Search papers, labs, and topics across Lattice.
100 papers published across 3 labs.
Arabic LLMs can speak the language of finance, but they often fail to reason about it, especially when it comes to causality and generation.
Forget painstakingly collecting real CAD data – Zero-to-CAD lets you bootstrap CAD program generation from multi-view images using a million-scale dataset synthesized entirely by an LLM agent.
A BiLSTM with a custom slang dictionary rivals AutoML in classifying the sentiment and emotion of messy, real-world Indonesian e-commerce reviews.
Training on semantically equivalent chart renderings in Python, R, and LaTeX unlocks surprisingly effective multi-lingual chart-to-code generation from a single model.
Forget painstakingly curating datasets – STELLAR-E auto-generates high-quality, domain-specific LLM benchmarks, rivaling real-world data in evaluation quality.
Forget painstakingly collecting real CAD data – Zero-to-CAD lets you bootstrap CAD program generation from multi-view images using a million-scale dataset synthesized entirely by an LLM agent.
A BiLSTM with a custom slang dictionary rivals AutoML in classifying the sentiment and emotion of messy, real-world Indonesian e-commerce reviews.
Training on semantically equivalent chart renderings in Python, R, and LaTeX unlocks surprisingly effective multi-lingual chart-to-code generation from a single model.
Forget painstakingly curating datasets – STELLAR-E auto-generates high-quality, domain-specific LLM benchmarks, rivaling real-world data in evaluation quality.
Even the largest language models still struggle to connect information across dispersed code segments, achieving only 74% accuracy on a new benchmark designed to test multi-hop code comprehension.
Finally, a dataset exists to train and benchmark algorithms for automatically detecting airway bifurcations in 3D CT scans, a crucial step towards understanding respiratory diseases.
Low-cost stereo vision can rival LiDAR for real-time windrow detection, paving the way for more accessible autonomous farming solutions.
Robots can now leverage human intuition for manipulation tasks, learning from a massive video dataset to improve motion plausibility and robustness, even when conditions change.
Simulate once, deploy anywhere: SPLIT lets you train tactile perception models on synthetic data and transfer them across different sensors without retraining.
Open-source diffusion models can now achieve state-of-the-art illumination control rivaling closed-source alternatives, thanks to a novel training pipeline and dataset.
LLMs can be systematically debugged and improved by treating training data as code, allowing for targeted "patches" that fix concept-level gaps and reasoning errors.
Unlock the secrets of the deep: OceanPile, a massive, meticulously curated multimodal dataset, finally brings the power of foundation models to the vast and underexplored ocean.
You can boost insurance claim prediction accuracy by combining simple environmental features with location data, even when you lack detailed individual-level spatial information.
Automatically generate data unit tests that actually catch the data errors that matter for your specific downstream tasks.
Ditching noisy SGD trajectories for smooth Bezier curves unlocks better dataset condensation, especially when data is scarce.
Forget memorizing table headers: TaNOS unlocks surprisingly robust numerical reasoning by pre-training on operation sketches and correctness-guaranteed programs.
Compact datasets in n-dimensional space can be transformed into linearly separable sets using diffeomorphisms and shallow, wide neural networks, challenging the need for complex architectures in certain classification tasks.
Stop punishing your model for disagreeing with corrupted data – Trust-SSL learns better representations by treating alignment with degraded views as a residual learning problem, not a hard constraint.
LLMs struggle to answer human-generated questions about multi-chart images, highlighting a critical gap in their ability to reason about real-world data visualizations.
Channel-free HAR is now possible: a single model can perform activity recognition across diverse IoT sensor setups without needing fixed channel arrangements, thanks to metadata-conditioned fusion.
Fixing miscalibrated black-box predictions with a simple post-hoc calibration step can significantly boost the accuracy and efficiency of semisupervised mean estimation.
LLMs aren't just Western-centric; they have a peculiar obsession with Japan, and this bias is amplified by English-language prompting.
A new synthetic aerial imagery dataset provides pixel-perfect depth, controlled illumination, and multi-scale imagery, unlocking joint research across geometric understanding, domain robustness, and resolution enhancement.
LLMs' apparent success at program repair crumbles when faced with slightly altered versions of known bugs, revealing a reliance on memorization rather than true understanding.
LLMs' factual knowledge is surprisingly brittle: simply changing an entity's surface form in a question (e.g., using an abbreviation instead of the full name) can drastically alter the answer.
Forget scaling laws – AgenticQwen proves that clever training with dual data flywheels can enable small language models to rival giants in real-world agentic tasks.
Even when you think you've scrubbed 90% of the PII, your anonymized text might still leak two-thirds of a person's identity.
Static analysis tools miss a staggering 87% of real-world Python vulnerabilities when they're introduced across multiple commits, even when the full codebase is available.
Training a video reshooting model on internet-scale monocular videos is now possible, thanks to a clever self-supervision trick that generates multi-view training data from a single video.
Seemingly innocuous choices in loss functions and training regimes can significantly hinder visual geometry estimation, even for state-of-the-art methods.
Frozen vision foundation models can be surprisingly effective at improving out-of-domain object detection by stabilizing relational modeling and semantic-spatial alignment in the detector.
Compressing expansive contexts like a convex mirror allows deep learning models to achieve robust ground filtering across diverse landscapes, even in complex urban scenes.
Annealing-based data augmentation lets you train a YOLOv10 detector to spot more fish in murky underwater images.
Data loading bottlenecks can strangle your GPU utilization down to 10%, but a few smart optimizations can unlock a 6x speedup.
COFs can withstand defects surprisingly well: mechanical properties remain stable even with defects, but thermal conductivity plummets, revealing design trade-offs.
Stop generating text-to-SQL training data that *runs* but is semantically wrong: this new framework finally aligns synthesis with database semantics.
LLMs can rewrite bad job descriptions and category-aware MoEs can better match candidates, leading to a 19.4% boost in recruitment click-through rates and millions saved.
Training on EVENT5Ws allows event extraction models to generalize across geographical contexts, suggesting a path towards truly universal event understanding.
Synthetic data can significantly boost controllable human video generation, but only if you carefully select which synthetic samples to use.
Poisoning attacks got you down? This defense flips the script by using the attacker's own clustering behavior against them, achieving near-perfect attack mitigation with minimal accuracy loss.
By unifying generative and discriminative approaches, UniGenDet achieves superior image generation and detection, suggesting that these tasks benefit from a symbiotic relationship previously hindered by architectural divergence.
Identity encoders can now achieve human-level performance in recognizing stylized faces, bridging the gap between artistic expression and identity consistency.
LLMs can now generate realistic online discussions, opening the door to studying deliberation dynamics at scale without real-world ethical and data access hurdles.
Turns out, coding agents in the wild are only writing useful code 44% of the time, and are introducing more security vulnerabilities than human developers.
Finally, a meta-learning approach that uses readily available negative control samples can close the persistent domain gap in biomedical imaging, making deep learning models practically usable across different experimental batches.
Spectral analysis of client feature representations can identify and relabel noisy data in federated learning, outperforming existing noise-tolerant loss and loss-dynamic approaches.
Systematic coverage gaps in retrieval evaluations can lead to misleading assessments, but semantic stratification offers a clearer, more trustworthy framework for understanding retrieval performance.
Differentially private federated learning gets a boost: PINA achieves 2.9% higher accuracy than state-of-the-art methods by using a novel two-stage approach with privacy-preserving initialization and normality-driven aggregation.
Overcome the scarcity of rare flight diversion events by using optimized generative models to create synthetic data that substantially improves prediction accuracy.
A groundbreaking dataset suite reveals the intricate dynamics of decentralized prediction markets, offering unparalleled insights into collective forecasting behavior.
Tabular anomaly detection gets a serious upgrade: uLEAD-TabPFN leverages frozen PFNs to model complex feature dependencies, outperforming existing methods by a significant margin, especially in high-dimensional spaces.
RL-based sample selection beats traditional active learning for transfer learning when data is scarce and imbalanced, especially in clinical settings.
Multilingual data quality classifiers can outperform monolingual ones, but only with careful tuning of the decision boundary, challenging the assumption that scale alone guarantees improved filtering.
Forget scaling laws: a model trained on a carefully curated subset of visual instruction data can beat models trained on datasets orders of magnitude larger.
BDD suites are drowning in duplicated steps—cukereuse finds that 80% are exact duplicates—and this tool offers a way to automatically clean them up.
LLMs still struggle with basic skills in low-resource languages, even when they excel at reasoning.
Discover expertise and collaborators in battery research at a global scale, grounded in semantic understanding rather than just citations.
Achieve near-perfect (96.35% Dice) maxillary sinus segmentation from X-rays with limited labeled data by distilling knowledge from GAN-refined pseudo-labels.
Querying strategically with energy-based models lets you actively learn in the wild, even when you don't know what you don't know.
Bridging the gap between human manipulation and robotic control, JoyAI-RA unlocks enhanced cross-embodiment behavior learning through multi-source pretraining.
Autonomous driving validation gets a shot in the arm: OVPD offers a proving ground dataset that fuses real vehicle dynamics with controllable virtual environments, enabling more realistic and diagnosable testing.
Key contribution not extracted.
Standardizing sign language data preprocessing with SignDATA enables reproducible research and explicit control over extractor choice, normalization, and privacy.
Synthetic counseling dialogues can be made significantly more realistic and useful for fine-tuning by grounding them in structured Client Psychological Graphs that capture the interplay of a client's thoughts, emotions, and behaviors.
COMPASS outperforms traditional multilingual fine-tuning by effectively leveraging semantic gaps to enhance cross-lingual transfer and model adaptability.
Synthetic video can now drive dexterous robotic manipulation without the need for high-quality 3D demonstrations, enabling zero-shot generalization across diverse tasks.
Domain-specific continual pre-training lets a 7B model punch *way* above its weight, beating a 24B generalist on medical tasks by 3.5x.
HOI video synthesis gets a major realism boost: CoInteract's dual-stream training and region-specialized experts produce interactions that are both structurally stable and physically plausible.
By clustering users based on the geometry of their feature spaces *before* training, FB-NLL sidesteps the vulnerability of existing federated learning methods to noisy labels and corrupted updates.
Active learning's promise of efficient annotation falls flat in chemical reaction extraction, with strong pretraining and structured decoding creating instability.
Arabic LLMs can speak the language of finance, but they often fail to reason about it, especially when it comes to causality and generation.
Forget training a GNN for every new graph: NodePFN learns universal node classification from synthetic graph priors, generalizing across diverse datasets without graph-specific training.
Escaping the curse of noisy data in semi-supervised learning: S$^2$MAM adaptively selects features and tunes similarity metrics, leading to more robust and interpretable models.
TabGRAA flips the script on tabular data synthesis, turning static statistical replication into a dynamic, self-improving generation process.
Small language models can achieve strong performance in specialized scientific domains like quantum field theory with targeted fine-tuning and synthetic data generation.
Generative models for mobility data, previously thought to be private, are vulnerable to membership inference attacks, highlighting the need for more robust privacy evaluations.
Despite growing concerns about data contamination, current black-box methods are essentially useless for detecting if an LLM has been trained on specific copyrighted material.
Noisy multimodal preference datasets are holding back reward model performance, but DT2IT-MRM offers a scalable curation strategy that achieves state-of-the-art results.
LLMs can now reason far better in low-resource domains, thanks to a new method that aligns their thinking with high-resource domains using "reasoning representation alignment."
Synthesizing realistic defect data with diffusion models and Perlin noise can dramatically improve unsupervised anomaly detection, achieving near-perfect AUROC scores on industrial surfaces.
Synchronized aerial imagery unlocks dense, geometrically consistent BEV semantic mapping of dynamic road scenes, even from ego-centric sensors alone.
Edge-deployable VLMs can now achieve surprisingly strong performance in Japanese language and real-world vision tasks, rivaling larger models.
Reasoning, not just text, is the key to detecting AI-generated content: REVEAL leverages interpretable reasoning chains to significantly outperform existing AIGC detectors.
A Bolu unveils the hidden structure within Sardinian improvisational poetry, revealing recurring patterns that challenge our understanding of oral creativity and offer a new dataset for NLP research on minority languages.
Lightweight LLMs can achieve Knowledge Graph construction performance rivaling GPT-3.5, suggesting a path to more efficient and accessible KG creation.
NLI models can be significantly debiased with minimal accuracy loss by simply downweighting examples where biased models exhibit high confidence.
Training LLMs on data detoxified with HSPD slashes toxicity by more than half, outperforming existing methods that only address toxicity during or after training.
NER performance on user-generated content isn't just about noise – it's fundamentally limited by information density, and targeted augmentation can unlock significant gains.
Open vs. closed debates miss the point: AI is fundamentally reshaping the economics of research metadata, creating new risks and opportunities that require careful governance of the space between free data and commercial products.
Achieve state-of-the-art 3D medical image generation by reformulating deterministic prediction as a multi-objective drifting problem, outperforming GANs, flow-matching, and SDEs in fidelity, realism, and efficiency.
Training-free diffusion models can now harmonize satellite imagery across diverse domains, enabling scalable remote-sensing synthesis without retraining.
Simple histogram adjustments can dramatically improve the real-world robustness of plant disease classifiers, especially in uncontrolled lighting conditions.
Forget dataset-specific tuning: IonMorphNet learns generalizable ion image morphologies, boosting peak picking accuracy by 7% across diverse mass spectrometry imaging datasets.
Forget synthetic data – Unposed-to-3D learns to reconstruct realistic, simulation-ready 3D vehicles directly from real-world driving images.
Achieve higher quality 3D Gaussian Splatting reconstructions of objects from mobile phone captures using fewer images, thanks to a novel object-centered capture guidance system.
Generative models can be surprisingly effective for texture filtering when fine-tuned with a two-stage supervised and reinforcement learning approach.
Autonomous vehicles can now see better in the dark and rain thanks to a lightweight module that fuses per-frame perception with a learned global map, boosting geometric understanding and localization.
Decomposing holistic visual cues into subtle, spatially-associated discrepancies allows for state-of-the-art ultra-fine-grained classification even with limited training data.
Soybean leaves have intricate vein structures that unlock state-of-the-art ultra-fine-grained visual categorization, even with limited data.
Discarding disagreement as noise in health-literacy annotation can mask critical social-scientific effects, which only surface when analyzing data stratified by inter-annotator agreement.
