Yaowen Ye, Jacob Steinhardt, Sarah Schwettmann, Cassidy Laidlaw
Training data can have unexpected side effects on language model behavior outside the training domain: insecure code data can induce broad misalignment, and outdated bird names can induce a 19th-century persona. Understanding such generalization is hard because only dozens of isolated examples exist, each constructed by hand or discovered by accident. We propose automatically discovering such datasets with surprising side effects (DSSEs). We define three objectives for DSSEs: they should (1) elicit a target behavior, (2) read like ordinary post-training data, and (3) be surprising to human auditors reading the data. We operationalize each with LLM judges, and search for high-scoring datasets via evolution, reinforcement learning, and subliminal learning approaches. After re-scoring and filtering with held-out metrics, our pipeline produces a collection of 6,575 DSSEs across 8 behaviors, many of which generalize across model families and sizes. We find our automated metrics track ML researchers' judgments, and frontier models outperform humans at predicting side effects. These datasets form a testbed for studying generalization and lay the groundwork for predicting training side effects before model deployment. Explore them at https://dsse.transluce.org.
Trung-Khang Tran, Tuan A. Hoang, Viet-Hoang Tran, Tan M. Nguyen
Conformal Prediction (CP) provides a model-agnostic framework for constructing finite-sample prediction sets with coverage guarantees and has become a tool for reliable decision-making in high-risk settings. Recent advances have extended Split Conformal Prediction to increasingly complex learning problems, including multilabel, high-dimensional, and functional outputs. In this work, we study CP for Distribution-to-Distribution Regression, a setting of growing interest across many application domains. We show that a naive application of Split CP yields prediction sets that, while marginally valid, are non-interpretable, e.g., lacking a well-defined volume or a tractable sampling procedure, and of limited practical value. To overcome this, we propose a new conformal framework that operates in an orthogonal basis space, producing interpretable prediction sets while preserving coverage guarantees. We further develop an adaptive variant with asymptotic conditional coverage under mild assumptions. Empirical results on synthetic and real-world datasets validate the effectiveness of the proposed approach. The code is publicly available at https://github.com/tuanhacs/conformal_d2d.
Model merging combines multiple fine-tuned models without additional training, and arithmetic methods have emerged as a simple yet effective way to merge models without data. However, existing arithmetic methods treat task vectors in parameter space and apply a single element-wise rule to directions that can differ in mergeability. In a shared directional basis, we find a consistent two-regime structure: a low-dispersion common bulk shared across tasks and a high-dispersion task-dominated tail that carries disproportionate energy. These findings motivate Common--Unique Subspace Decomposition (CUSD), a data-free pre-merging method that separates task vectors into common and unique directional components, merges the common components with a downstream arithmetic operator, and retains unique residuals separately. Experiments across $11$ base models, four model families, and four arithmetic-based merging operators show that CUSD improves 42 of 44 model--operator combinations and achieves up to 22.5% improvement over the corresponding baselines. Our code is publicly available at https://github.com/stwang666/CUSD
Optical Character Recognition (OCR) is essential for digitizing historical archives, yet standard evaluation metrics such as Character Error Rate (CER) do not necessarily reflect whether meaningful information, such as names, dates, and identifiers, is preserved. In this work, we investigate entitylevel analysis of OCR errors as a complementary perspective for evaluating OCR quality in historical documents. Using a publicly available benchmark of 175 documents from the Berrutti Archive, augmented with 5,209 manually annotated entity mentions that will be released upon acceptance, we compare sequence-recognition OCR systems, OCR-specialized vision-language models, and general-purpose vision-language models. Beyond global transcription accuracy, we analyze entity preservation and recoverability under varying character-level edit tolerances. Our experiments show that OCR rankings depend on the evaluation criterion: while Qwen 3.6 achieves the highest entity coverage across all tested tolerances, several systems with lower exact-match coverage recover substantially when small errors are tolerated, suggesting that many of their errors correspond to localized character-level deviations rather than fundamentally incorrect entity transcriptions. Conversely, some systems remain unable to recover many entities even under relaxed matching. These findings show that OCR quality cannot be fully characterized by aggregate transcription metrics alone and highlight the importance of evaluating how OCR errors affect the preservation and recoverability of meaningful information.
Hugo Frezat, Ronan Fablet, Guillaume Balarac, Julien Le Sommer
In this paper, we propose a generic algorithm to train machine learning-based subgrid parametrizations online, i.e., with $\textit{a posteriori}$ loss functions, but for non-differentiable numerical solvers. The proposed approach leverages a neural emulator to approximate the reduced state-space solver, which is then used to allow gradient propagation through temporal integration steps. We apply this methodology on a chaotic two-timescales Lorenz-96 system and a single layer quasi-geostrophic system with zonal dynamics, known to be highly unstable with offline strategies. Using our algorithm, we are able to train a parametrization that recovers most of the benefits of online strategies without having to compute the gradient of the original solver. We found that training the neural emulator and parametrization components separately with different loss quantities is necessary in order to minimize the propagation of approximation biases. Experiments on emulator architectures with different complexities also indicates that emulator performance is key in order to learn an accurate parametrization. This work is a step towards learning parametrization with online strategies for climate models.
Léo Nicollier, Marc Pic, Pablo Musé, Enric Meinhardt-Llopis, Gabriele Facciolo
Despite recent advances, self-supervised learning (SSL) models and Joint-Embedding Predictive Architectures (JEPAs) remain susceptible to spurious correlations. Many anti-collapse regularizers constrain the global representation distribution to match a prescribed target, such as a multivariate Gaussian or the uniform distribution on the sphere. However, global matching does not prevent groups defined by a bias attribute from occupying distinct regions of the representation space. We show that existing encoder-only debiasing regularizers, including EnD and FSCL, can be interpreted as restricted forms of conditional distribution matching. Building on this interpretation, we introduce Unbiased Open World (UOW), which matches the representation distribution within each observed bias group to a common target distribution. At the population optimum, using a discriminative conditional discrepancy guarantees independence between the representation and the targeted attribute. We instantiate UOW in Gaussian and spherical spaces and evaluate it across controlled synthetic settings and real-world fairness benchmarks. Across Synthetic Engraving, Colored MNIST, CelebA, and Waterbirds, conditional matching substantially reduces bias compared with global regularization while preserving useful downstream representations. Jointly conditioning on multiple bias attributes further reduces intersectional disparities missed by separate marginal constraints. Finally, experiments on CelebA show that models can be trained from random initialization using UOW as a pretraining objective, provided checkpoints are selected according to fairness-aware validation criteria.
We ask how hard it is to build a ReLU network when its parameter values are given in advance. An instance supplies an appropriate encoding of a target behavior, a multiset of rational numbers, and a bound on the number of neurons (output included). The question is whether one can choose the graph of a ReLU network within the bound, and place the numbers on its edges and biases so that it meets the target. This separates the numbers from the architecture. Matching two labeled examples exactly is already $\mathsf{NP}$-complete, so we ask what changes when the target must hold on every input or on average. Agreeing with a Boolean circuit on every Boolean input is $Σ_2^{\mathsf P}$-complete, already for four neurons. Agreeing with a given ReLU network on every real input lies in $\mathsf{DP}$ for every fixed neuron bound and is $\mathsf{DP}$-complete for three neurons. Approximating a ReLU network on the unit cube within absolute error one everywhere is $Σ_2^{\mathsf P}$-complete, even with no hidden neurons and all numbers equal and positive. Deciding whether some network has expected squared loss at most a threshold is $\mathsf{NP}^{\mathsf{PP}}$-complete for uniform Boolean inputs. For uniform inputs on the unit cube and for standard Gaussian inputs with a positive threshold, the same holds for two neurons when a positive risk gap, given in binary and possibly exponentially small, separates yes from no instances. Without a gap, the cube version is complete for guessing a polynomial-length string, then comparing a sum of exponentially many rationals with a threshold, all computable in polynomial time from the input and the guess.
Mayan Sharma, Kriti Saini, Devansh Behl, Suganya R, S. Vijayaragavan
Risk-budgeted throughput forecasting converts a point forecast of LEO satellite capacity into a conservative bound whose overestimation rate is held at a user-specified budget, so that admission control can act on it safely. We reproduce the state of the art, BG-CFQS, on the three Starlink traces its authors used and on an independent fourth dataset, and report four findings. First, its risk guarantee is reproducible but conditional on the calibration split: under a random split we recover every published quantity, including risk-pass on 3 of 3 datasets, but under a contiguous temporal split, which is what a deployed terminal faces, it holds on 1 of 3. Second, independent of any split, the method cannot serve a risk budget below 0.15, the low end of its candidate quantile set. On all four datasets the achieved rate is identical at budgets of 0.05, 0.10 and 0.15, overshooting by 1.4x to 3.7x at the tightest. Third, static conformal calibration cannot improve a ranking, only shift an operating point; we confirm this exactly on latency spike detection, where online calibration improves AUPRC by up to 51%. Fourth, group-conditional online recalibration splits into two mechanisms. The online part replicates: it tracks the budget within 1% from 0.05 to 0.35, where static calibration is 16 to 21% off. The group-conditional part does not: on three of four datasets, conditioning on any available covariate, including measured serving-satellite geometry, is worse than not conditioning, and our attempt to predict which datasets benefit failed. What is new here is empirical, not algorithmic: we claim the findings, not the methods.
Daisy R. Bradley, Nathan A. Hinchliffe, Daniel J. Pitchforth, Matthew R. Jones, Elizabeth J. Cross
Machine learning plays an increasingly vital role in engineering, but the corresponding increase in compute time is not without environmental cost. Physics-informed machine learning or "grey-box" models have been developed to overcome some of the limitations of traditional black-box learners, utilising the physical insight that an engineer would have about the structure they are modelling and have shown promising results in the structural engineering field among many others. This work explores whether an additional advantage could be a reduced environmental impact, considering the relationship between training data quantity and training time, linking this duration to carbon emissions from computing. In a structural health monitoring context, four physics-informed machine learning approaches - spanning Gaussian processes and neural networks - are evaluated: residual modelling, input augmentation, hybrid modelling, and constrained learning. The emissions for training each of the models to reach a given error threshold is compared, and in most examples, shown to be lower for the physics-informed models (with input augmented models being an exception). This reduction in training emissions further compounds the environmental savings achieved by collecting and storing less data. Although promising results, we cannot expect a silver bullet and the case studies demonstrate that a trade-off is needed between the increased complexity that comes from introducing physics into a machine learner, against the gain from reduced training data requirements.
Machine learning approaches to subgrid-scale (SGS) modeling are now well established in atmospheric and oceanic applications. Among these, online end-to-end learning, where the differentiable solver participates in the training, has shown particular promise. Yet existing studies are largely restricted to idealised periodic domains, with no mechanical boundaries, precluding them from addressing the dynamics of bounded rotating flows relevant to planetary interiors. Here we consider two-dimensional quasi-geostrophic turbulence in a rapidly rotating annular bounded domain. We examine three configurations varying the geometry of the container and the rotation rate. The system exhibits key features such as zonal jets, Rossby waves, and, in the spherical shell geometry, a slow quasi-periodic inward drift of the jets. The spectral properties of the zonal and non-zonal flow can be understood in the framework of zonostrophic turbulence theory. We develop a differentiable solver for this system, which allows us to train SGS models online, over a time span of one turnover time, using coarse-grained data from direct numerical simulations. In all cases, a SGS model trained on a single turnover time accurately reproduces global integrated diagnostics, energy spectra as well as long-term dynamical behaviours, such as jet migration, occurring on timescales which exceed the training window by one order of magnitude. The online-trained model further outperforms classical hyperdiffusivity and Leith closure schemes, for which reducing the radial resolution is impractical. The resulting speed-up paves the way to further investigations of long-term geophysical fluid processes beyond the reach of direct numerical simulations.
Jonas Teufel, Luca Torresi, André Eberhard, Pascal Friederich
Numerical molecular representations underpin virtual screening, property prediction, and materials discovery. Conventional fingerprints are efficient and deterministic but lose structural information through hash-based compression at low dimensionalities, whereas graph neural networks recover expressiveness through compute-intensive, task-specific training. Here, we introduce hyperdimensional fingerprints (HDF), which compose the chemical attributes of atoms and their multi-hop environments through algebraic operations on high-dimensional vectors, producing deterministic representations without any training. Across diverse property-prediction benchmarks, HDF ranks consistently among the best representations across datasets and models. Crucially, HDF encodes structure more efficiently: at low dimensionalities, its distances correlate with graph edit distance more strongly than Morgan fingerprints', yielding effective similarity metrics and competitive nearest-neighbor regression at just 64 components. In Bayesian optimization, HDF-based surrogates are more sample-efficient than Morgan fingerprints with a Tanimoto kernel for most of the tested objectives. These results establish HDF as compact, training-free representations for property prediction, similarity search, and sample-efficient optimization.
Ana Leticia Garcez Vicente, Gijs van Seeventer, Saber Salehkaleybar
Estimating causal effects from observational data requires identifying valid adjustment sets. This task is especially challenging in realistic settings where latent confounding and feedback loops are present. Existing approaches typically assume acyclicity or rely on global causal structure learning, limiting applicability and computational efficiency. In this work, we study a local, data-driven method for covariate selection based on conditional independence information under a pre-treatment assumption. While this method is known to be sound and complete in acyclic causal models, its validity in the presence of cycles has remained unclear. Our main contribution is to show that these guarantees extend to cyclic causal models. In particular, our result relies on the invariance of conditional independence assertions under $σ$-acyclification. These findings establish a unified, cycle-agnostic perspective on covariate selection, which allows for causal effect estimation using any consistent estimator. Therefore, showing that the method applies across cyclic and acyclic settings without modification. Empirically, we validate this on extensive synthetic data, showing reliable performance in cyclic causal models.
Encrypted traffic classifiers are routinely reported at 95-99% accuracy. We show that on one of two public corpora, the same model and code score macro-F1 0.992 under the evaluation protocol standard in the literature and 0.567 under a protocol that keeps flows from one packet capture on one side of the split. Nothing changes but the partition (gap 0.426 over eight seeds, p = 0.008 after Holm-Bonferroni correction). The inflation is not a property of random splitting as such. On a month of backbone QUIC traffic, we evaluate five split protocols at eight seeds each: random, temporal and client-disjoint splitting are statistically indistinguishable from session-disjoint, while server-disjoint splitting costs 0.192 macro-F1 (d = 17.0, p = 0.008). Which axis matters depends on how the corpus was collected, in a way invisible from its label space. We give a diagnostic that identifies the right axis before any model is trained: single-identifier probes. A decision tree on the capture identifier alone reaches 0.983 under a random split and exactly 0.000 under session-disjoint splitting. Under this evaluation, a metric-trained flow embedding leads four deep baselines but does not beat gradient-boosted trees on twenty packet sizes. Five of our own pre-registered architectural claims fail: distance-based rejection of unknown applications beats softmax thresholding, but ablating every metric objective makes rejection better, because class separation correlates at r = -0.88 with rejection quality. Few-shot enrollment is flat within a corpus but, across corpora, five labelled flows per class recover 91% of the deficit left by failed zero-shot transfer. All splits, results, code and negative findings are released.
Christian Klötergens, Vijaya Krishna Yalavarthi, Lars Schmidt-Thieme, Tom Hanika
Tabular Foundation Models (TFMs) are the state of the art for many tabular prediction tasks. They are transformers that approximate the Bayesian posterior predictive distribution under a pretraining prior. TFMs become multivariate predictors autoregressively, by sampling one target and appending it to the features. This paper is the first to evaluate whether this setup induces a valid stochastic process. Informally, we ask: *Do we get the same answer to the same question asked in two different ways?* The standard Kolmogorov consistency conditions cannot be checked on TFMs directly, since TFMs are unable to model certain conditional distributions. We therefore propose two adapted, testable requirements. Marginalization consistency demands that marginalized conditionals equal directly predicted marginals, and factorization consistency that both factorization orders yield the same joint. Since the data-generating distribution satisfies both, a perfect TFM would as well, and every violation lower-bounds how far the model's predictions are from the ground truth. Every TFM we evaluate violates both requirements, for classification and regression, on all datasets.
Chain-of-Thought (CoT) prompting improves LLM accuracy on complex tasks but often increases token usage and inference cost. Existing ``Budget Forcing'' methods reduce cost via fine-tuning with heuristic length penalties, suppressing both essential reasoning and redundant filler. We recast efficient reasoning as a lossy compression problem under the Information Bottleneck (IB) principle, and identify a key theoretical gap when applying naive IB to transformers: attention violates the Markov property between prompt, reasoning trace, and response. To resolve this issue, we model CoT generation under the Conditional Information Bottleneck (CIB) principle, where the reasoning trace $Z$ acts as a computational bridge that contains only the information about the response $Y$ that is not directly accessible from the prompt $X$. This yields a general Reinforcement Learning objective: maximize task reward while compressing completions under a prior over reasoning traces, subsuming common heuristics (e.g., length penalties) as special cases (e.g., uniform priors). In contrast to naive token-counting approaches, we introduce a semantic prior that measures token cost by surprisal under a language model. Crucially, the prior is queried only for token-level log-probabilities, adding negligible overhead to the training loop. Empirically, our CIB objective prunes reasoning redundancy while preserving fluency and logic, improving accuracy at moderate compression and enabling aggressive compression with minimal accuracy drop. These gains generalize across model families and task domains, confirming CIB as a domain-agnostic CoT compression framework.
Iterative neural channel estimators can improve recovery under pilot contamination, but repeated network evaluations increase inference cost. This paper develops observation-anchored consistency and Consistency Mid-Training (CMT) estimators for one-step channel estimation in multiple-input multiple-output orthogonal frequency-division multiplexing (MIMO-OFDM) systems. Both estimators refine pilot-derived least-squares channel estimates in the delay-beam domain. Consistency learning couples endpoint predictions along noisy-to-clean paths. CMT then refines a compact consistency-trained student to predict a flow-matching teacher's endpoint from intermediate trajectory states. Across the evaluated SNR range, the compact CMT student reduces uncoded symbol error rate (SER) by 9.6-20.2% relative to the one-step flow-matching baseline (0.44-0.98 dB in logarithmic SER reduction), while using 74.5% fewer parameters and 73.0% fewer network multiply-accumulates.
Naïl B. Khelifa, Richard E. Turner, Ramji Venkataramanan
Score-based diffusion models are typically trained by minimizing the $L^2$ score matching error, and standard theoretical analyses rely on this quantity to bound the sampling discrepancy between the learned and target distributions. Recent work has shown that the score admits an orthogonal decomposition into gradient and solenoidal components in which only the gradient part matters for exact sampling. We apply this decomposition to the score error of trained models and study its quantitative consequences. Since only the gradient component of the error enters the marginal Fokker-Planck dynamics, a model can incur arbitrarily large $L^2$ score error while perfectly matching the target. Starting from this observation, we present three results. First, we derive an upper bound on the Kullback-Leibler divergence that depends only on the observable gradient component of the error. It tightens the standard Girsanov bound for generic score networks and identifies that bound's looseness as the cost of working with path-space rather than marginal-space dynamics. Second, we prove that no non-trivial function of the $L^2$ score error can uniformly lower bound any divergence between the learned and target distributions. Third, we give an estimator of the gradient component, computable from denoising score matching residuals via a dual Sobolev identity, which empirically correlates substantially better with sample quality than the full $L^2$ error.
Yun Young Choi, Asung Kil, Sun Woo Park, Minho Lee, Seokhwan Kim
Several widely used families of Graph Neural Networks carry a positive-definite geometric prior. Graph convolution read as heat diffusion needs a positive semi-definite operator, hyperbolic models place features on positive-definite Riemannian manifolds, and dot-product and GAT-style attention scores use positive-definite inner products. On heterophilic graphs, where edges often signify dissimilarity or structural repulsion, that prior is a mismatch: no positive-definite form can assign a negative squared distance to a direction the labels separate. To measure the mismatch, we introduce the Graph Signature Index ($s_{\mathcal{G}}$), the inertia of a class-compatibility matrix built from label topology alone, and we prove that no value of the edge or node homophily index in $(0,1)$ rules out a negative eigenvalue. We then propose the Pseudo-Riemannian Attention Network (PRAT), whose attention logit is a signed squared pseudo-distance under a learnable diagonal metric initialized from $s_{\mathcal{G}}$. The design is diagnose-then-learn: the signature supplies a sign structure that unconstrained training keeps outside near-zero directions, and the projection learns the rest of the metric around it. Across fifteen node-classification benchmarks PRAT outperforms or matches the strongest heterophily-adapted models. A controlled ablation ties that performance to the sign structure itself: with architecture and protocol fixed, forbidding the metric to be indefinite costs accuracy on the graphs $s_{\mathcal{G}}$ marks as indefinite and leaves the others unchanged.
Jorge García-Carrasco, Javier Sanchis, Álvaro Navarro, Alejandro Maté, Juan Trujillo
Open-weight language models that fit on a single consumer GPU are becoming capable on a growing range of tasks, and they matter because they can run locally without sending the code to a hosted service. Vulnerability reproduction, turning a vulnerability description into an input that crashes the vulnerable build, is a natural target for them, since it is an automated analysis task closer to test generation than to code comprehension and it runs against a build that must be executed rather than read. In order to assess whether current models can perform it, we evaluate twelve configurations of local models, spanning three generations, two architectures and three quantization levels, on ten CyberGym vulnerability-reproduction tasks, three runs each under one harness, and release the full artefact. We show that every configuration, down to the 4B model, solves most of the six easier tasks, and configurations separate only on the remaining four. For the harder tasks, the 27B model is required. Quantization did not change what the models solved in any matched pair, but it changed how fast they got there, and the mixture of experts reached the same scores in less wall clock. We also examine the agent trajectories. Almost every run we examined, including those that fail, names the vulnerable function and states the correct hypothesis, and what separates a solved task from an unsolved one is whether that hypothesis becomes a working input.
Reward alignment for diffusion- and flow-based generative models is often formulated over stochastic trajectories, making it less suitable for few-step deterministic generators. Initial noise optimization provides a trajectory-free alternative, but existing gradient-free methods require costly per-sample search at inference. We introduce LeNO (Learning Noise Optimization), which amortizes reward-guided initial noise optimization into a learned residual transport while keeping the pretrained generator frozen. LeNO learns an anchor-conditioned distribution over noise corrections using Local GRPO and scalar rewards, eliminating iterative test-time optimization. For Gaussian initial noise with norm-preserving projection, we derive a tractable upper bound on the KL divergence between the corrected noise distribution and the original prior, providing a principled regularization objective. We evaluate LeNO on few-step text-to-image and protein structure generation, including expensive black-box designability rewards. Across both domains, LeNO improves target rewards with a single noise correction at inference while largely preserving generation quality and diversity.
C. F. Maximilian Nagy, Onur Celik, Emiliyan Gospodinov, Florian Seligmann, Weiran Liao, Aryan Kaushik, Gerhard Neumann
Action chunking improves exploration and accelerates value propagation in long-horizon reinforcement learning, but naively applying off-policy methods to the temporally extended action space at reduced decision frequency offsets these gains, leading to poor sample efficiency. Existing online action chunking methods address these issues through computationally expensive critic-only approaches or by relying on offline data. We introduce SEAR, a sample-efficient off-policy algorithm that enables online reinforcement learning with action chunks by addressing both challenges. To handle high-dimensional action sequences, SEAR employs a causal transformer critic trained with multi-horizon targets that provide a training signal for every prefix of an action chunk, effectively increasing the useful gradients per sample. To restore decision frequency, SEAR operates with a receding horizon and random replanning, ensuring uniform state coverage while combining the fast value propagation of large chunks with the reactivity of small ones. SEAR outperforms state-of-the-art online reinforcement learning methods including SimbaV2 on challenging Metaworld manipulation tasks. Beyond online RL, applying SEAR to the offline-to-online method QC improves its performance on the demanding OGBench cube-triple tasks.
Zezhou Wang, Hanzhang Zhou, Panrong Tong, Quyu Kong, Pengxiang Li, Steven Hoi, Zhuzhong Qian, Yue Wang
Scaling GUI reinforcement learning requires environments that preserve real-app behavior. Yet coupling behavior inference with coding can turn partial traces into runnable mocks with incorrect state transitions. We introduce \textbf{Trace2World}, which separates global behavior reconstruction from implementation. It reconciles traces into a shared, source-linked app model, then compiles an executable environment. Independent replay of source workflows localizes mismatches for bounded implementation repair. Validated worlds then support learner-adaptive task and environment augmentation within source-supported behavior. From 1,570 traces, Trace2World constructs 50 environments. On 200 workflows across ten apps, 91.1% are reproduced end to end, versus 59.8% for the strongest matched task-local baseline. Source-grounded validation retains 94.0% of simulated training gains on real devices, versus 54.7% under self-validation. Learner guidance raises benchmark success from 56.9% to 61.1%. Training in Trace2World further improves the model by 20.2 points on our benchmark, 5.4--16.2 points across four external GUI benchmarks, and 16.3 points on 100 real-device tasks. Reliable reconstruction separates what behavior source evidence supports from whether the implementation preserves it.
Sumeet Ramesh Motwani, Tejas Krishnan, Alistair Letcher, Charles London, Suhaas Bhat, Huitian Jiao, Philip Torr, Riashat Islam 等 13 位作者
Reinforcement learning with verifiable rewards (RLVR) has led to significant gains in math and code generation. Extending these gains to open-ended domains remains challenging: ground-truth verification is unavailable, human annotation is expensive, and learned reward models are prone to reward hacking. The generation-verification gap, where judging good outputs is easier than producing them, offers a potential path to improvement. However, rubric-based judges are often imprecise and stationary, so policies can overfit or reward hack them. We introduce Rubric Curriculum RL (RcRL), which exploits this gap without new data, human annotations, or stronger teacher models. We combine pairwise preferences over rollouts with a curriculum over rubric criteria, yielding a more discriminative signal than absolute scoring and a non-stationary objective. Whereas baselines plateau within a few hundred steps, RcRL preserves output entropy and keeps improving past 1000 steps. On creative writing, RcRL achieves a 70.5% human win rate against the base model, with gains across benchmarks and judges. On HealthBench, RcRL improves over the base model by 14.6% (full set) and 25.4% (consensus subset), outperforming all training baselines, including instance-specific rubrics.
Decision models read a document and return a probability for each of a fixed set of options instead of generating text. Several appeared in 2026, but none has published results outside English. We describe Zuse, three decision models (4B, 9B and 27B) that train a low-rank adapter and a pointer head on frozen open-weight bases, targeting 27 European languages, and use them to study how this design behaves across languages and tasks. On a broad European suite of 11 sources excluded from training, covering 9 task groups and 27 languages, a development set that also guided model selection, the 27B model reaches 87.5% accuracy with a pooled expected calibration error of 0.016, and 0.035 when weighted per source. Post-hoc analyses of these single-seed results give three findings. First, miscalibration is structured by task rather than language: on parallel reading and topic items in 26 languages the confidence gap varies about as much as sampling noise, while across sources it spans -0.07 to +0.09, with knowledge tasks overconfident and acceptability judgements underconfident at all three sizes, so one fitted temperature cannot correct both. Second, the share of decisions that can be automated at a 5% error budget follows chance-corrected skill (Spearman 0.97 over 33 source and model pairs) more closely than raw accuracy (0.91). Third, reach beyond the training languages is set by the base model: across 210 languages, changing the training run moved scores by at most about 2 points within a size, while languages outside the base's documented list trail by 23 to 28 points at every size. On English public suites Zuse is behind Jev and Kev, including Kev models on the same bases. We also report two pitfalls in building multi-source training data: label-word collision between topic schemes, and an injection augmentation that taught a shortcut. Zuse can be tried at https://bineric.com/zuse.
Federico Pavesi, Noémie Jaquier, Antonio Candelieri
Probabilistic meshless methods solve partial differential equations (PDEs) by placing a Gaussian process prior on the solution and conditioning it on the PDE at a finite set of collocation nodes. The posterior provides both an approximation of the solution and a Bayesian quantification of its error. While these methods are well established on Euclidean domains, their extension to manifolds lacks, to our knowledge, any convergence analysis. We extend them to compact manifolds and study their convergence in two settings. First, for known manifolds whose Laplace-Beltrami eigenpairs are available, the prior can be evaluated exactly, but the PDE operator generally cannot be applied to it in closed form. We replace this operator by a standard generalized finite difference discretization on a point cloud sampled from the manifold, and show that the resulting method is consistent at the points of the cloud, at a rate governed by their fill distance. Second, for implicit manifolds, known only through a point cloud, both the prior and the operator must be approximated. We propose a new generalized finite difference scheme with a free stencil radius, and show that it retains the consistency of the standard construction while controlling the growth of its coefficients, and that an appropriate radius yields uniform convergence of the discrete Gram matrix. In both settings, we derive error bounds for the posterior mean which separate the collocation error intrinsic to the method from the error due to approximating the operator and, for implicit manifolds, the prior. Experiments on the circle, the sphere and the flat torus confirm our theoretical claims. On real-world cortical thickness data, our approach attains the accuracy of established discretizations, while conditioning on fewer than a tenth of the observations and providing a credible interval at every vertex.
Can independent components be recovered when all of them are Gaussian? Yes, if their variances change across environments; but how many environments, and how many samples as the number of components $d$ grows? We answer both for the joint diagonalization of $K$ covariances. Each component has a signature, its variances across the environments. Separating component $j$ costs the Riesz 2-energy $E_j$ of this signature, a sum, over the other components, of inverse pairwise Fisher informations. For Gaussian sources no estimator in $O(d)$ beats $cE_j/(n+E_j)$ at any $n$ and $d$, and above an explicit sample size a spectral estimator attains $E_j/n$ up to a constant, for every mixing and without the signatures (given a lower bound on the pairwise information). Since signatures live in $K-1$ dimensions, the rate obeys a dimension law: $d^2/n$, $d\log d/n$ and $d/n$ for $K=2$, $3$ and $K\ge4$ well-spread environments. What counts is the dimension the signatures span, not their number, and $\log d$ well-designed environments make $d\log d$ samples suffice. Projecting the covariances onto one linear combination pays $d^2/n$ at first order instead. In simulations up to $d=1000$ the predictions hold and joint diagonalization with $K\ge4$ attains the rate once $n/d$ is in the hundreds; on task fMRI the energy says which components are hard to separate.
Mingi Choi, Gunhee Kim, Jisoo Kim, Taeksoo Kim, Taeyun Ha, Jongbin Lim, Hanbyul Joo
Scaling real-world dexterous-grasp data collection is limited by the sustained human effort required to operate the robot, verify grasp outcomes, and reset objects between trials. Simulation can generate grasp candidates at scale, but their success on hardware remains uncertain. We present AutoDex, a system that collects real-world dexterous-grasp data without continuous human operation. Given generated grasp candidates, AutoDex plans and executes grasps, labels their outcomes, and resets the object to continue collection across stable poses. Multi-view object tracking supports labeling under hand-object occlusion, while residual-torque collision monitoring and pose-directed reset enable the loop to operate without continuous human supervision. Using AutoDex, we collect over 3,500 grasp trials across 100 objects with Allegro and Inspire hands, recording both successes and failures together with synchronized visual observations and robot states. AutoDex reduces human effort by more than an order of magnitude, requiring an estimated 0.2 human-hours to collect 500 trajectories compared with 9.3 for teleoperation. As AutoDex collects trials, it builds a library of physically validated grasps for subsequent execution. In new object placements and obstacle configurations, retrieval from this library achieves 76% success, compared with 34% for candidates screened only in simulation. Code and data are available at https://snuvclab.github.io/AutoDex/.
Benedikt Seiter, Anya Fries, Julius von Kügelgen, Jonas Peters
Principal component analysis (PCA) is one of the most widely used unsupervised dimension reduction techniques. We study PCA for data from multiple related domains. Since principal components generally differ across domains, one way to obtain a shared low-rank embedding is to perform PCA on the pooled data. However, this approach can focus on spurious directions that exhibit high variation in only a few domains. To find a robust embedding that still explains most variance in unseen but similar domains, we propose instead to focus on shared directions of variation. To this end, we introduce Anchor PCA which trades off overall explained variance with agreement between the shared and domain-specific low-rank embeddings. Anchor PCA amounts to PCA on a modified target matrix and thus can be solved efficiently. Moreover, we show that Anchor PCA recovers a maximal invariant subspace and admits a minimax reconstruction interpretation under bounded domain-specific covariance inflations. On simulated and real-world data with temporal distribution shift, we demonstrate, respectively, that Anchor PCA recovers the maximal invariant subspace and yields embeddings that explain more variance on unseen domains than the pooling baseline and a worst-case alternative. Taken together, these findings establish Anchor PCA as a promising approach to robust unsupervised dimension reduction from multi-domain data.
Recent text-to-image (T2I) generators can synthesize realistic images, but still struggle with compositional prompts involving multiple objects, counts, attributes, and relations. We introduce EPIC (Efficient Predicate-Guided Inference-Time Control), a training-free inference-time refinement framework for compositional T2I generation. EPIC casts refinement as predicate-guided search: it parses the original prompt once into a visual program of object variables and typed predicates, then verifies each candidate image against this program using visual evidence extracted from that image. A candidate is judged to satisfy the prompt when all predicates pass; otherwise, unmet predicates guide the next step, routing localized failures to targeted editing and layout-level or persistent failures to resampling. On GenEval2, EPIC improves the prompt-level Soft-TIFA GM score from 34.16% for single-pass generation to 71.46%. With the same generator, editor, and image-model execution budget, EPIC outperforms the strongest baseline by 19.23 points. It also uses 31% fewer image-model executions, 72% fewer multimodal large language model (MLLM) calls, and 81% fewer MLLM tokens per prompt.
Eric Balkanski, Paul Dütting, Federico Fusco, Silvio Lattanzi, Ashkan Norouzi-Fard, Ola Svensson, Morteza Zadimoghaddam
Online algorithms with bounded recourse face three intertwined obstacles: uncertainty about the future, computational limits, and a limited budget for changes. We ask a fundamental question that separates consistency from the other two: what does solution stability cost when the entire future is known and computation is unlimited? We answer this question for monotone submodular maximization, where an algorithm must maintain a solution of at most $k$ arrived elements, with a constant number of changes per step. We determine the optimal approximation thresholds: $3/4$ for deterministic algorithms and $ξ^\star \approx 0.7942$ for randomized algorithms. For every $\varepsilon>0$, we get within $\varepsilon$ of these thresholds with worst-case recourse polynomial in $1/\varepsilon$, independent of $k$ and the stream length. We prove matching impossibility results even for coverage functions. We attain the sharp bounds by optimally balancing two opposing forces: preparing for future arrivals and protecting the current solution. For randomized algorithms, we use addition robustness against prefixes of the known future, which we analyze tightly via an interval-selection problem. For deterministic algorithms, we introduce certificates, paths in a fractional relaxation that maintain high value, which we round into an update schedule.