Data Curation & Synthetic Data - Weekly Roundup

A staggering 79.4% of LLM-generated clinical content is redundant, challenging the assumption that volume equates to information richness.

SurrogateShield not only prevents PII leakage but also enhances the semantic quality of LLM responses, outperforming conventional redaction methods.

LightOnOCR-2-1B sets a new benchmark for Sinhala OCR, achieving a remarkable 1.05% CER while outperforming both open-source and commercial alternatives.

The curated deforking map reveals 5.41% of fork families that GitHub's platform graph completely overlooks, highlighting the limitations of traditional repository metrics.

MoRE achieves a staggering 44 percentage point increase in deployment success rates by seamlessly integrating behavior mode redirection into policy weights, eliminating the need for inference-time adjustments.

A staggering 79.4% of LLM-generated clinical content is redundant, challenging the assumption that volume equates to information richness.

SurrogateShield not only prevents PII leakage but also enhances the semantic quality of LLM responses, outperforming conventional redaction methods.

LightOnOCR-2-1B sets a new benchmark for Sinhala OCR, achieving a remarkable 1.05% CER while outperforming both open-source and commercial alternatives.

The curated deforking map reveals 5.41% of fork families that GitHub's platform graph completely overlooks, highlighting the limitations of traditional repository metrics.

MoRE achieves a staggering 44 percentage point increase in deployment success rates by seamlessly integrating behavior mode redirection into policy weights, eliminating the need for inference-time adjustments.

Touch is not just an add-on; it fundamentally enhances object representation, leading to dramatic improvements in physical property estimation and manipulation tasks.

RAHA achieves superior cross-modal retrieval performance by leveraging hyperbolic geometry to better capture the low-dimensional semantics of image-text pairs.

Achieving 94.2% precision in item knowledge production at an unprecedented scale, Oxygen AIIC transforms e-commerce item management.

Data mixing, especially with instruction-heavy data, emerges as the crucial factor for optimizing VLM training, challenging traditional filtering approaches.

Achieving R² scores over 0.99 for damage sizing, this framework redefines the potential of deep learning in structural health monitoring with minimal experimental data.

CIRCLE achieves state-of-the-art performance in cold-start continual learning without ever fitting a backbone to image data, revolutionizing the approach to class-incremental learning.

Achieving state-of-the-art fiber bundle segmentation with three times less manual annotation by harnessing synthetic data from dMRI tractography could revolutionize histological analysis.

Directional inconsistency can destabilize LLM training, but geoalign curates rollouts to improve performance and stability, outperforming several established methods.

A novel pipeline generates realistic synthetic clinical notes, enabling safe AI development in healthcare without real patient data.

Over 10% of LLM agent configurations are exact duplicates across repositories, highlighting a critical lack of management in coding environments.

Fine-tuning a TTS model with LoRA can boost Khmer speech synthesis quality significantly, while revealing that adaptation may not benefit already well-supported languages like Korean.

PRISM reveals that a two-dimensional representation of PE files can match the detection performance of larger, traditional models while drastically reducing dimensionality.

ATGBuilder transforms how we construct Activity Transition Graphs by integrating LLM-generated UI summaries and widget-trigger data, achieving superior accuracy in navigation modeling.

SmellNet-V reveals that olfactory identities can be effectively paired with visual data, leading to a 7% improvement in smell classification accuracy.

M2C transforms SAM3 into an auto-promptable annotator that achieves state-of-the-art few-shot segmentation with minimal expert intervention.

BEACON's performance in entity matching reveals significant sensitivity to data constraints, challenging assumptions about its robustness in low-resource settings.

Identifying challenging reasoning examples using just the first 100 tokens can drastically reduce data curation costs while enhancing model performance.

GGR transforms the landscape of open-set semi-supervised learning by ensuring that auxiliary gradients enhance rather than conflict with supervised updates.

KPN achieves superior fault detection accuracy and stability in CCGTs, even with limited labeled data, outpacing conventional few-shot learning approaches.

Synthetic 4D cardiac MRI sequences can boost segmentation performance by over 1.4%, showcasing a powerful augmentation strategy for training robust AI models in medical imaging.

NebulaExp reveals that a meticulously curated dataset and innovative reinforcement learning strategies can boost LLM performance significantly, achieving up to 4.43 points improvement in instruction-following tasks with minimal data.

Distinguishing negative samples can boost LLM reasoning performance on ARC-like tasks by providing critical near-miss alternatives.

Cleaning logs can drastically enhance the performance of model inference and anomaly detection by eliminating irrelevant noise, leading to more accurate insights from software systems.

Reproducibility in quantum software datasets can drop dramatically, with 93.6% of failures tied to dependencies that demand code changes rather than simple version adjustments.

PhysRAG not only elevates visual fidelity in video generation but also ensures compliance with physical laws, setting a new benchmark for realism in AI-generated content.

DeCoFlow achieves a remarkable 98.40% AUROC in continual anomaly detection while ensuring zero parameter forgetting, revolutionizing how we handle sequential data in industrial applications.

The ABC framework empowers researchers with the largest open-source teleoperation dataset and a complete toolkit to accelerate advancements in behavior cloning for robotic manipulation.

Annotation costs can be drastically reduced by shifting from manual labeling to correcting automated hypotheses in speaker diarization tasks.

Manipulative betting ads on social media are not just misleading; they can significantly harm users' mental well-being, and our new dataset equips researchers to combat this issue effectively.

Meta-optimizing an AI data scientist can dramatically enhance the quality of synthetic datasets, outperforming traditional methods.

Synthetic data augmentation can actually worsen performance in well-specified models, but it becomes a powerful tool for correcting ranking errors in misspecified settings.

FedReLa achieves significant accuracy gains for minority classes in federated learning without requiring knowledge of global class distributions.

Small models can achieve competitive performance with innovative data generation and distillation techniques, challenging the notion that bigger is always better in model training.

OncoSynth slashes treatment effect estimation errors by up to 66% in oncology, transforming how synthetic data can inform precision medicine.

Generating realistic patient embeddings can yield performance on par with full datasets, even in scenarios with missing classes.

Forecasting cellular load can be drastically improved by incorporating population dynamics and mobility data, leading to a 60% increase in prediction accuracy.

Efficiently removing undesirable concepts from image generations without sacrificing quality could redefine how we manage biases in generative models.

Curation of training data for brevity can yield a staggering 35x improvement in inference efficiency without sacrificing accuracy in VLMs.

Explicitly marking regions of interest in brain MRIs leads to a staggering 37.54% improvement in anomaly detection accuracy, transforming how diagnoses are made.

EchoStyle achieves high-quality video stylization without the content leakage and style drift that plague existing methods, even outperforming many proprietary systems.

Bridging the gap between informal and formal mathematics, TheoremGraph reveals 18.3 million dependencies that can enhance mathematical search and reasoning.

ALDM achieves unprecedented MRI synthesis quality in few-shot settings, outperforming traditional methods and setting a new benchmark for clinical data augmentation.

LLMs overwhelmingly rely on third-party sources for brand information, with Wikipedia dominating citations across most languages.

Data resources are not just tools; they are the driving force behind the evolution of research methods in Library and Information Science.

LLMs fail to generate reliable requirements from code, revealing critical limitations in their current capabilities for empirical Requirements Engineering.

Pixel-level quality assessment can significantly enhance the reliability of fundus image evaluations, with the new EFIQA-CP method leading the way in explainability and performance.

Label-preserving self-saliency mixup can significantly boost model performance while maintaining semantic integrity in data augmentation.

The largest surgical video-language dataset to date not only enhances surgical video understanding but also sets new benchmarks for AI-driven surgical reasoning.

The variance fields in synthetic stereo data reveal a hidden correlation that could drastically enhance disparity estimation performance in neural networks.

Vision-language models can achieve over 96% accuracy in identifying dangerous actions by children, far surpassing traditional deep learning methods.

LEVIRDet-159 not only sets a new standard in dataset scale but also enables a foundation model that generalizes effectively across diverse remote sensing contexts.

Achieving state-of-the-art image synthesis detection accuracy with a minimalist method that runs efficiently on low-end devices could revolutionize how we combat image-based deception.

Synthetic personas generated by PROFILEFOUNDRY provide a robust framework for evaluating LLMs while ensuring privacy and temporal consistency.

Data-informed agents achieve flawless decision-making accuracy, showcasing the transformative potential of structured data exchange in autonomous ecosystems.

Data Asset Management Capability is the linchpin that drives both standard conformity and benefit realization, revealing a critical chain mechanism for data quality enhancement.

Explicit health data leakage is prevalent in fertility tracking apps, but some manage to monetize without compromising user privacy.

Over 99% of malicious Go module versions remain retrievable via proxies long after GitHub has taken them down, exposing a critical gap in supply chain security.

Orchestrating imperfect schema converters can yield usable results in 72% of real-world data model conversion tasks, exposing critical gaps in the converter ecosystem.

High-end 3D printers can significantly improve the consistency of tactile sensor readings, but cross-printer generalization remains a hurdle.

Lacuna outperforms existing tools in literature retrieval and citation tasks, setting a new standard for research mapping in machine learning.

Cost-efficient relevance evaluation can be achieved without sacrificing accuracy, thanks to a novel calibrated model cascade that halves compute costs while enhancing annotation quality.

PurPL's typestate system ensures that sensitive data is only used for its intended purposes, dynamically adapting to changes in compliance requirements.

Achieving 92.61% accuracy in bearing fault diagnosis with just 10% labeled data reveals a transformative approach to predictive maintenance in industrial settings.

Training data diversity is the secret sauce that boosts agentic model performance, with OpenThoughts-Agent achieving a notable accuracy leap over existing benchmarks.

QC-SMOTE outperforms traditional methods by ensuring synthetic samples are generated with high reliability, adapting to the local data structure to avoid noise and overlap pitfalls.

Partial data augmentation can match the statistical benefits of full augmentation, challenging the notion that complete symmetry is necessary for optimal learning.

A standardized dataset of over 9,000 high-voltage fault simulations could revolutionize the benchmarking of power system protection methods.

FlowPipe achieves a remarkable 11.96% accuracy improvement and 12.5x faster training convergence for data preparation pipelines by leveraging LLMs and advanced flow generation techniques.

Natural identifiers can transform LLM privacy audits by eliminating the need for retraining and inaccessible datasets, making post-hoc assessments practical and scalable.

Unlearning in federated systems can now be achieved in seconds without sacrificing model integrity, thanks to a novel filter-based approach.

Achieving high accuracy in FinFET modeling with minimal training data could revolutionize device characterization and circuit simulation.

Long-tailed distributions in federated graph learning can be effectively tackled with a dual decoupling approach that boosts minority node performance without compromising majority class accuracy.

Optimal batch selection for entity resolution is NP-hard, yet this study reveals a practical solution that significantly enhances recall while controlling costs.

CANDLE reduces Arabic text noise with a 5.37% error rate while slashing inference costs by optimizing tokenizer efficiency.

Transforming noisy web text into a high-quality petroleum engineering dataset boosts retrieval performance by over 44% on benchmark tasks.

WATER-S is a game-changing dataset that boosts WordArt recognition capabilities by hundreds of times, enabling unprecedented accuracy in complex text layouts.

Achieving high-fidelity TEM image synthesis with as few as 15 samples could revolutionize data availability in semiconductor metrology.

GradAudit uncovers that advanced VLLMs can inadvertently expose up to 40% more unauthorized training data than previously detected by existing methods.

Training on SignNet-1M boosts sign language model robustness by improving generalization across diverse real-world conditions without sacrificing performance.

Spurious correlations in foundation models can be effectively disentangled using a dual-branch approach, achieving superior bias mitigation with minimal parameter adjustments.

Freezing decoder layers can significantly enhance the transferability of HTR models across different historical datasets, challenging conventional wisdom about layer freezing in fine-tuning.

The Bengal-HP_RU dataset reveals the critical gap in head pose research, showcasing the rich diversity of Bengali individuals that has been largely overlooked.

Minimizing interpolation during data generation can significantly boost the effectiveness of imitation learning, yielding higher success rates in complex manipulation tasks.

A two-stage framework for mispronunciation detection in low-resource Arabic achieves a groundbreaking F1-score of 0.7201, outperforming previous methods by over 63%.

HDS achieves 44% faster training iterations while improving model performance, redefining efficiency in LLM pre-training.

Real-time cloth manipulation success rates soar as robots learn to refine actions using a simulator-in-the-loop approach, outperforming traditional methods.

Unbalanced Optimal Transport enables accurate tree counting from noisy satellite data, outperforming traditional methods in dense forest environments.

Burnyard transforms malware analysis by enabling efficient, secure emulation that avoids the pitfalls of traditional sandboxing.

POLAR achieves superior performance in adaptive data acquisition by harnessing pretrained models, requiring dramatically fewer training samples than existing methods.

Structuring Arabic-English dictionaries for machine readability could revolutionize NLP applications in Arabic language processing.

Quality-aware data selection can significantly enhance the performance of long-document summarization models, outperforming random sampling even at matched training sizes.

A new gold-standard Marathi POS tagging dataset reveals that even under-resourced languages can achieve high accuracy with the right tools and methodologies.

Decomposing annotation tasks can significantly reduce the cognitive burden on annotators, leading to better quality outputs at lower costs.

Robust representations of ancient Greek letterforms reveal intricate stylistic evolutions and inter-letter relationships that traditional methods fail to capture.

VieSpeaker demonstrates that you can build a robust speaker recognition dataset without relying on visual cues, unlocking new possibilities for under-resourced languages.