PAPER PLAINE

Many real-world recognition tasks involve categories the system has never encountered—new medical diagnoses, emerging languages, novel products. The bias toward familiar categories can cause serious misidentifications. This fix is modular and works with existing systems without redesigning them, making it practical to deploy across medical imaging, document processing, and other fields where recognizing truly new things matters.

Variational autoencoders are already deployed in energy grids, hospitals, and financial systems to forecast and detect anomalies in time-dependent data. Without theoretical guarantees, there's no principled way to know when these models are reliably learning patterns versus just memorizing training data. This work provides that foundation, letting practitioners understand when these models are safe to trust on new real-world sequences.

Language models power chatbots, search engines, and writing tools that real users interact with through imperfect inputs—voice transcription errors, typos, garbled PDFs. This work reveals that a model might produce reasonable-sounding output while its internal workings are partially corrupted, masking fragility that only shows up under specific stresses. Understanding these hidden failure modes is essential for deploying language models in safety-critical contexts and for building systems that degrade gracefully rather than confidently producing wrong answers.

This could let researchers quickly identify what data a neural network has seen, which matters for detecting model theft, understanding training data composition, and auditing AI systems for bias or unwanted memorization. It also suggests a simpler way to compare and categorize trained models based on their internals alone.

Machine learning teams routinely use synthetic data to fix imbalanced datasets, but have lacked clear guidance on which generation methods work best. This framework provides a principled way to evaluate synthetic data quality before deploying it, potentially saving teams from investing time in high-fidelity generators that won't improve actual performance. It also shows that simpler oversampling methods are sometimes more reliable than complex generative models—a finding that could reshape how practitioners choose their augmentation strategy.

Synthetic data is cheap and fast to generate, but mistakes in using it waste time and money on flawed conclusions. This framework lets practitioners know precisely when they can trust their results, making it practical to combine real and synthetic data without guessing about reliability. For surveys, medical studies, and other research constrained by small sample sizes, this could make the difference between usable insights and misleading ones.

Distributional reinforcement learning helps AI agents make better decisions in uncertain environments by understanding the full range of what might happen, not just the typical outcome. This proof provides the first formal guarantee that the method works reliably, which gives engineers confidence to use it in real systems where data is expensive or risky to collect—like robotics or medical applications.

Indigenous languages like Baniwa are disappearing as fewer people speak them, and digital tools almost never exist for them because companies focus on major languages. This work shows that endangered language communities don't need massive datasets to build functional speech recognition tools—they can leverage existing AI systems trained on major languages. This creates a practical pathway for documenting and preserving languages that otherwise leave no digital record.

Video platforms make billions in ad revenue based on which videos they show you. Better watch-time predictions mean more accurate recommendations, which keeps people engaged longer and generates more advertising impressions. The production test showed statistically significant increases in user engagement, translating directly to platform value.

Policy makers often need to predict outcomes like employment, pollution, or disease spread across regions and years, but existing methods force them to choose between oversimplified models or treating each location and outcome separately. This framework lets analysts automatically discover which connections actually exist in their data, making forecasts more accurate and revealing hidden regional interdependencies that single-location models would miss.

Probability estimation is central to compression, language modeling, and machine learning. This method works reliably across different alphabet sizes and data quantities without needing to adjust knobs by hand—a major practical advantage. The work also reveals a precise mathematical pattern: prediction cost scales directly with how fast new items appear in your data, which offers clearer insight into what makes some estimation problems harder than others.

Companies increasingly automate invoice sorting for tax compliance and financial reporting, but relying on large cloud-based AI creates cost and security risks. This work shows that smaller, in-house models can match or beat expensive alternatives while keeping sensitive financial data private—and only need about 100 labeled invoices per new client to work well. Understanding *how* these models make decisions through geometry rather than black-box predictions also lets accountants spot when something has gone wrong.

Spotting fraudulent transactions, network intrusions, or equipment failures in databases matters enormously for banks, security teams, and manufacturers. Current tools rely almost entirely on one detection approach, leaving blind spots. This work shows that dusting off an older technique and combining it with modern methods fills those gaps, making anomaly detection measurably more reliable without needing to retrain existing systems.

Many real-world prediction tasks face uncertain conditions where you don't know upfront whether your available tools will work well. This algorithm handles both cases simultaneously, making it practical for financial forecasting, weather prediction, and fraud detection systems that can't afford to fail or slow down when assumptions break down.

This result clarifies a fundamental question about quantum computing's advantage in machine learning. If quantum data sources become available, quantum computers could provide speedups that aren't possible with purely classical training data—establishing a new dimension of quantum advantage beyond just computational power. This helps researchers understand where quantum machine learning will matter most in practice.

Financial errors move fast and cost real money. A trading system that silently acts on yesterday's permissions or misread instructions can execute unauthorized trades before anyone notices. SAGE-Fin forces the system to prove at runtime that each proposed action is still authorized, catching drift and stale data before execution. The platform's independent operations team found it practical enough to deploy on live customer requests, suggesting it can reduce the window where bad trades slip through.

VARMA models are theoretically superior to simpler alternatives because they capture complex patterns with fewer parameters, but companies and researchers have avoided them for anything beyond small datasets due to prohibitive computational cost. This breakthrough removes that barrier, letting practitioners use better forecasts for demand planning, weather modeling, and air quality without switching to inferior approximations or waiting for impractical computation times.

Topological diagrams are used to find meaningful patterns in complex datasets, from brain imaging to materials science. Currently, researchers treat these diagrams as fixed snapshots. This framework enables diagrams to evolve and adapt, opening the door to cleaner data summaries that retain what scientists actually care about—reducing noise without losing signal. This could make it easier to compare, compress, and understand high-dimensional data across fields.

Generative flow networks are increasingly used to solve hard combinatorial problems in drug discovery, chip design, and optimization. This geometric approach makes training faster and more reliable by letting the algorithm adapt to the underlying problem structure rather than using one-size-fits-all updates, potentially cutting the computational cost of finding good solutions substantially.

Many high-stakes decisions—hiring, healthcare, academic grading—rely on assessment where there's no perfect right answer to check against. This method lets organizations systematically improve their judgment calls by combining what humans do well (comparing options) with what AI does well (spotting patterns). The theoretical guarantees mean practitioners can trust the approach even when human agreement is messy or incomplete.

Hospital records are inherently messy: labs are ordered when doctors suspect problems, vital signs stream continuously but get recorded sporadically, and timing itself contains medical information. Current methods either throw away the timing and detail or ignore bias in what gets measured when. This approach handles all of that without requiring researchers to manually engineer summaries, making it possible to reliably answer causal questions from real electronic health records—which matters because many treatment decisions in medicine depend on understanding what actually works, not just what correlates with recovery.

Many real-world problems—genomics, climate modeling, financial forecasting—involve hundreds or thousands of variables, making traditional statistical methods impractical. BAND makes these high-dimensional estimation problems computationally feasible, enabling faster and more reliable probability forecasts in domains where previous methods would simply take too long to produce usable results.

Reservoir computing is already used for real-time forecasting in weather, finance, and engineering systems because it trains faster than deep neural networks. This frequency-based design makes it work better while also becoming interpretable—you can actually understand why it makes specific predictions—rather than operating as a black box. That combination of speed, accuracy, and explainability matters for applications where you need both reliable forecasts and the ability to trust why the system made them.

Quantum computers are still in early stages, and researchers have worried that scaling up quantum machine learning models would hit the same wall as classical models: they'd memorize training data and fail on anything new. This work suggests that quantum circuits might naturally avoid that trap, making it more plausible to build larger, more capable quantum learning systems. It doesn't solve quantum machine learning yet, but it removes one major barrier to scaling up.

Modern statistics often combines traditional techniques with machine learning, but most existing methods assume more about the data structure than they actually need—creating a drag on accuracy. This work tightens that requirement, directly improving the precision of estimates in fields ranging from medical research to economics where pinpointing the true effect of a treatment or variable is critical.

Running an AI chatbot for 38 million customers costs enormous amounts in computing power and causes frustrating delays. This method slashed those costs in half and made responses 50 times faster while keeping personalization intact—enabling companies to actually deploy AI services that would otherwise be too expensive to run at scale. The technique works because it mathematically proves each user gets grouped with genuinely similar requests, not arbitrary bucketing.

This method addresses a real problem in fields like medicine and policy: you need to decide what treatment or action to take for a specific person or situation, but you can't run experiments on everyone. By finding similar past cases and learning from their outcomes, the approach lets organizations make better decisions faster without waiting for new experimental data. It's particularly useful when past records are plentiful but noisy—common in healthcare, criminal justice, and business operations.

Most visualization tools for categorical data either collapse it into simplified charts that lose information or produce abstract plots that disconnect results from the original data—making it hard to trace why a pattern appeared and whether it's real. cGAP solves this by keeping the raw data visible while layering interpretable geometric structure on top, letting researchers across genetics, biomedicine, and social science spot meaningful patterns without sacrificing transparency or accuracy.

Scientists and engineers building machine learning systems need to quantify uncertainty—to know when predictions are reliable and when they're guesses. This work provides a principled way to choose which uncertainty measures to use: instead of debating which is 'correct,' practitioners can now derive the right uncertainty decomposition directly from their loss function and the specific problem they're solving. That foundation makes uncertainty quantification less arbitrary and more tied to the actual goal of the model.

Earthquake risk assessments that guide building codes and insurance depend on accurate, honest estimates of what we don't know about subsurface conditions. This work lets engineers use practical testing setups (multiple-sized arrays) while still rigorously quantifying uncertainty — rather than forcing a choice between depth of measurement or confidence in results. The method produces engineering metrics like site period that are more stable than the raw measurements, making them more useful for actual structural design decisions.

This reveals a critical vulnerability in how AI systems are trained and deployed. If someone controls the training process, they can sabotage a model in ways that look completely legitimate to any auditor or security check. This creates a significant asymmetry in power between those who train AI systems and those who use them — trustworthiness becomes nearly impossible to verify mathematically, even with full access to the model's internals.

Backpropagation's gradient calculations become unreliable in very deep networks, limiting how large and capable these systems can be. This gradient-free alternative sidesteps that problem entirely and works without requiring architectural tricks like batch normalization that researchers currently use as workarounds. If practical, it could unlock simpler ways to train neural networks and offer new insights into how these systems learn.

Training procedures that wander through unsafe regions — even briefly — can cause real failures in deployed systems, from autonomous vehicles misclassifying obstacles to medical AI producing dangerous predictions. This work provides concrete mathematical guarantees about when and how badly a model can stray during training, letting practitioners either trust the process or redesign it to avoid specific high-risk zones altogether.

AI systems currently struggle when faced with hard problems outside their normal experience. TREK fixes this by letting models learn from external teachers—whether a more powerful AI, a human-created solution, or the model itself with extra context—then polish those solutions through standard training. This means better reasoning AI without requiring expensive human feedback or labeled training data, directly improving performance on math competitions and complex real-world tasks.

Many real datasets have built-in boundaries: household income surveys that don't record above a threshold, medical tests with detection limits, or equipment readings that plateau. This method lets analysts make reliable predictions from such data without either throwing information away or pretending the boundaries don't exist. It's particularly useful when the relationship between data source and real-world conditions has shifted, a common problem when applying models trained on one dataset to different populations.

Physics simulations are computationally expensive—whether for engineering, climate modeling, or drug discovery. These more stable neural networks could run orders of magnitude faster while remaining reliable, enabling scientists to explore more scenarios and design variations in the time it currently takes to run a single full simulation.

As AI systems are increasingly deployed online with learned reward models, this finding reshapes how teams should set up "safe" training. Instead of maximizing conservatism, practitioners need to find a calibrated middle ground — one that maintains alignment without accidentally creating vulnerability. Getting this balance wrong could undermine safety efforts across reasoning tasks where online learning is used.

Generative AI models power image synthesis, drug discovery, and scientific simulation. Current methods introduce systematic errors precisely where accuracy matters most: in the tails of probability distributions where rare but important outcomes live. This framework cuts estimation error in those critical regions, potentially improving the reliability of AI-generated samples across applications from medicine to materials science.

For scientists working with specialized sequences in biology and chemistry, this cuts the cost of selecting training data from hours of computation to seconds. The technique also reveals whether models learn from data patterns, prediction errors, or how those interact—giving researchers insight into what makes transfer learning succeed or fail in their domain.

Companies analyzing thousands of reviews need to know not just what rating to expect, but why customers gave it. This method delivers both at once, letting product teams spot the actual language patterns driving customer satisfaction rather than treating prediction and understanding as separate problems that require different tools.

Networks with group relationships (like email threads with multiple participants or chemical reactions involving many atoms) are common but difficult to analyze. This technique makes it practical to feed these complex structures directly into machine-learning systems, which could improve applications ranging from recommendation engines to molecular modeling without requiring researchers to manually decide how to combine information from different relationship types.

Healthcare systems and marketing platforms constantly evaluate whether new treatment or customer strategies would work better than current ones, but incomplete record-keeping undermines these tests. This method makes it possible to learn from flawed historical data without bias, meaning hospitals could confidently test new care protocols and companies could validate strategy changes using the messy real-world data they actually have.

Conditional independence testing underpins decisions in machine learning, fairness auditing, and causal inference. When these tests give false positives—declaring variables independent when they're not—they can lead to flawed models and unfair automated decisions. This method works reliably even when the assumed model has small errors, which is almost always the case in practice, making it directly usable in real applications rather than just theoretical settings.

Identifying true causes rather than just correlations is essential in fields from medicine to economics, where treating a symptom won't help if you don't know what causes it. TSCD's ability to work with fewer experiments saves time and resources, while its efficiency means it can handle systems with hundreds of variables—making it practical for real-world problems like understanding gene networks or economic supply chains.

Federated learning powers real-world systems like predictive keyboards, health apps, and industrial sensors that must learn from private data without sending it to a central server. Better generalization means these systems work reliably when deployed to new users or environments, rather than degrading because they memorized quirks of their training group. This directly improves the practical performance of privacy-preserving AI across smartphones, hospitals, and distributed networks.

Monitoring insect populations by listening to their natural sounds could replace labor-intensive manual surveys, making it cheaper and faster to track how ecosystems are changing. The system works with unlabeled field recordings, which means researchers don't need expensive expert annotation of every audio clip, making large-scale ecological monitoring practically feasible for conservation programs.

Protein design is expensive and time-consuming, requiring many lab experiments to find variants with desired properties. This method cuts the amount of experimental data needed, potentially accelerating the discovery of proteins for drugs, industrial enzymes, and other applications. It's especially valuable when screening many related proteins at once, letting researchers predict behavior across multiple properties simultaneously rather than testing each one separately.

Solar operators currently make blind decisions about maintenance, storage, and grid commitments at a plant's launch. Better cold-start forecasts let them optimize operations immediately rather than waiting months for real data to accumulate, reducing waste and improving grid reliability. The method works across different climates and plant types, making it practical for rapid deployment worldwide.

Most modern statistical analyses rely on machine learning to handle complex nuisance tasks, from estimating treatment effects in medicine to calculating causal impacts in policy. This work shows how to extract more reliable answers from the same amount of data—without requiring stronger assumptions or running more experiments. For practitioners, it means sharper confidence intervals and more trustworthy conclusions when combining flexible machine learning with rigorous statistical inference.

Machine learning engineers spend significant time tuning hyperparameters and choosing algorithms without knowing in advance whether their choices will lead to fast or slow solutions. This framework lets them estimate convergence speed from a simple formula, cutting down trial-and-error and making it easier to decide whether a particular configuration is worth pursuing before investing computational resources in training.

When a medical AI says it's 95% confident in a diagnosis, that confidence needs to mean something. Poorly calibrated models mislead doctors, lenders, and regulators about how much they can trust a decision. This benchmark provides a standardized way for practitioners to pick the right fix for their specific problem, and gives researchers a shared testing ground so better methods don't get lost among dozens of competing approaches.

Diffusion models power today's most capable image and text generators, but engineers have been working largely in the dark about why they're so statistically efficient. This theoretical proof validates the practical intuition that these models naturally exploit hidden structure in real data—like the fact that natural images, despite having millions of pixels, lie on much lower-dimensional manifolds. It means companies building generative AI can trust that the approach is fundamentally sound, not just empirically lucky.

Ranking systems appear everywhere—search engines rank web pages, platforms rank sellers or content, sports leagues rank teams. If someone can deliberately skew which comparisons happen more often, they can artificially boost their own ranking. This work provides a practical fix that prevents such manipulation without needing to know in advance which comparisons an adversary will target.

Fine-tuning is one of the fastest ways to adapt AI models to specific tasks, but practitioners have long assumed that pushing training loss too low causes the model to overfit and fail. This work shows that apparent overfitting can actually improve real-world output quality, challenging a core assumption in how models are trained and opening a path to better performance with minimal computational cost.

Earthquake early-warning systems guide emergency response and evacuation decisions. Better forecasts of which regions will experience intense clustering in a given week could help authorities pre-position resources and issue more reliable alerts. The model's strongest gains come in predicting extreme weeks (5+ earthquakes), exactly when forecasts are hardest to make and most consequential for public safety.

Popular optimizers like Adam and AdamW are built on AdaGrad's principles, so understanding why AdaGrad works under chaotic, noisy training conditions explains why these widely-used tools perform reliably in practice. This closes a gap between theory and practice: machine learning practitioners have long observed these algorithms working well on messy real-world tasks, but the math didn't fully explain why until now.

When doctors, banks, or regulators rely on machine learning to make decisions, they need to know which factors the model actually used—not just a ranking that changes every time the model is retrained. RoSHAP makes those explanations stable and trustworthy. The method also lets companies use fewer data inputs while keeping the same prediction accuracy, reducing complexity without sacrificing performance.

Hospitals need accurate timelines to predict which patients are deteriorating—crucial for conditions like sepsis where hours matter. Current systems force doctors to choose between rich but fuzzy narratives or precise but gappy data tables. This method uses both, meaning clinicians get both the full picture of what happened to a patient and the exact timing of when it happened, improving risk prediction and care decisions.

AI-as-a-judge systems are increasingly used to automatically grade student work, evaluate code, and validate outputs in production systems. Making these systems smarter about when to engage expensive reasoning directly cuts computational waste while maintaining accuracy—crucial for companies running these evaluations at scale where every percentage point of wasted compute multiplies across millions of judgments.

Many fields rely on proxy measurements for speed—clinical trials using biomarkers instead of patient outcomes, industrial testing using sensor readings instead of final quality checks. Current methods fail when conditions drift between experiments. This adjustment works on top of existing corrections and requires only summary-level historical data, making it practical to implement across domains while reducing the risk of biased conclusions.

Attention sinks waste computational resources and can degrade model performance by forcing the network to concentrate on irrelevant tokens. Understanding the root cause opens the door to cleaner, more efficient models—the architectural tweak the researchers tested could reduce training time and improve how language models process information, with potential benefits for speed and accuracy in real applications.

Sampling from complex probability distributions is central to machine learning, physics simulations, and Bayesian statistics. Current methods either require extensive training or many expensive probability evaluations. This hybrid approach cuts the computational cost of generating high-quality samples, which directly speeds up inference in scientific computing, drug discovery, and probabilistic machine learning models where every probability calculation is expensive.

Surveys and tests that adapt their questions based on previous answers can extract more reliable information while asking fewer questions—cutting costs and reducing respondent fatigue. This method makes adaptive surveying practical for real applications like market research, psychological assessment, and opinion polling, especially when you're starting fresh with a new population and can't rely on historical data. The approach also produces interpretable results: you learn not just what someone thinks, but which persona type they resemble, offering actionable insights alongside raw answers.

In fields like medicine, finance, and environmental monitoring, obtaining ground-truth labels is costly or time-consuming. This framework lets organizations leverage multiple off-the-shelf AI models they already have, extracting more reliable statistical conclusions from the labeled data they can afford to collect. The guaranteed best-expert performance means the approach never does worse than just using a single good predictor.

Current machine learning practice forces a choice: either waste 10–20% of your data on a validation set to estimate real performance, or train blindly and risk deploying an overfit model. This algorithm could eliminate that trade-off, letting practitioners use all their data while still getting reliable estimates of how their model will perform in the real world. The method was tested on image classification tasks and consistently narrowed the gap between training and test performance compared to standard training approaches.

Training machine learning models is expensive in both time and computational energy. This approach cuts training time dramatically—by 23× on large text tasks—without sacrificing accuracy. It also handles messy real-world data better: when labels contain errors, the method outperforms standard approaches by 2.6% on standard benchmarks, making it immediately useful for practitioners working with imperfect datasets.

As AI systems become more powerful and are deployed for high-stakes decisions, knowing whether alignment methods actually work is critical. This work provides a way to verify that an AI system trained to follow human preferences will genuinely do so, rather than discovering failures after deployment. The new method is especially useful for handling tricky cases where multiple different responses are equally correct—a common problem in real-world AI alignment.