, accountability, and security. While detecting AI-generated code is critical, existing datasets and benchmarks are narrow, typically limited to binary human-machine classification under in-distribution settings. To bridge this gap, we introduce $\emph{AICD Bench}$, the most comprehensive benchmark for AI-generated code detection. It spans $\emph{2M examples}$, $\emph{77 models}$ across $\emph{11 families}$, and $\emph{9 programming languages}$, including recent reasoning models. Beyond scale, AICD Bench introduces three realistic detection tasks: ($\emph{i}$)~$\emph{Robust Binary Classification}$ under distribution shifts in language and domain, ($\emph{ii}$)~$\emph{Model Family Attribution}$, grouping generators by architectural lineage, and ($\emph{iii}$)~$\emph{Fine-Grained Human-Machine Classification}$ across human, machine, hybrid, and adversarial code. Extensive evaluation on neural and classical detectors shows that performance remains far below practical usability, particularly under distribution shift and for hybrid or adversarial code. We release AICD Bench as a $\emph{unified, challenging evaluation suite}$ to drive the next generation of robust approaches for AI-generated code detection. The data and the code are available at https://huggingface.co/AICD-bench}.

cs.LG↗

arXiv2026

DP-Fusion: Token-Level Differentially Private Inference for Large Language Models

Large language models (LLMs) do not preserve privacy at inference-time. The LLM's outputs can inadvertently reveal information about the model's context, which presents a privacy challenge when the LLM is augmented via tools or databases containing sensitive information. Existing privacy-preserving methods at inference-time have significant limitations since they (i) lack provable guarantees or (ii) have a poor utility/privacy trade-off. We propose DP-Fusion, a Differentially Private Inference (DPI) mechanism for LLMs that provably bounds the influence a set of tokens in the context can have on the LLM's output. DP-Fusion works as follows: (1) label a subset of sensitive tokens, (2) infer the LLM without any sensitive tokens to obtain a baseline, (3) infer the LLM with the sensitive tokens, and (4) blend distributions so that the final output remains within a bounded distance of the baseline distribution. While this per-token influence bound also mitigates jailbreak-style prompt injection, we focus on \emph{document privatization}, where the goal is to paraphrase a document containing sensitive tokens, e.g., personally identifiable information, so that no attacker can reliably infer them from the paraphrased document while preserving high text quality. The privacy/utility trade-off is controlled by $ε$, where $ε=0$ hides sensitive tokens entirely, while higher values trade off privacy for improved text quality. We show that our method creates token-level provably privatized documents with substantially improved theoretical and empirical privacy, achieving $6\times$ lower perplexity than related DPI methods.

cs.CL↗

arXiv2026

MuSLR: Multimodal Symbolic Logical Reasoning

Multimodal symbolic logical reasoning, which aims to deduce new facts from multimodal input via formal logic, is critical in high-stakes applications such as autonomous driving and medical diagnosis, as its rigorous, deterministic reasoning helps prevent serious consequences. To evaluate such capabilities of current state-of-the-art vision language models (VLMs), we introduce the first benchmark MuSLR for multimodal symbolic logical reasoning grounded in formal logical rules. MuSLR comprises 1,093 instances across 7 domains, including 35 atomic symbolic logic and 976 logical combinations, with reasoning depths ranging from 2 to 9. We evaluate 7 state-of-the-art VLMs on MuSLR and find that they all struggle with multimodal symbolic reasoning, with the best model, GPT-4.1, achieving only 46.8%. Thus, we propose LogiCAM, a modular framework that applies formal logical rules to multimodal inputs, boosting GPT-4.1's Chain-of-Thought performance by 14.13%, and delivering even larger gains on complex logics such as first-order logic. We also conduct a comprehensive error analysis, showing that around 70% of failures stem from logical misalignment between modalities, offering key insights to guide future improvements. All data and code are publicly available at https://llm-symbol.github.io/MuSLR.

cs.CV↗

arXiv2026

COMMUNITYNOTES: A Dataset for Exploring the Helpfulness of Fact-Checking Explanations

Fact-checking on major platforms, such as X, Meta, and TikTok, is shifting from expert-driven verification to a community-based setup, where users contribute explanatory notes to clarify why a post might be misleading. An important challenge here is determining whether an explanation is helpful for understanding real-world claims and the reasons why, which remains largely underexplored in prior research. In practice, most community notes remain unpublished due to slow community annotation, and the reasons for helpfulness lack clear definitions. To bridge these gaps, we introduce the task of predicting both the helpfulness of explanatory notes and the reason for this. We present COMMUNITYNOTES, a large-scale multilingual dataset of 104k posts with user-provided notes and helpfulness labels. We further propose a framework that automatically generates and improves reason definitions via automatic prompt optimization, and integrate them into prediction. Our experiments show that the optimized definitions can improve both helpfulness and reason prediction. Finally, we show that the helpfulness information is beneficial for existing fact-checking systems.

cs.CL↗

arXiv2026

Beyond "Not Novel Enough": Enriching Scholarly Critique with LLM-Assisted Feedback

Novelty assessment is a central yet understudied aspect of peer review, particularly in high volume fields like NLP where reviewer capacity is increasingly strained. We present a structured approach for automated novelty evaluation that models expert reviewer behavior through three stages: content extraction from submissions, retrieval and synthesis of related work, and structured comparison for evidence based assessment. Our method is informed by a large scale analysis of human written novelty reviews and captures key patterns such as independent claim verification and contextual reasoning. Evaluated on 182 ICLR 2025 submissions with human annotated reviewer novelty assessments, the approach achieves 86.5% alignment with human reasoning and 75.3% agreement on novelty conclusions - substantially outperforming existing LLM based baselines. The method produces detailed, literature aware analyses and improves consistency over ad hoc reviewer judgments. These results highlight the potential for structured LLM assisted approaches to support more rigorous and transparent peer review without displacing human expertise. Data and code are made available.

cs.CL↗

arXiv2026

YaPO: Learnable Sparse Activation Steering Vectors for Domain Adaptation

Steering Large Language Models (LLMs) through activation interventions has emerged as a lightweight alternative to fine-tuning for alignment and personalization. Recent work on Bi-directional Preference Optimization (BiPO) shows that dense steering vectors can be learned directly from preference data in a Direct Preference Optimization (DPO) fashion, enabling control over truthfulness, hallucinations, and safety behaviors. However, dense steering vectors often entangle multiple latent factors due to neuron multi-semanticity, limiting their effectiveness and stability in fine-grained settings such as cultural alignment, where closely related values and behaviors (e.g., among Middle Eastern cultures) must be distinguished. In this paper, we propose Yet another Policy Optimization (YaPO), a \textit{reference-free} method that learns \textit{sparse steering vectors} in the latent space of a Sparse Autoencoder (SAE). By optimizing sparse codes, YaPO produces disentangled, interpretable, and efficient steering directions. Empirically, we show that YaPO converges faster, achieves stronger performance, and exhibits improved training stability compared to dense steering baselines. Beyond cultural alignment, YaPO generalizes to a range of alignment-related behaviors, including hallucination, wealth-seeking, jailbreak, and power-seeking. Importantly, YaPO preserves general knowledge, with no measurable degradation on MMLU. Overall, our results show that YaPO provides a general recipe for efficient, stable, and fine-grained alignment of LLMs, with broad applications to controllability and domain adaptation. The associated code and data are publicly available\footnote{https://github.com/MBZUAI-Paris/YaPO}.

cs.AI↗

arXiv2026

MuDRiC: Multi-Dialect Reasoning for Arabic Commonsense Validation

Commonsense validation evaluates whether a sentence aligns with everyday human understanding, a critical capability for developing robust natural language understanding systems. While substantial progress has been made in English, the task remains underexplored in Arabic, particularly given its rich linguistic diversity. Existing Arabic resources have primarily focused on Modern Standard Arabic (MSA), leaving regional dialects underrepresented despite their prevalence in spoken contexts. To bridge this gap, we present two key contributions. We introduce MuDRiC, an extended Arabic commonsense dataset incorporating multiple dialects. To the best of our knowledge, this is the first Arabic multi-dialect commonsense reasoning dataset. We further propose a novel method adapting Graph Convolutional Networks (GCNs) to Arabic commonsense reasoning, which enhances semantic relationship modeling for improved commonsense validation. Our experimental results demonstrate that this approach consistently outperforms the baseline of direct language model fine-tuning. Overall, our work enhances Arabic natural language understanding by providing a foundational dataset and a new method for handling its complex variations. Data and code are available at https://github.com/KareemElozeiri/MuDRiC.

cs.CL↗

arXiv2026

How Does Prefix Matter in Reasoning Model Tuning?

Recent alignment studies commonly remove introductory boilerplate phrases from supervised fine-tuning (SFT) datasets. This work challenges that assumption. We hypothesize that safety- and reasoning-oriented prefix sentences serve as lightweight alignment signals that can guide model decoding toward safer and more coherent responses. To examine this, we fine-tune three R1 series models across three core model capabilities: reasoning (mathematics, coding), safety, and factuality, systematically varying prefix inclusion from 0% to 100%. Results show that prefix-conditioned SFT improves both safety and reasoning performance, yielding up to +6% higher Safe@1 accuracy on adversarial benchmarks (WildJailbreak, StrongReject) and +7% improvement on GSM8K reasoning. However, factuality and coding tasks show marginal or negative effects, indicating that prefix-induced narrowing of the search space benefits structured reasoning. Token-level loss analysis further reveals that prefix tokens such as "revised" and "logically" incur higher gradient magnitudes, acting as alignment anchors that stabilize reasoning trajectories. Our findings suggest that prefix conditioning offers a scalable and interpretable mechanism for improving reasoning safety, serving as an implicit form of alignment that complements traditional reward-based methods.

cs.CL↗

arXiv2025

Uncertainty Quantification for LLMs through Minimum Bayes Risk: Bridging Confidence and Consistency

Uncertainty quantification (UQ) methods for Large Language Models (LLMs) encompass a variety of approaches, with two major types being particularly prominent: information-based, which focus on model confidence expressed as token probabilities, and consistency-based, which assess the semantic relationship between multiple outputs generated using repeated sampling. Several recent methods have combined these two approaches to boost UQ performance. However, they sometimes fail to outperform much simpler baseline methods. Our work discusses the fundamental approach to constructing uncertainty measures that directly links uncertainty with the minimum Bayes risks achieved by LLM decoding. Building on these findings, we propose a novel approach to integrating model confidence with output consistency, resulting in a family of efficient and robust UQ methods. Our investigation reveals distinctive characteristics of LLMs as probabilistic models, which help to explain why these UQ methods underperform in certain tasks. Based on these findings, we propose a new way of synthesizing model confidence and output consistency, leading to a family of efficient and robust UQ methods. We evaluate our approach across various tasks such as question answering, abstractive summarization, and machine translation, demonstrating sizable improvements over state-of-the-art UQ approaches.

cs.CL↗

arXiv2025

From Generation to Detection: A Multimodal Multi-Task Dataset for Benchmarking Health Misinformation

Infodemics and health misinformation have significant negative impact on individuals and society, exacerbating confusion and increasing hesitancy in adopting recommended health measures. Recent advancements in generative AI, capable of producing realistic, human like text and images, have significantly accelerated the spread and expanded the reach of health misinformation, resulting in an alarming surge in its dissemination. To combat the infodemics, most existing work has focused on developing misinformation datasets from social media and fact checking platforms, but has faced limitations in topical coverage, inclusion of AI generation, and accessibility of raw content. To address these issues, we present MM Health, a large scale multimodal misinformation dataset in the health domain consisting of 34,746 news article encompassing both textual and visual information. MM Health includes human-generated multimodal information (5,776 articles) and AI generated multimodal information (28,880 articles) from various SOTA generative AI models. Additionally, We benchmarked our dataset against three tasks (reliability checks, originality checks, and fine-grained AI detection) demonstrating that existing SOTA models struggle to accurately distinguish the reliability and origin of information. Our dataset aims to support the development of misinformation detection across various health scenarios, facilitating the detection of human and machine generated content at multimodal levels.

cs.CL↗

arXiv2025

UnsafeChain: Enhancing Reasoning Model Safety via Hard Cases

As large reasoning models (LRMs) grow more capable, chain-of-thought (CoT) reasoning introduces new safety challenges. Existing SFT-based safety alignment studies dominantly focused on filtering prompts with safe, high-quality responses, while overlooking hard prompts that always elicit harmful outputs. To fill this gap, we introduce UnsafeChain, a safety alignment dataset constructed from hard prompts with diverse sources, where unsafe completions are identified and explicitly corrected into safe responses. By exposing models to unsafe behaviors and guiding their correction, UnsafeChain enhances safety while preserving general reasoning ability. We fine-tune three LRMs on UnsafeChain and compare them against recent SafeChain and STAR-1 across six out-of-distribution and five in-distribution benchmarks. UnsafeChain consistently outperforms prior datasets, with even a 1K subset matching or surpassing baseline performance, demonstrating the effectiveness and generalizability of correction-based supervision. We release our dataset and code at https://github.com/mbzuai-nlp/UnsafeChain

cs.CL↗

arXiv2025

OpenFactCheck: Building, Benchmarking Customized Fact-Checking Systems and Evaluating the Factuality of Claims and LLMs

The increased use of large language models (LLMs) across a variety of real-world applications calls for mechanisms to verify the factual accuracy of their outputs. Difficulties lie in assessing the factuality of free-form responses in open domains. Also, different papers use disparate evaluation benchmarks and measurements, which renders them hard to compare and hampers future progress. To mitigate these issues, we propose OpenFactCheck, a unified framework for building customized automatic fact-checking systems, benchmarking their accuracy, evaluating factuality of LLMs, and verifying claims in a document. OpenFactCheck consists of three modules: (i) CUSTCHECKER allows users to easily customize an automatic fact-checker and verify the factual correctness of documents and claims, (ii) LLMEVAL, a unified evaluation framework assesses LLM's factuality ability from various perspectives fairly, and (iii) CHECKEREVAL is an extensible solution for gauging the reliability of automatic fact-checkers' verification results using human-annotated datasets. Data and code are publicly available at https://github.com/yuxiaw/openfactcheck.

cs.CL↗

arXiv2025

OpenFactCheck: A Unified Framework for Factuality Evaluation of LLMs

The increased use of large language models (LLMs) across a variety of real-world applications calls for automatic tools to check the factual accuracy of their outputs, as LLMs often hallucinate. This is difficult as it requires assessing the factuality of free-form open-domain responses. While there has been a lot of research on this topic, different papers use different evaluation benchmarks and measures, which makes them hard to compare and hampers future progress. To mitigate these issues, we developed OpenFactCheck, a unified framework, with three modules: (i) RESPONSEEVAL, which allows users to easily customize an automatic fact-checking system and to assess the factuality of all claims in an input document using that system, (ii) LLMEVAL, which assesses the overall factuality of an LLM, and (iii) CHECKEREVAL, a module to evaluate automatic fact-checking systems. OpenFactCheck is open-sourced (https://github.com/mbzuai-nlp/openfactcheck) and publicly released as a Python library (https://pypi.org/project/openfactcheck/) and also as a web service (http://app.openfactcheck.com). A video describing the system is available at https://youtu.be/-i9VKL0HleI.

cs.CL↗

arXiv2025

PTPP-Aware Adaptation Scaling Laws: Predicting Domain-Adaptation Performance at Unseen Pre-Training Budgets

Continual pre-training (CPT) for domain adaptation must balance target-domain gains with stability on the base domain. Existing CPT scaling laws typically assume a fixed pre-training budget, which limits their ability to forecast adaptation outcomes for models trained at different tokens-per-parameter (PTPP). We present \emph{PTPP-aware} adaptation scaling laws that make the pre-training budget an explicit variable, enabling accurate \emph{prediction} of adaptation loss at unseen \ptpp. On a multilingual setup (English/Arabic $\rightarrow$ French), PTPP-aware formulations trained on early stages (\ptpp{}=\{15,31\}) predict target loss at \ptpp{}=279 and outperform a PTPP-agnostic \dcpt{} transfer baseline on metrics (Huber-on-log, MAE$_\mathrm{rel}$, calibration slope); full diagnostics (RMSE, MAPE) are in the appendix. Beyond forecasting, we show a practical use case: planning replay ratios and adaptation token budgets that satisfy target and forgetting constraints under compute limits.

cs.LG↗

arXiv2025

Unconditional Truthfulness: Learning Unconditional Uncertainty of Large Language Models

Uncertainty quantification (UQ) has emerged as a promising approach for detecting hallucinations and low-quality output of Large Language Models (LLMs). However, obtaining proper uncertainty scores is complicated by the conditional dependency between the generation steps of an autoregressive LLM because it is hard to model it explicitly. Here, we propose to learn this dependency from attention-based features. In particular, we train a regression model that leverages LLM attention maps, probabilities on the current generation step, and recurrently computed uncertainty scores from previously generated tokens. To incorporate the recurrent features, we also suggest a two-staged training procedure. Our experimental evaluation on ten datasets and three LLMs shows that the proposed method is highly effective for selective generation, achieving substantial improvements over rivaling unsupervised and supervised approaches.

cs.CL↗

arXiv2025

HALF: Harm-Aware LLM Fairness Evaluation Aligned with Deployment

Large language models (LLMs) are increasingly deployed across high-impact domains, from clinical decision support and legal analysis to hiring and education, making fairness and bias evaluation before deployment critical. However, existing evaluations lack grounding in real-world scenarios and do not account for differences in harm severity, e.g., a biased decision in surgery should not be weighed the same as a stylistic bias in text summarization. To address this gap, we introduce HALF (Harm-Aware LLM Fairness), a deployment-aligned framework that assesses model bias in realistic applications and weighs the outcomes by harm severity. HALF organizes nine application domains into three tiers (Severe, Moderate, Mild) using a five-stage pipeline. Our evaluation results across eight LLMs show that (1) LLMs are not consistently fair across domains, (2) model size or performance do not guarantee fairness, and (3) reasoning models perform better in medical decision support but worse in education. We conclude that HALF exposes a clear gap between previous benchmarking success and deployment readiness.

cs.CL↗

arXiv2025

Sherkala-Chat: Building a State-of-the-Art LLM for Kazakh in a Moderately Resourced Setting

Llama-3.1-Sherkala-8B-Chat, or Sherkala-Chat (8B) for short, is a state-of-the-art instruction-tuned open generative large language model (LLM) designed for Kazakh. Sherkala-Chat (8B) aims to enhance the inclusivity of LLM advancements for Kazakh speakers. Adapted from the LLaMA-3.1-8B model, Sherkala-Chat (8B) is trained on 45.3B tokens across Kazakh, English, Russian, and Turkish. With 8 billion parameters, it demonstrates strong knowledge and reasoning abilities in Kazakh, significantly outper-forming existing open Kazakh and multilingual models of similar scale while achieving competitive performance in English. To ensure effective and responsible alignment, we leverage translated instruction datasets, a Kazakhstan-specific instruction dataset that is automatically constructed and manually verified, and Kazakh-specific safety data. We release Sherkala-Chat (8B) as an open-weight model, along with a detailed description of its training, alignment, and evaluation, to support research and real-world applications for Kazakh speakers.

cs.CL↗

arXiv2025

BLAZER: Bootstrapping LLM-based Manipulation Agents with Zero-Shot Data Generation

Scaling data and models has played a pivotal role in the remarkable progress of computer vision and language. Inspired by these domains, recent efforts in robotics have similarly focused on scaling both data and model size to develop more generalizable and robust policies. However, unlike vision and language, robotics lacks access to internet-scale demonstrations across diverse robotic tasks and environments. As a result, the scale of existing datasets typically suffers from the need for manual data collection and curation. To address this problem, here we propose BLAZER, a framework that learns manipulation policies from automatically generated training data. We build on the zero-shot capabilities of LLM planners and automatically generate demonstrations for diverse manipulation tasks in simulation. Successful examples are then used to finetune an LLM and to improve its planning capabilities without human supervision. Notably, while BLAZER training requires access to the simulator's state, we demonstrate direct transfer of acquired skills to sensor-based manipulation. Through extensive experiments, we show BLAZER to significantly improve zero-shot manipulation in both simulated and real environments. Moreover, BLAZER improves on tasks outside of its training pool and enables downscaling of LLM models. Our code and data will be made publicly available on the project page.

cs.RO↗