Dan A. Calian

Reinforcement-learning algorithms have traditionally been designed by hand, and discovering ones automatically that rival them has been a long-standing open problem. This work shows it is possible. By meta-learning from the experience of a population of agents across many complex environments, the method discovers an update rule — governing both what an agent predicts and how it learns from those predictions — that matches or exceeds the best hand-designed algorithms and generalizes to environments never seen during discovery.

The quality of a foundation model depends heavily on its training data, yet curation usually means hand-tuned data mixtures or hand-crafted filters. DataRater instead learns which data is worth training on. It uses meta- gradients to estimate the value of each individual training example, with the objective of improving training efficiency on held-out data. Across a range of model scales and datasets, filtering with the DataRater markedly improves compute efficiency.

Gradient-based bilevel optimization — the engine behind meta-learning, hyperparameter optimization and algorithm discovery — requires differentiating through an optimization process. This produces expensive “gradient-of-a-gradient” computations involving second-order and mixed derivatives, which standard automatic differentiation handles inefficiently. MixFlow-MG uses mixed-mode differentiation to build more efficient computational graphs for these problems, giving over 10× memory savings and up to 25% faster wall-clock time in modern meta-learning setups.

Deploying RL agents in the real world often means respecting system constraints, but constraint thresholds are frequently mis-set — sometimes so that no policy can solve the task without violating them. Since such violations are usually undesirable rather than catastrophic, this paper takes a soft-constrained approach: it uses meta-gradients to learn a good trade-off between expected return and constraint violations. The method consistently outperforms baselines across four MuJoCo domains.

Adversarially robust training appears to need far more data than standard classification, leaving a gap between models trained on the original data and those given extra real data. This work shows that samples from a generative model trained only on the original data can close much of that gap. It identifies when generated data helps, achieves state-of-the-art robustness on CIFAR-10, CIFAR-100, SVHN and TinyImageNet, and finds — surprisingly — that even random, non-realistic samples can improve robustness.

The technical report for the Gemini 2.X model family, including Gemini 2.5 Pro and 2.5 Flash. Gemini 2.5 Pro is a “thinking” model that reaches state-of-the-art performance on frontier coding and reasoning benchmarks, with strong multimodal understanding and very long context (up to three hours of video). The lighter Flash models trade some capability for much lower compute and latency, so that together the family spans the Pareto frontier of capability versus cost.

LINAC is an input-transformation defense that encodes pixel intensities through an implicit neural representation. Classifiers using it appear robust to small perturbations without adversarial training and with little loss of clean accuracy, and the random seed used to build the representation acts as a private key. The paper also proposes a new bypass attack (PBA) for such key- based defenses — showing that apparent robustness can depend on the breadth of attacks used to evaluate it.

Neural networks remain vulnerable to common corruptions such as blur, noise and fog. Prior methods defend against these on average, while the adversarial- robustness literature targets worst-case perturbations. This paper brings the two together with AdversarialAugment, which optimizes image-to-image models to generate adversarially corrupted augmentations, improving both average-case performance on corruption benchmarks (CIFAR-10-C) and worst-case robustness on CIFAR-10 and ImageNet.

Adversarial training tends to suffer from “robust overfitting”: robust test accuracy declines as training proceeds. This paper shows that, contrary to earlier findings, combining data augmentation with model-weight averaging significantly improves robust accuracy. Spatial composition augmentations work best, reaching state-of-the-art robustness on CIFAR-10 without any external data.

Control policies must often satisfy constraints once deployed, yet real systems drift through wear, miscalibration and non-stationarity. This perturbs the dynamics, so a policy trained in one setting may both earn less reward and violate its constraints in another. The paper formalizes this “constrained model misspecification” and presents an algorithm that is robust to it, demonstrated on simulated MuJoCo tasks from the Real- World RL suite.

Recommender systems often balance competing objectives — engagement against diversity and freshness — by imposing constraints on their rankings. This can be cast as a linear program, but off-the-shelf solvers are too slow to run on live traffic. NodeHopper exploits the structure of the dual problem to build a fast solver that handles these constrained ranking problems an order of magnitude faster than general-purpose solvers, making them deployable under tight latency budgets.

Self-attention in Transformers scales quadratically with the number of pixels, which is prohibitive for large images. SCRAM is a fast randomized algorithm that computes image attention maps in O(n log n) time by exploiting the spatial coherence and sparsity of real-world attention maps. It adapts the PatchMatch correspondence algorithm to find each query’s most compatible key, building a sparse approximation to the underlying non-local-means operation.

Photorealistic compositing needs an accurate capture of scene lighting, which is difficult for casual photographers. This paper estimates a high-dynamic- range light probe from a single ordinary outdoor photograph by treating a person’s face as a light probe, using inverse rendering with data-driven priors learned from a large set of HDR sky maps. It recovers high-frequency, very-high-dynamic-range lighting, and contributes a new dataset of faces paired with ground-truth light probes.

Convincing mobile augmented reality requires lighting virtual objects consistently with the real scene, which usually demands costly radiometric calibration and complex shading. This work introduces a physical “shading probe” that captures the scene’s shading directly, rather than its incident light. This enables an efficient reformulation of the rendering equation that delivers fast, plausible shading — including local global-illumination effects — on mobile devices.

Local search can sharpen evolutionary learning, and it matters most on the very large search spaces found in large-scale data mining. This paper adapts memetic operators — which use the supervised-learning signal to heuristically edit classification rules — to the BioHEL iterative learning system, and extends them to handle continuous attributes and noise. The resulting configurations are evaluated against other machine-learning methods on large synthetic and real-world datasets.

3D printers handle static models well, but printing posable, articulated objects — a hand with working knuckles, say — usually requires manual assembly, and earlier joint designs could not hold a pose without parts fusing during printing. This paper presents a workflow that converts a rigged 3D model into a printable, non-assembly model whose joints have enough internal friction to hold poses against gravity. It fits posable joints automatically and lets the user specify rotational limits.