From Machine Learning to Large-Scale EO Products: Good Practices and Guidelines for Making Maps
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "From Machine Learning to Large-Scale EO Products".
Jane: Here's the gist of the paper. Recent years have seen a rapid expansion in the production of large-scale geospatial maps derived from Earth observation (EO) data, driven largely by advances in machine learning (ML) and large computing infrastructure <ref:2607.24532#pg2>.
Tom: Although the barrier to generating such maps has dropped substantially, established best practices have yet to emerge, and design decisions made early in the pipeline can quietly propagate errors into the final product <ref:2607.24532#pg2>.
Jane: Producing a technically sound and scientifically credible product remains challenging <ref:2607.24532#pg2>.
Tom: The EO Data Infrastructure Landscape Researchers must navigate a complex ecosystem of satellite data sources, derived data products, and access platforms <ref:2607.24532#pg2>.
Title and authors: Tom: So we're looking at this paper from arXiv called From Machine Learning to Large-Scale EO Products: Good Practices and Guidelines for Making Maps. It’s basically a guide on how to actually build these massive geospatial maps using machine learning, which is a big deal because the technology lets us create maps at resolutions that were impossible just a few years ago.
Jane: Right, so it talks about how ML has shifted things away from models that are only calibrated locally toward single models that generalize globally across really diverse data. But it immediately points out that this new workflow brings in challenges that aren't just for the machine learning folks anymore.
Lu: The paper emphasizes that the whole pipeline is tightly coupled, meaning every decision you make at one step, like preprocessing, actually shapes what the model learns and how good its final performance will be. It shows how choices made early on can quietly introduce errors into the final map product.
Meng: I'm thinking about that coupling; when you’re shipping something to production at scale, those early decisions aren't just academic; they directly impact operational costs because every extra step compounds the cost of deployment.
Lalam: Exactly. It moves the focus from just getting a model trained to managing the entire ecosystem, which includes data access, preprocessing pipelines, and then the huge engineering challenges of scaling inference and handling artifacts.
Tom: So when we look at what this paper suggests are good practices—it covers everything from how you access data to how you handle those big compute issues during scale-up. It’s less about one single trick and more about setting a solid workflow foundation for anyone trying to make these large maps.
Jane: It’s really helpful that it breaks down the access landscape, showing that there are different trade-offs depending on whether you use a canonical provider or a re-distribution platform like Google Earth Engine. That distinction matters because it can lead to distribution shifts in your data, which is something we need to keep an eye on when we talk about real world deployment.
Lu: And the paper lays out this practical pattern of treating data access as a two stage process: model development uses smaller volumes that you can get from almost any provider, but inference at scale is dominated by the massive amount of data movement required. That distinction is huge for planning our infrastructure.
Meng: Data movement is exactly where I worry about things breaking at scale; if the data pipeline isn't designed right, your whole system grinds to a halt because you're bottlenecked by how fast you can move those massive datasets around.
Title and authors: Lalam: And it brings up that example of Google Earth Engine automatically correcting for radiometric offsets in Sentinel-two data or applying a full preprocessing chain to Sentinel-one GRD data before redistributing the product as given by ESA, which can cause distribution shifts <ref:2607.24532#pg1>. That’s a concrete illustration of how platform choices affect the output quality.
Tom: So it’s not just about choosing one tool; it’s about understanding that the processing pipelines themselves can differ dramatically between platforms, and we have to account for those differences if we want our map to be reliable everywhere.
Jane: It sets up the next big area they focus on, which is how you handle uncertainty quantification because a map isn't just a static picture; it has to be scientifically credible so people can actually use it to make decisions. The paper stresses that you need dedicated methodological attention for uncertainty quantification and independent validation.
Lu: They discuss the sources of uncertainty—it comes from upstream input data, the labels you create, and then the model itself, which is epistemic uncertainty. And they point out that there are trade-offs with methods for quantifying this; no single method is cheap, scalable, and well calibrated all at once.
Meng: That's a tough spot to be in when you're trying to ship something reliable; you can’t just throw the most complex uncertainty metric at it if it makes inference take three times longer or requires data storage that triples the cost.
Lalam: So they recommend focusing on coverage and reliability plots, standardized residuals—which are like z-scores with an ideal mean of zero and a standard deviation of one—and proper scoring rules to get a handle on this. It’s about choosing what’s actually useful for guiding action rather than just throwing numbers at the problem.
Tom: Moving into the next part, they talk about how you actually build the training dataset from raw imagery, and that introduces spatial autocorrelation which can inflate your evaluation metrics if you aren't careful. This means data selection isn't neutral; it’s encoding assumptions right into your entire training pipeline.
Jane: That’s a crucial point because it ties directly into how we filter our input data. The paper suggests one common filter is discarding products above a certain threshold of cloud cover, but they also mention that the spatial strategy you choose—whether you use single time steps or composites—directly determines what the model actually learns and how well it generalizes.
Lu: And they touch on preprocessing complexity directly translating into operational cost at scale; every extra processing step you add compounds the operational cost of deployment significantly. That’s a practical reality when you think about running these models across millions of pixels or square kilometers.
Title and authors: Meng: If I’m thinking about deploying this, I need to know exactly what that compounding cost looks like, because if the preprocessing gets too heavy, we might have to drop the resolution or the frequency of updates just to make it run on our hardware.
Lalam: They also detail how you structure your training data formats and splits; they mention container formats like HDF5 or NetCDF4 versus key-value chunk stores like Zarr, and for splitting, they suggest spatial K-fold block cross-validation to divide the area into non-overlapping spatial blocks.
Tom: The paper also discusses how you define your grid—whether you use a simple Web Mercator projection or something more complex like Discrete Global Grid Systems such as HEALPix—because that grid choice defines the index over which data is sampled and split, and it affects both sampling and evaluation validity.
Jane: Before we get to the tricky stuff about how to make this work in reality, they spend time on uncertainty quantification again, but this time they talk about the structure of those errors. They point out that map errors are spatially correlated, so ignoring that spatial structure in your uncertainty measurements can lead to substantially underestimated confidence intervals.
Lu: That is a very important caveat because if you treat every error as independent noise when it’s actually related to where the pixels are located, your confidence metrics will be misleading. The paper recommends characterizing that residual spatial correlation in validation regions instead of just looking at averages.
Meng: That makes sense because if we don't know the spatial correlation of our errors, we can’t truly say how sure we are about a prediction in a specific area, which is critical when you're trying to deploy something for real-world use.
Lalam: So for uncertainty, they suggest focusing on coverage and reliability plots and standardized residuals—those z-scores with an ideal mean of zero and standard deviation of one—and proper scoring rules to get the right picture.
Tom: Now we move into the engineering side: map production. They treat this as an engineering detail, but they stress that your choices here directly affect the final quality and usability of what you ship out. They look at compute environments being shaped by data gravity, identifying regimes like data-to-compute versus compute-to-data.
Jane: And then there are artifact mitigation strategies they suggest, which are practical engineering fixes, like using overlapping patches with weighted blending or padding and cropping to center predictions. But they also warn us that these mitigation methods can complicate the process of propagating uncertainty through the final result.
Lu: The paper also points out that for inference at scale, you should prioritize architectures with higher throughput, like convolutional encoder-decoders over Vision Transformers, because this is what makes a model wall-to-wall inference tractable when dealing with global scale problems.
Title and authors: Meng: If we're talking about deployment on edge devices or high-throughput servers, I need to know if the model we train actually has the throughput to handle the data flow without creating massive latency issues.
Lalam: And for sharing these maps out, they suggest adhering to FAIR principles by supporting both large-scale batch downloads and interactive visualization through formats like Cloud Optimized GeoTIFFs and Zarr, which makes it much more usable than just dumping a big file somewhere.
Tom: Finally, they talk about map validation. They state that the gold standard for accuracy assessment is a probability-based sampling design where every location in the mapped domain has a known, non-zero chance of being selected for verification. This lets you get an unbiased estimation of accuracy over the full map domain.
Jane: That’s important because it means you aren't just checking spots that are convenient or easy to access; you’re testing the whole thing systematically against reference data covering every single location in the mapped area.
Lu: The paper concludes by summarizing what we learned: spatial autocorrelation demands those blocked splits to avoid inflated metrics, and your gridding choices affect both sampling and evaluation validity. It also really hammers home that uncertainty quantification is not optional.
Meng: I agree with that; you can't skip the validation step just because you’re trying to rush a deployment timeline; the cost of an inaccurate map is too high when you’re dealing with real-world applications.
Lalam: And they make sure to state their limitations clearly, saying that while they provide guidelines, the practical knowledge needed for navigating these stages is still quite dispersed across different communities and data providers.
Tom: So the main thing to remember from From Machine Learning to Large-Scale EO Products: Good Practices and Guidelines for Making Maps is that there’s no single magic answer; you need a thoughtful, multi-stage workflow that accounts for data access, training design, uncertainty handling, and rigorous validation.
Jane: It’s a lot of moving parts, but by treating the entire process as tightly coupled stages—from pre-processing decisions to final artifact mitigation—we can build maps that are not just technically sound but also scientifically credible for real use.
Lu: The implication is that we need better shared knowledge across the EO and ML communities because those domain-specific challenges are what make this workflow so difficult to master without these guidelines.
Meng: For production teams, it means investing time upfront in designing a robust pipeline structure rather than just chasing the highest accuracy score on a small test set.
Lalam: I’m confident that by following these good practices for making maps, we can move past simply generating pretty pictures and start building tools that are actually trustworthy for complex applications.
The paper's summary: Tom: So, this paper is basically laying out a set of best practices for anyone trying to make those massive geospatial maps using machine learning.
Jane: Right, so it’s not just about training a model; it’s about managing the whole process from start to finish—data access, preprocessing, training design—and then scaling up the engineering side.
Lu: The core idea they push is that decisions you make early on in the pipeline quietly shape the final map product and can cause errors to get baked in there.
Meng: And for me, it's about operational reality. They talk about how preprocessing complexity directly translates into how much it costs to run the model at scale, which is a huge factor when you’re deploying something globally.
Lalam: It also covers the whole spectrum of uncertainty—where that uncertainty comes from and what methods you can actually use to quantify it without making things too complicated.
Tom: They spend a lot of time talking about how data selection isn't neutral; picking what you keep, like filtering out cloudy images, encodes assumptions right into your training pipeline.
Jane: And they suggest specific ways to handle that; like using spatial K-fold block cross-validation when splitting your data for training and validation so you get a realistic evaluation instead of an overly optimistic one.
Lu: They also address how you define the grid—like choosing between Web Mercator or something more complex like HEALPix—because that choice affects both how you sample your data and how valid your final evaluation is.
Meng: And they stress that for deployment, you have to think about compute environments based on things like data gravity, figuring out if you’re moving data to the compute or the other way around.
Lalam: They also talk about artifact mitigation strategies during map production, like using overlapping patches with weighted blending or padding and cropping to handle border effects.
Tom: It sounds like they conclude that there isn't just one right way; the practical knowledge needed is scattered across different communities, but following these guidelines helps you build a foundation for a scientifically credible product.
Jane: Exactly. The paper makes it clear that accuracy assessment is essential, and you need to state where your methods fall short instead of just ignoring those limitations.
Lu: They point out that uncertainty isn't optional; you have to characterize the spatial structure of those errors because map errors are spatially correlated, and if you ignore that, your confidence intervals get way too wide.
Meng: So for production teams, it means investing time upfront in a robust pipeline structure before you start chasing accuracy scores on small test sets.
Tom: It’s about building a solid workflow from the ground up rather than just rushing to ship something that looks good on paper but breaks in the field.
Jane: This is really important because it shows that we need better shared knowledge across the EO and machine learning communities if we want to make these maps trustworthy for real-world use.
Lu: The implication for future work they hint at is focusing on how to combine all these stages—from raw data ingestion to final distribution—into a seamless system.
Meng: For me, it means prioritizing pipeline design over just finding the highest accuracy number you can get in a lab setting.
The paper's improvements: Tom: So, this paper isn't just giving us rules; it’s actually suggesting concrete improvements to how we handle those map products once they are trained.
Jane: Right, so instead of just saying "do this," they are pointing out specific technical fixes for the problems we talked about earlier, like how to better mitigate those sensor artifacts.
Lu: They suggest using data augmentation specifically to simulate things like narrow no-data seams in Sentinel-one imagery, which helps the model learn to be resilient against those specific sensor issues.
Meng: That's smart because if you can train a model on simulated bad data, it becomes much more robust when you deploy it to real traffic or real satellite feeds.
Tom: And on the engineering side, they talk about artifact mitigation again, like using overlapping patches with weighted blending or padding and cropping predictions at the edges of the map.
Jane: But they also give us a heads-up that these mitigation strategies can complicate how we calculate uncertainty later on.
Lu: They’re pointing out that you have to be careful; while you’re fixing artifacts, you might accidentally make it harder to measure the true error distribution later.
Meng: That's a real concern for me because if uncertainty gets messy during the production phase, we can’t actually trust the map for safety applications.
Tom: They also emphasize that when doing inference at scale, we should prioritize model architectures with higher throughput, like convolutional encoder-decoders over Vision Transformers.
Jane: That makes sense for global scale; if you want it to run wall-to-wall without massive latency issues, the architecture has to be designed for speed.
Lu: So it’s a trade-off between model complexity and operational tractability; you pick one based on whether your inference needs to be fast or if you can afford a bit more computational load.
Meng: I definitely lean toward faster architectures if we're talking about things that need to run continuously, because the cost of running slow models across massive datasets is just too high.
Tom: And finally, they reinforce the validation side by stressing that you need to explicitly state what your method *doesn't* do when you assess accuracy against reference data.
Jane: They want us to be honest about the gaps; it’s not enough to just say how accurate a map is at a few points, we need to know where it stops working entirely.
Lu: And they conclude by saying that the practical knowledge isn't just theoretical; it’s about building this whole system together, acknowledging all these coupled dependencies.
Meng: So for anyone building this, the message is clear: don't just train a model and hope for the best; design the whole pipeline with these specific failure points in mind.
Conclusion: Tom: So, to wrap up, this paper lays out a practical blueprint for building large scale EO products using machine learning by focusing on robust workflow practices across data access, training design, and final validation.
Jane: It’s really about moving past just getting a cool result to actually shipping something that scientists and engineers can trust and use for real-world action.
Lu: The authors stress that the practical knowledge isn't scattered; it’s about treating the entire pipeline as one coupled system where early decisions have long-term consequences.
Meng: For production teams, this means spending time designing a structure that handles those data movement bottlenecks and operational costs right from the start.
Lalam: I’m confident that by following these guidelines for "From Machine Learning to Large-Scale EO Products: Good Practices and Guidelines for Making Maps," we can move past just generating pretty pictures and start building tools that are actually trustworthy.
Tom: It’s about making sure we account for spatial autocorrelation in the training data so our evaluation metrics don't get inflated by chance.
Jane: And they make it very clear that uncertainty quantification is not optional; you have to characterize those errors spatially because they aren't just random noise.
Lu: They conclude that accuracy assessment remains essential, and we need to state the limitations plainly instead of ignoring where the method stops working.
Meng: I agree with that; you can’t rush a deployment timeline and skip the validation step when you're dealing with things that could impact real-world safety.
Tom: The implication is that building these maps requires a holistic approach, not just optimizing one part of the machine learning process in isolation.
Jane: It shows us what we need from our community: better shared knowledge on how to navigate these complex stages together.
Lu: The authors point toward future work focusing on unifying all these stages into a single system for easier management and deployment.
Meng: That unified view would be huge for making sure that everything from the raw data ingestion to the final distribution is running smoothly in production.
Lalam: It’s clear they want us to build a pipeline where every step is considered and accounted for, which is exactly what we need for reliable AI applications.
ETH Zurich Institute for Photogrammetry and Remote Sensing Department of Computer Science and Technology University of Cambridge Swiss Federal Research Institute WSL Asterisk Labs EcoVision Lab Inst. for Earth System Science and Remote Sensing Leipzig University Chair of Data Science in Earth Observation TU Munich RISE Research Institutes of Sweden Climes Swedish Centre for Impacts of Climate Extremes Climate AI Nordics
cs.LG
Submitted: 2026-07-27
Updated: 2026-11-06
Comments: This work has been submitted to the IEEE for possible publication
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 85/100
The gist: Recent years have seen a rapid expansion in the production of large-scale geospatial maps derived from Earth observation (EO) data…
Key concepts
- Data Access Pattern
- This describes how researchers interact with satellite data. The paper suggests a two-stage approach: modest data volumes for model development and massive data movement for large-scale inference. Choices here significantly impact downstream processing costs and feasibility.
- Spatial Autocorrelation
- When training models on geospatial imagery, nearby pixels often have similar values, which inflates evaluation metrics. The paper recommends using spatial K-fold block cross-validation to ensure a realistic evaluation instead of an overly optimistic one.
- Uncertainty Quantification (UQ)
- UQ transforms a static map into a credible tool by measuring error. It involves analyzing input uncertainty, label uncertainty, and model uncertainty. The paper advises using coverage plots and standardized residuals to properly characterize the spatial structure of these errors.
Terminology
Summary
Recent years have seen a rapid expansion in the production of large-scale geospatial maps derived from Earth observation (EO) data, driven largely by advances in machine learning (ML) and large computing infrastructure <ref:2607. p. 024532#pg2>. Although the barrier to generating such maps has dropped substantially, established best practices have yet to emerge, and design decisions made early in the pipeline can quietly propagate errors into the final product <ref:2607. p. 024532#pg2>. Producing a technically sound and scientifically credible product remains challenging <ref:2607. p. 024532#pg2>.
The EO Data Infrastructure Landscape
Researchers must navigate a complex ecosystem of satellite data sources, derived data products, and access platforms <ref:2607.24532#pg2>. The choices made at the data access level have downstream consequences that are easy to underestimate <ref:2607. p. 124532#pg2>. A practical pattern is to treat access as a two-stage process: model development typically involves modest data volumes downloadable from almost any provider, while inference at scale is dominated by data movement <ref:2607. p. 224532#pg2>. For example, in their COPERNICUS/S2 HARMONIZED collection, Google Earth Engine automatically corrects for the radiometric offset that was introduced for Sentinel-2 data post Processing Baseline 04. p. 200; they also apply a full preprocessing chain (thermal noise removal, data calibration, multi-looking and range-doppler terrain correction) to Sentinel-1 GRD data <ref:2607. p. 224532#pg2>.
Data Selection and Preprocessing
Data selection and preprocessing choices are not neutral; they encode assumptions that propagate into the training pipeline and ultimately shape the final map product <ref:2607.24532#pg2>. Key considerations include:
-
Individual observations vary in usability, with a common filter being discarding products above a certain threshold of cloud cover <ref:2607. p. 224532#pg2>.
-
Spatio-temporal strategy determines what the model can learn and how well it generalizes, with strategies including single time-steps, time-series, composites, and time series of composites <ref:2607. p. 324532#pg2>.
-
Preprocessing complexity directly translates into operational cost at scale because every additional processing step compounds the operational cost of deployment <ref:2607. p. 224532#pg2>.
Machine Learning Dataset Construction and Model Training
Constructing a training dataset from EO imagery introduces challenges such as spatial autocorrelation that inflates evaluation metrics <ref:2607.24532#pg2>. Key decisions in this stage involve:
** Data formats:**
- Two broad storage patterns exist: container formats with an internal index (HDF5, NetCDF4, GeoTIFF, COG) and key-value chunk stores (Zarr) <ref:2607. p. 424532#pg2>.
Dataset splits:
- Spatial K-fold block cross-validation divides the area into non-overlapping spatial blocks and assigns each to a fold <ref:2607. p. 424532#pg2>.
Spatial gridding:
- A grid defines the spatial index over which data is sampled and split, with candidate grids ranging from Web Mercator to Discrete Global Grid Systems such as HEALPix <ref:2607. p. 424532#pg2>.
Uncertainty Quantification
Uncertainty is what transforms a map from a static product into a scientifically credible tool that guides action <ref:2607.24532#pg2>. Sources of uncertainty arise from upstream (input) uncertainty, label uncertainty, and model (epistemic) uncertainty <ref:2607. p. 524532#pg2>.
Methods for UQ present trade-offs, as no method is simultaneously cheap, scalable, and well-calibrated <ref:2607. p. 524532#pg2>. The paper recommends focusing on coverage and reliability plots, standardized residuals (z-scores with ideal mean 0 and std 1), and proper scoring rules <ref:2607. p. 524532#pg2>.
Map Production and Distribution
Producing a map from a trained model is often treated as an engineering detail, but the choices made at this stage directly affect the quality and usability of the final product <ref:2607.24532#pg2>. Key engineering considerations include:
-
Compute environments are shaped by data gravity, identifying regimes like data-to-compute and compute-to-data <ref:2607. p. 624532#pg2>.
-
Artifact mitigation strategies involve either overlapping patches with weighted blending or padding and cropping to center predictions <ref:2607. p. 624532#pg2>.
-
Sharing should adhere to the FAIR principles, supporting both large-scale batch downloads and interactive visualization via formats like Cloud Optimized GeoTIFFs and Zarr <ref:2607. p. 624532#pg2>.
Map Validation
Map validation assesses the final product against systematically and independently collected reference data covering the full mapped domain <ref:2607.24532#pg2>. The gold standard for map accuracy assessment is a probability-based sampling design where every location in the mapped domain has a known, non-zero chance of being selected for verification <ref:2607. p. 724532#pg2>. This allows for unbiased estimation of accuracy over the full map, not just at convenient or available locations <ref:2607. p. 724532#pg2>.
Conclusion
The practical knowledge needed to navigate these stages is often dispersed across communities, and the choice of data provider is not neutral <ref:2607.24532#pg2>. The paper highlights that spatial autocorrelation demands blocked splits to avoid inflated metrics, gridding choices affect both sampling and evaluation validity, and uncertainty quantification is not optional <ref:2607. p. 724532#pg2>. Accuracy assessment remains essential, and its limitations should be stated rather than ignored <ref:2607. p. 824532#pg2>. The paper refers readers to the online version for additional depth <ref:2607. p. 124532#pg2>.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Data selection should explicitly address artifacts like
Sentinel-2 products, specifically, exhibit artifacts from incomplete orbits and split partial products,
to avoid producingpixel duplication, bias dataset statistics.
This ensures the training signal is not corrupted by known sensor issues. -
Model design must account for operational constraints by prioritizing architectures with
higher throughput
like convolutional encoder-decoders over Vision Transformers when designing models for global scale inference. This choice directly impacts whether a model can runwall-to-wall inference tractable.
-
During training, instead of just preprocessing, use data augmentation to simulate corruptions, such as
simulating narrow no-data seams in Sentinel-1 GRD mosaics,
which helps the modellearn to be invariant to them.
This builds resilience against specific sensor artifacts. -
When splitting datasets for training and validation, employ
spatial K-fold block cross-validation
with a buffer zone to avoid leakage, as this leads to aRealistic evaluation
rather than anOptimistic evaluation.
-
For uncertainty quantification, explicitly address the spatial structure of errors by characterizing residual spatial correlation in validation regions because
Map errors are spatially correlated,
and ignoring this leads tosubstantially underestimated confidence intervals.
-
Map production must incorporate artifact mitigation strategies like
(a) overlapping patches with weighted blending
or(b) padding and cropping
to handle border effects, while noting that these methods can complicate uncertainty propagation.
Sources
- AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data
- TESSERA: Temporal Embeddings of Surface Spectra for Earth Representation and Analysis
- TEMPO: Global Temporal Building Density and Height Estimation from Satellite Imagery
- OlmoEarth: Stable Latent Image Modeling for Multimodal Earth Observation
- Tiling and Stitching Segmentation Output for Remote Sensing: Basic Challenges and Recommendations
- Foundation Models for Generalist Geospatial Artificial Intelligence
- ECHOSAT: Estimating Canopy Height Over Space And Time
- Estimating Canopy Height at Scale
- Capturing Temporal Dynamics in Large-Scale Canopy Tree Height Estimation
- Making Convolutional Networks Shift-Invariant Again
Related papers
- A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
- FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis