Papers

Software agents built on large language models act by issuing tool calls, some of which cannot be undone. To control theory the model is an opaque feedback controller, replaced every few months, whose closed loop cannot be analyzed directly. We propose a runtime filter, or shield, between any such controller and its environment that certifies a bound on the probability of reaching an unsafe condition during a bounded task. Given a predicate abstraction and candidate successor sets, everything else comes from logs: the loop becomes a finite controlled Markov chain, the estimated kernel is surrounded by a distribution-free confidence region, and robust reach-avoid value iteration runs over it. The certificate is probably approximately correct and uniform over controllers respecting the filter, so one model may replace another without recomputation. In all tested datasets the true risk stayed within certificate and budget; in a file-system sandbox the filter stays sound against a hazard the logs barely covered, where a plug-in baseline certifies safety and fails; and it holds seven controllers, three of them open-weights models, to zero unsafe outcomes in 1050 episodes.

Organizations pairing a human and a machine on a task decide, often by default, which executes each instance and what happens when both do. We account for what this attains under the squared rule, with performance distributed over instances to capture generalization difficulty. A gross stake lies between the skills' mean and best per instance. It splits at the best single skill into a systematic part, known from separate assessments and secured by commitment, and a residual, the expected value of perfect information on which skill each instance favors, known only from paired observation and split further by classes. Skills set the stake; workflows decide the take, via two channels. Routing directs work, weighting or selecting responses; with binary truth it is capped by the best skill per instance. Reading responses as joint evidence is not routing; what they jointly reveal caps it. Synergy lets skills interact before responding; bounded by no skill either way, its lift over independent responses must clear a break-even and can fall below one. As algorithms, arrangements form a ladder stating what each reads, holds and records, and a procedure whose checks keep earlier gains when class means are known and lose little on finite records given a margin. The theory extends to many skills, each worth what it adds where it beats all others held, and to competing skills, where at parity a single failure's allowed cost decides. In a public chest-radiograph study with 227 radiologists, they are near parity with the model and the stake is mostly residual; assisted reading, no better on average originally, returned less than averaging both responses; recalibrating the model alone beats any deployable arrangement of raw responses. Organizations can learn from their records what is at stake, which arrangement collects it and what evidence each step needs; pricing costs and errors is future work.

Optical music recognition (OMR) for medieval chant lacks training data linking manuscript images to musical symbols across notational families. Using existing Volpiano transcriptions from the Cantus Database, we created 30,567 image-target pairs from ten sources in eleven exports, including 339 verse fragments. Their target sequences contain 1,952,143 tokens. The pipeline locates staff systems, infers clefs and maps known pitches to staff positions. Six detector families handle the different layouts and notations. Where heuristics failed, two small U-Nets learned staff lines and clef letters from manual annotations. Reviews of sampled export lines found zero to four pitch or clef errors, but used different error definitions and do not provide a uniform audit of the corpus. A CRNN trained from scratch achieved a symbol error rate of 9.6 % on 3,086 test lines. The earlier model, trained mainly on Austrian sources of the 15th and 16th centuries, scored 45.3 %. In three separate training runs, each excluding one manuscript, error rates on the excluded manuscript were 21.5-31.0 %. Model-assisted alignment recovered a further 2,686 lines.

This paper presents GpFlash, a Gaussian process library that is focused on fast training and posterior evaluations, minimal memory usage, and scalability to high-dimensional problems. GpFlash was compared to the popular GP libraries scikit-learn, GPy, GPflow, GPyTorch, and SMT on a MacBook Pro and the supercomputer Trillium. GpFlash was found to use less memory and to be faster for training and posterior evaluations relative to the other GP libraries for all cases considered. For example, when using $\nx = 10^4$ training points in a 10-dimensional space GpFlash required 9.6 to 35 times less memory and was 2.6 to 54 times faster than the other GP libraries for the evaluation of the marginal log likelihood and its gradient. The memory reduction was achieved by only using one matrix of size $\nx \times \nx$ to store different symmetric matrices in its lower and upper triangular entries. The computational speedup was the result of designing algorithms to access memory efficiently, using BLAS and LAPACK methods extensively, and factoring kernel equations so that expensive exponential calculations could be reused. GpFlash was more than 440 times faster than all of the other GP libraries considered when evaluating the posterior mean together with its gradient for the case with $10^4$ training points and 10 dimensions. GpFlash is available at https://github.com/marchildon/gpflash.

How the relationship between image-text aligned and vision-only models changes across development remains unknown. We applied time-resolved representational similarity analysis to EEG data from 8-year-olds, 12-year-olds, and adults. CLIP showed significant model-brain alignment already in children, whereas vision-only models aligned more strongly later in development. Residual analyses revealed complementary structure, with CLIP retaining alignment beyond an architecture-matched vision-only model from age 8. Much of the model-brain alignment reflected semantic structure, with EEG representations of category and animacy strengthening toward adulthood, while alignment with low-level image properties remained weak across age. Together, these findings suggest that image-text aligned and vision-only models capture complementary aspects of developing visual representations and may provide useful constraints for AI systems that aim to learn and adapt over time.

For repository-grounded coding agents, a functionally correct patch is not necessarily an acceptable one: developers also constrain how the task must be carried out-modification scope, workflow order, validation, tool use, and reporting. This capability, which we call engineering constraint following, remains largely invisible to current evaluations. We introduce ANCHOR (Assessing Naturalistic Constraint Hierarchies and Ordering in Repository-Grounded Agentic Coding), a benchmark built on naturalistic, multilingual repository tasks spanning eight language groups. ANCHOR combines (i) a hierarchical taxonomy of 10 categories and 78 leaf types distilled from real developer interactions, (ii) an automated pipeline that instantiates task-grounded atomic constraints and composes them into typed AND/THEN/PRIORITY relations, and (iii) a trajectory-grounded evaluation framework that checks both the code diff and the execution trajectory with rule-based and LLM judges. Across 19 agent configurations spanning 2 scaffolds and 10 foundation models on 358 tasks, the strongest measured system, Claude Opus 4.8 under Cline, reaches 70.1% atomic compliance and 63.7% relation compliance, while 19.0% of tasks pass every scored checklist item under the strict all-item FCCR criterion. A five-engineer audit of constraints from 50 instances reports 94% feasibility, 94% repository consistency, and 96% realism. In the primary paired set (14 configurations), the mean signed change in task completion after adding constraints is +0.7 percentage points; individual changes vary in both directions. These results distinguish functional success from constraint compliance, with failures concentrated in workflow ordering, tool-use discipline, and reporting, and substantial interactions between models and scaffolds.

Target speaker extraction (TSE) aims to recover a target speaker's speech from a speech mixture using an enrollment utterance. Recent codec-based generative approaches formulate TSE as conditional speech token generation and achieve promising perceptual quality. However, most existing systems rely on unified codec tokens that entangle linguistic content, speaker characteristics, and acoustic details, forcing a single model to jointly handle target selection, content semantic, and acoustic reconstruction. In this paper, we investigate whether disentangled semantic-acoustic codec representations are better suited to TSE. We first show that simply replacing unified codec tokens with semantic-acoustic tokens is insufficient and can even degrade performance when both remain handled by a unified generative model. Motivated by this observation, we propose a task-aware discriminative-generative framework, where a discriminative branch extracts target-related semantic information, while a generative branch predicts coarse acoustic tokens and refines them under semantic guidance. This specialization separates what is said from how it sounds while coordinating both factors in a factorized codec space. Experiments on Libri2Mix show that our method reduces dWER from 0.388 to 0.118 compared with a unified-codec baseline, while maintaining competitive perceptual quality and speaker consistency. Ablation studies confirm that the gains arise from task-aware semantic-acoustic modeling rather than representation disentanglement alone.

Research on video classification models has witnessed significant growth recently, accompanied by a significant increase in the demand for datasets to train such models. Existing video dataset curation approaches share a common bottleneck: candidate videos must be downloaded before any relevance judgment is made, which results in storage, time and annotation costs. In this work, we present a filtering step, using Large Language Models (LLMs) to pre-screen videos based on their metadata, before actually downloading these videos, thereby saving on these costs. We evaluate the LLMs against a strict human majority vote derived from four independent annotators and validate the reliability of the annotations through inter-annotator agreement analysis (Gwet's AC1). We compare the LLM pre-filter against two no-LLM baselines, namely two keyword heuristics. Via a Leave-One-Rater-Out (LOO) mechanism and an Add-One-Rater-In (AORI), we evaluate the performance of the LLMs as virtual annotators. Finally, we report cost comparisons relative to existing approaches. To ensure reproducability, we made our implementation available on github.

Real-time bimanual humanoid teleoperation requires accurate wrist tracking while keeping arm configurations predictable. Because a 7-DoF arm admits multiple elbow configurations for the same wrist pose, wrist-only tracking can cause elbow drift or abrupt configuration changes. Fixed anthropomorphic constraints can stabilize elbow posture, but they can also block the redundant adjustments needed near singular configurations. We propose a human-like redundancy-resolution framework that combines adaptive elbow guidance with singularity-robust wrist tracking. Shoulder-wrist geometric retargeting generates robot-scale wrist targets. An adaptive elbow-offset criterion uses the offset direction to guide posture and scales this guidance by the offset magnitude, so the redundant degree of freedom is progressively released as the human arm straightens. A hierarchical velocity solver prioritizes wrist tracking and projects the adaptive elbow guidance into the wrist-task null space. In a fixed-wrist-target ablation, adaptive guidance reduces mean elbow-direction error by 54.4% compared with removing this guidance. In a workspace-boundary stress test, singularity-robust tracking reduces peak raw joint-velocity demand by 97.9% compared with the undamped pseudoinverse. Real-world bimanual tasks and cross-platform retargeting show applicability across robot morphologies.

We derive explicit formulas for the dual Moore–Penrose generalized inverse (DMPGI) of dual complex matrix sums under some conditions. We first express the DMPGI of a matrix partitioned into column blocks in terms of these blocks and associated projectors, and then derive two formulas for matrix sums. Two examples illustrate the results. We further apply the formulas to rigid-body motion estimation with two observation points subject to a first-order perturbation of their positions that preserves the rank-deficient observation geometry. The DMPGI recovers the observable motion components from the velocity data.

Diffusion and flow-matching models may struggle to accurately represent distributional tails, especially for heavy-tailed data. Existing approaches often use heavy-tailed noise or modify the generative dynamics, requiring dedicated retraining and more difficult estimation. We take a simpler approach: rather than relearning the full distribution, we correct only the model's behavior in the extreme regime. Our method builds on the radial-angular decomposition of multivariate heavy tails, where radius and direction become asymptotically independent. This allows us to preserve the directional structure learned by the pretrained model and correct only its radial tail. We estimate the tail onset from reference and generated samples, then rescale generated extremes so that their radial survival matches the target distribution. The resulting method is a lightweight post-training correction applied to a frozen diffusion or flow-matching model, requiring neither retraining nor changes to its sampling dynamics. Experiments on challenging synthetic heavy-tailed distributions and real weather and wildfire datasets show that this simple correction substantially improves extreme-tail coverage using only a small number of fitting samples.

Face recognition systems often experience performance degradation under large pose variations due to changes in facial appearance and visibility of discriminative features. This study presents a multi-task deep learning framework that jointly addresses pose-invariant face recognition and facial landmark localization within a unified architecture. To enable simultaneous optimization of identity recognition and facial structure learning, the original Cross-Pose Labeled Faces in the Wild (CPLFW) dataset was adapted into a multi-task learning dataset by associating facial images with identity labels and five-point landmark annotations. The proposed framework consists of two task-specific branches operating on a common facial input: a ResNet50-based branch for identity recognition and an EfficientNetV2B0-based branch with spatial attention for facial landmark localization. The two objectives are jointly optimized using a weighted multi-task loss, allowing the model to learn complementary semantic and spatial facial representations. The proposed framework was evaluated on the prepared CPLFW-based multi-task dataset. Unlike the original CPLFW benchmark, which focuses primarily on identity recognition, the proposed dataset adaptation incorporates five-point facial landmark annotations to enable joint recognition and localization learning. Experimental results demonstrate that the model achieves a validation recognition accuracy of 81.58% and a facial landmark localization mean squared error of 96.38. The results indicate that combining identity classification and landmark regression within a single optimization framework provides an effective approach for learning robust facial representations under cross-pose conditions The proposed method demonstrates the potential of multi-task learning and attention-based feature refinement for unconstrained facial analysis.

We prove that, for every fixed rational $c>0$, computing a $2^{-n^c}$-approximate Nash equilibrium in an $n$-player two-strategy anonymous game is $\mathsf{PPAD}$-complete. This holds even when one strategy always pays $1/2$ and, for all but two players, the other is constant outside at most five opponent counts. The reduction uses coefficient estimates to simulate circuit constraints approximately. Exact computation requires eliminating the simulation error: with three strategies, a separate construction uses polynomial roots to implement circuit gates exactly. This yields $\mathsf{FIXP}$-completeness of exact equilibrium computation under separable-linear reductions and $\mathsf{FIXP}_a$-completeness of approximation in distance to an exact equilibrium.

Standardizing a rare feature can help stochastic gradient descent learn its effect within a finite budget, but it also changes the noise cost of learning. We characterize the lowest rare-subgroup risks attainable by standardized and unscaled training when each chooses a nonnegative constant learning rate and a deterministic stopping time under the same observation cap. For two independent binary features, we derive both optimized risks as prevalence vanishes while the expected number of rare observations remains finite. Uniform control over rates and stopping times reduces each limiting oracle to a scalar minimization. With positive noise, the preferred scaling changes exactly once as we move signal to the rare feature while the limiting risk at initialization stays fixed and positive. The formulas also explain how noise and budget affect the comparison. Separate finite-prevalence bounds favor standardization on a rare-dominant signal family and unscaled training on a common-dominant family. The results concern final iterates with population preprocessing and fresh observations. Exact-moment searches assess the approximation at finite prevalence. In a controlled study, both orderings persist under observation reuse, estimated preprocessing, and intercept fitting when tuning choices are fixed before evaluation. Selecting a predictor after observing training can reverse an ordering, even when selection uses exact population risk.

Real-world systems often serve many objectives but can deploy only a few options, whose performance must be learned from noisy feedback. Modeling each option as an arm in a multi-objective bandit, we introduce few-for-many best-arm identification, which seeks a small set of arms minimizing the worst weighted degradation from the objective-wise optima, rather than the best arm for every objective. We consider both known objective groups and latent groups specified only by a representative budget, and propose GFM-Elim and AFM-Elim with instance-dependent sample-complexity guarantees. For Gaussian rewards, we further derive a lower bound for the latent-group setting and show that it is asymptotically attainable with high probability by an oracle-based Track-and-Stop procedure. Our results show that sharing observations across objectives and ignoring irrelevant objective-wise distinctions can substantially reduce the sampling cost, as confirmed by experiments on synthetic and real-world tasks.

In-context learning (ICL) applies large language models to new tasks by prepending a few labeled demonstrations to a query. Prior work has studied how demonstration ordering affects ICL, but the interaction between label arrangement patterns and model biases has not been systematically examined. We define label collapse as the failure mode in which the model generates query-independent labels, and show that it is architecturally universal at the tested 8-9B scale: five models spanning autoregressive (Llama-3, Qwen3.5) and discrete diffusion (LLaDA, iLLaDA, LLaDA-Instruct) families all exhibit label collapse even under the class-balanced alternating arrangement, with some label word pairs collapsing in virtually every tested condition and others flipping their collapse direction with arrangement. Critically, accuracy alone cannot detect this collapse: models can achieve near-chance accuracy while generating query-independent labels, a score indistinguishable from balanced performance, so the failure can persist undetected in deployment. We measure separation, the log-odds of the two label words at the query position (the quantity that label-bias calibration operates on), as a complementary diagnostic that exposes collapse hidden beneath surface accuracy, together with a measurement protocol that guards label extraction against tokenizer and formatting artifacts. For autoregressive models, collapse persists where the majority-label-bias account does not apply, and arrangement patterns can reverse the collapse direction; for diffusion models, we present a systematic evaluation of ICL failure modes, complementing contemporaneous studies of dLLM context comprehension and demonstration-order sensitivity, and identify open questions about how iterative denoising interacts with label biases.

As language-model agents write code, move money and change records on behalf of users, deciding whether an action is authorized has become a central safety problem. Guardrails are often placed in front of these actions, yet few evaluations ask whether their decisions follow authorization: blocking unauthorized actions while still allowing the same actions under valid permission. We introduce \bench{}, a benchmark that contrasts authorization evidence for the same operation in its language track. It combines expert-annotated natural-language scenarios with executed and verified operations, plus externally sourced native-policy requests. \bench{} checks each decision for correctness, usability and resulting effect, and tracks each request across guardrail changes. Our experiments show that many blocked operations stem from failed reviews, not correct denials. Repairing the output format restores usable decisions but changes which authorized requests are honored, and adding a second reviewer removes some errors while rejecting valid authorization. Guardrails should therefore be evaluated on both the unauthorized actions they stop and the authorized actions they let through.

Syntactic language models (SLMs) that generate sentences along with their syntactic tree structures have shown promise in both syntactic generalization and downstream tasks. Prior studies, however, have mostly evaluated smaller models trained on less pretraining data and tested on narrower task suites. In addition, the design space of tree linearization and attention masks remains underexplored. We conduct a large-scale empirical study of 12 controlled configurations—one terminal baseline and 11 SLM variants spanning different linearizations and masks—at scales up to 1B parameters and 100B training tokens, evaluated across 11 downstream tasks and two syntactic benchmarks. The experimental results show that SLMs with standard causal attention match or exceed structural-mask variants across downstream and syntactic benchmarks. They require no architectural modifications and remain compatible with FlashAttention. We also observe that when scaling up to the 1B-parameter scale, the top-performing SLM variant no longer shows a clear advantage over a compute-matched non-structural baseline, suggesting that SLMs may become less competitive with scale. Our code is available at https://github.com/qlwpc/TG-Interpolation.

Model providers can support image protection, but open-weight deployment introduces practical constraints. We propose four requirements: protection effectiveness, clean fidelity, compatibility, and tamper resistance. We incorporate long-horizon attack simulations to strengthen resistance to protection removal. Experiments with a single protection tool show improved resistance to re-alignment at a cost to clean fidelity, motivating further research toward practical model-side protection.

We study the two-player communication complexity of the Coin problem. Here, two players, denoted Alice and Bob, are given up to $n$ individual i.i.d. random coins, and asked to output whether there are more heads or tails. An immediate approach is for Alice to send Bob her exact numbers of heads and tails, yielding a deterministic protocol with one-way communication cost $O(\log n)$ bits. Indeed, this is provably optimal in the related multi-player and streaming settings [Braverman, Garg, Woodruff, FOCS'20]. Our contribution is to show that this is not the right answer in the one-way two-player setting. Instead, we prove a $\Omega(\log\log n)$ bit lower bound for the constant error-rate regime, and a $\Omega(\log(1/\delta))$ bit lower bound for arbitrarily small error rates, $\delta \geq n^{-0.1}$. We also complement these with a tight one-way protocol using $O(\log\log n + \log(1/\delta))$ bits of communication. Unlike the previous works, our lower bounds avoid intricate direct information-theoretic arguments. Instead, our $\Omega(\log(1/\delta))$ bound is via a standard fooling-set style argument, whilst our $\Omega(\log\log n)$ bound is via a novel yet simple reduction from the Gap-Greater-Than problem. This reduction is our main technical contribution, and can be seen as loosely inspired by the Robust Communication Complexity arguments of [Chakrabarti, Cormode, McGregor, STOC'08]. Finally, we also show that our constant-error lower bound and protocol can be adapted to the multi-round setting: up to polynomial factors in $r$, the $r$-round communication complexity of the coin problem is $\Theta\left((\log\log n)^{1/r}\right)$ bits.

When does a shared representation become insufficient for jointly supporting multiple tasks? This question lies at the heart of multi-task learning models. Although task interference in shared representations has been widely observed, its onset, measurable form, and capacity boundary remain poorly understood. In this paper, we revisit this problem through the lens of Bell's theorem. We show that if four binary task outcomes can be jointly represented by a common shared representation, their correlations must satisfy the classical CHSH bound $|S|\leq2$. We introduce NCnet, a controlled multi-task model for studying how representational capacity affects task compatibility. Experiments reveal a clear capacity-dependent transition. Once nontrivial task fitting emerges, constrained capacity can yield $|S|>2$. As capacity further increases, these exceedances progressively disappear and $S$ returns to the bounded region. This systematic pattern provides evidence that limited representational capacity is an important source of the observed incompatibility. We observe similar transitions across CNNs, PLMs, and LLMs, suggesting that the phenomenon extends beyond the controlled NCnet setting. We refer to this capacity-dependent transition as the Bell Boundary, providing a measurable diagnostic for representation bottlenecks, task incompatibility, and capacity allocation in multi-task models.

Deepfake speech detectors are typically evaluated as binary classifiers on fixed benchmarks, but deployed systems face evolving attacks, codecs, speakers, and recording conditions, with labels unavailable during operation. This makes it difficult to know when a detector operates in familiar conditions and can be relied on. We propose an unsupervised drift monitor that operates on the internal embeddings of a deployed detector. The monitor compares each input against bona fide and spoof references using regularized Mahalanobis distances. These distances define trusted, ambiguous, and novel regions of the detector embedding space, allowing unfamiliar inputs to be routed for review. We evaluate the monitor on four modern detectors on three datasets. The monitor raises alerts for dataset mismatches and substantially reduces the security-critical case of accepting deepfakes at the cost of reduced automatic coverage. The results also show that drift affects both classes: bona fide recordings can move away from the training reference, increasing alert rates even in in-domain settings. Our findings suggest that deployed deepfake speech detectors should be treated not only as a binary classification, but also as a reliability problem.

A release decision about a fine-tuned model is an operational commitment about the next rerun of its training recipe. Yet standard benchmark reporting gives only the mean accuracy of past seeds. These target fundamentally different questions: release decisions depend on catastrophic tail risk, whereas standard evaluation reports only central tendency. Because fine-tuning induces bimodal mode collapse, high average accuracy frequently conceals severe failure rates. Auditing $2\,124$ fine-tuned models across eleven language models, we show that shipping models based on mean accuracy violates a $10\%$ deployment failure budget $48.0\%$ of the time. Common heuristics like $\bar b - 2\sigma$ prove equally dangerous: by underestimating variance on small samples, they become most permissive when evidence is weakest. We introduce recipe cards, an evaluation framework that directly certifies rerun tail failure rates using exact finite-sample binomial bounds. Recipe cards reduce held-out budget violations to $0.6\%$, enforce a mathematical sample floor ($N \ge 22$) that prevents certification on thin evidence, and explicitly refuse unverified claims under domain and task shifts.

Architecture attribute predictors reduce the cost of neural architecture search by estimating accuracy or latency from a small set of evaluated candidates. The key difficulty is that neural architectures are operation-labeled DAGs: edge direction defines computational flow, but existing predictors often treat direction as a local structural cue. We propose DTA-GT, a Direction- and Topology-Aware Graph Transformer based on encoder-wide directional consistency, which preserves directed computational semantics across node initialization, global spectral encoding, pairwise interaction, and feature update. DTA-GT realizes this principle with direction- and topology-aware node initialization, Magnetic Laplacian spectral encoding, direction-dual structural attention, and a Direction-Sensitive MoE-FFN. Across accuracy and latency prediction on NAS-Bench-101 and NAS-Bench-201, DTA-GT consistently outperforms representative sequence-, GNN-, Transformer-, and hybrid predictors under limited-budget settings. Ablations and Magnetic spectral controls show that the improvements come from coordinated direction-aware stages and edge-orientation-sensitive spectral information, while mechanistic diagnostics confirm that the modules exhibit the intended directional behavior. These results support encoder-wide directional consistency as an effective design principle for architecture DAG representation.

In simulation-based inference (SBI), neural posterior estimators trained on a bounded range of sample sizes can become inaccurate on larger datasets because their parameterizations do not specify how posterior location, uncertainty and shape must change as observations accumulate. We propose center-estimated NPE (CE-NPE), which combines a statistical center with an explicit contraction rate and learns a local center correction and residual densities in posterior-scale coordinates. The center is a statistical estimate, such as a closed-form statistic or a simulator-based moment match, which does not require an evaluable likelihood. Our analysis explains how contraction amplifies center error and decomposes Gaussian posterior error into shape, mean and covariance contributions. On six regular benchmarks, CE-NPE gives the best reference-posterior agreement of all tested amortized estimators at every out-of-range size, up to five times the largest training size, without retraining; it also fits better within the training range on most of them. Matched controls and nonregular examples characterize the scope of these gains and identify the conditions on which they depend: an appropriate contraction rate, an accurate center, and a stable residual representation.

Matching one distribution to many targets can require repeated evaluations of every target. CAL-$\alpha$-MODM aggregates a nonlinear predictor into one barycenter, calibrates its residual derivatives, and checks steps against the original $f$-divergence objective. Under smoothness assumptions, calibration controls parameter-step model error without requiring targets to be close to the model. Exact-oracle stationarity and local convergence guarantees are established. At common matching, canonical geometry cancels the leading quadratic iteration error, giving cubic convergence with second-order calibration and sufficiently accurate inner solves. High-precision experiments support this prediction. Sensitivity-based initialization also accelerates preference continuation along regular stationary branches. On 300-target Jensen–Shannon problems, CAL-p2 achieves median paired speedups of $3.15\times$ over L-BFGS and $1.10\times$ over second-order Direct-TR, across 23/25 and 24/25 objective-comparable stationary pairs. The savings arise when cheaper predictor queries offset calibration overhead.

Progressive depth growth builds large networks by starting with shallow models and inserting layers during training. Recent advances in large language models have renewed interest in this longstanding approach, demonstrating benefits for training efficiency and reasoning alongside changes in internal computations. However, how growing affects representation quality and computational patterns in self-supervised vision transformers remains unclear. In this work, we systematically investigate progressive depth growth in two popular self-supervised families, namely masked autoencoders (MAE) and DINOv3, and analyze how growing affects downstream performance, intermediate representations, and internal computations. Across global and dense tasks, we show that growing improves downstream performance while reducing training FLOPs by 13.9–24.8%, with ImageNet linear-probe accuracy gains of up to 3.36 percentage points. Growing also improves downstream performance at intermediate layers and reshapes internal computational patterns, introducing recurring alternations between global and localized attention in MAE and earlier local-to-global attention progression in self-distilled DINOv3. Together, these findings establish progressive depth growth as an effective and efficient pretraining strategy for improving downstream performance across global and dense visual tasks with less computation, and advance our understanding of how growing shapes visual computation.

Global reanalyses are too coarse to resolve the convective systems that produce most West African monsoon rainfall. In this work, we use U-Net based architectures to downscale daily ERA5 fields (0.25{\deg}) to IMERG V07 satellite precipitation (0.1{\deg}) over West Africa for the June to October seasons of 2000 to 2024, using geopotential, temperature, humidity and wind at three pressure levels together with elevation and land fraction. Adding ERA5 precipitation as input slightly improves performance for this particular challenging region. We find that all the deterministic networks investigated for this work are too dry and lose almost all observed variance below a few hundred kilometers, which motivated our use of a GAN-based alternative. We also investigated permutation importance for the U-Net that ranks 500 hPa specific humidity first, consistent with the sensitivity of deep convection to mid-tropospheric moisture.

Calibration has been extensively studied for classifiers, but remains largely unexplored for crowdsourcing aggregation methods, which are increasingly used for making probabilistic decisions. To the best of our knowledge, we provide the first theoretical and empirical study of calibration in label aggregation. To our great surprise, we find that widely used aggregation methods can be severely miscalibrated, and in opposite directions: Majority Vote is underconfident while model-based aggregators are overconfident. We prove that the Bayesian aggregator is perfectly calibrated under its data-generating model. We also show that exact calibration of weighted majority votes requires highly restrictive conditions, which helps explain why Majority Vote is generally miscalibrated in realistic crowds. Empirically, we study calibration across $39$ human crowdsourcing datasets and LLM-based crowds. In both cases, we observe systematic miscalibration for aggregation models such as Majority Vote, Dawid–Skene, GLAD, and CrowdFM. Finally, we propose a post-hoc calibration procedure that aggregates the aggregators: for any aggregation function $\phi$, we introduce a transformation that first combines its predictive distribution with those of complementary aggregators, before applying temperature scaling. This simple and interpretable transformation makes temperature scaling more effective than when applied directly to $\phi$. It improves calibration, especially in the high-redundancy regime where there are many annotations per item, while generally preserving or even improving accuracy.

Large language models (LLMs) are increasingly used as judges to assign scalar scores that drive downstream decision-making, including tasks such as paper review and recommendation. However, LLM-as-a-Judge pipelines are vulnerable to score-hijacking prompt injection embedded in evaluated content, which corrupts scoring signals and distorts induced rankings. Existing defenses often enforce robustness by modifying the judge model itself, making safety dependent on a specific model's scale and training and requiring repeated effort when the judge is updated or replaced. We propose StableJudge, a lightweight defense framework that externalizes prompt-injection mitigation as a reusable pre-scoring capability. StableJudge centers on a compact 0.13B Sentry that localizes injected spans and performs targeted sanitization prior to evaluation, and is trained using adversarially evolving attacks to improve robustness beyond a fixed prompt set. To stabilize measurement, StableJudge derives deterministic, fine-grained scalar scores from the judge's output distribution. Across held-out scientific abstracts and out-of-distribution attack variants, StableJudge improves score recovery under attack while largely preserving general-purpose performance, demonstrating that robust LLM-based judging can be achieved by decoupling safety from judge scale via reusable defense capabilities. Code and data will be released.