Visual Question Answering (VQA) imposes substantial computational overhead on vision-language models (VLMs), motivating model-agnostic input reduction methods that can be readily applied across diverse VLMs, including in black-box settings. Simply enforcing a visual-token cap reduces computation, but it can discard information needed to answer the question. A single reduction strategy also struggles to accommodate different scenarios. Our key observation is that cropping and compression are complementary: cropping highlights small targets and local details, whereas whole-image compression better preserves context for large targets or distributed evidence. Based on this observation, we propose C$^2$-Routing, which adaptively selects cropping or compression according to the input image and question. Since both strategies may still discard key information, we further design a safety gate that uses the predicted joint correctness distribution under original and reduced inputs to guide fallback and protect accuracy. To accommodate consumer GPU resource constraints, all training and evaluation experiments run on RTX 3090 GPUs. To measure gains beyond token capping, we use full-image baselines with the same cap of 1,024 visual tokens. We conduct extensive evaluation on three VLMs and 18 datasets spanning local details, global scenes, scene text, and documents. With the visual budget already capped, C$^2$-Routing further reduces end-to-end FLOPs by 36.53% and visual tokens by 42.36% on average, with a mean relative task-score decrease of only 0.77%.
We study Bayesian inference of persistence from early observations of a heterogeneous susceptible--infected--susceptible (SIS) reaction--diffusion model with standard incidence. We show that the leading prevalence dynamics depend on the unknown coefficients only through the difference between the transmission and recovery rates and the infected diffusivity. The same pair determines the principal eigenvalue whose sign governs persistence, so early prevalence can inform the threshold even when the individual rates remain uncertain. We prove global weak well-posedness and quantitative parameter stability for the SIS system and use the resulting observation-map estimates to establish Bayesian well-posedness. For an initial infection of size $\varepsilon$, we derive a uniform second-order expansion that describes how the separate rates enter prevalence. Incidence depends explicitly on transmission at leading order. When the noise standard deviation is proportional to $\varepsilon$, we prove $O(\varepsilon)$ convergence in Hellinger distance of the nonlinear posteriors to their low-prevalence limits for both prevalence alone and joint prevalence and incidence data. We bound the principal eigenvalue using the growth rate of the infected $L^2$ norm between two times. This rate gives a lower bound for the full nonlinear system, and the upper bound separates spatial transients from nonlinear saturation. Uniform spectral estimates and bounds on infection size and measurement error give sufficient conditions for persistence and extinction. We also bound the error in posterior persistence probabilities under forward and eigenvalue approximation, accounting for posterior mass near the threshold. Numerical examples illustrate threshold inference and the predicted \(O(\varepsilon)\) convergence of the nonlinear posteriors to their low-prevalence limits in Hellinger distance.
Policy optimization is central to language model post-training, yet principled guidance remains limited. We revisit its theoretical foundations to identify algorithms that learn near-optimal policies using total samples and gradient steps governed by the base model's response coverage, a measure of how readily it generates optimal responses. We call such algorithms coverage-preserving. This guarantee captures the goal of amplifying capabilities accessible through sampling from the base model. (1) Coverage collapse. REINFORCE, the basis for PPO and GRPO, can drive the probability of optimal responses to near zero, even as reward improves. This coverage collapse stalls optimization, and convergence can require exponentially many steps despite good initial coverage. The failure arises even with exact gradients and persists under KL regularization to the base model. (2) Reference resets. We show that repeatedly resetting the KL reference to the most recent policy mitigates coverage collapse. We formalize this intervention in Approximate Policy Mirror Descent (APMD), a meta-algorithm that approximately solves a sequence of KL-regularized RL subproblems. APMD enables stronger regularization, making each subproblem easier to solve, while reference resets allow convergence to global optimality. (3) Coverage-preserving algorithms. For softmax linear policies over pretrained features, we use APMD to establish the first global convergence guarantees showing that widely used families of post-training algorithms are coverage-preserving. These include stylized PPO and GRPO variants with explicit KL regularization and reference resets, and squared-loss algorithms that implement them implicitly, including those used to train the Kimi models. These results clarify how regularization, clipping, and objective choice affect the sampling and optimization costs of post-training.
AUC and the Brier score are standard summaries of a binary risk score, yet they can rank models differently. We introduce a weighted AUC, AUC$_\pi$, which weights each pair of observations by how differently a decision maker with probability of acting $\pi$ treats them. We show that all three are built on weighted averages of one curve, the Youden index $J(t) = \mathrm{TPR}(t) - \mathrm{FPR}(t)$ across cutoffs $t$, and that the averages differ in the weights they put on the cutoffs. Each weights a cutoff by how much something rises across it: for AUC, the distribution of all scores; for the Brier score, the true probability $\mathbb{E}[Y \mid S]$; for AUC$_\pi$, the decision maker's probability of acting. Up to normalization, AUC$_\pi$ is the Brier skill score of the ideal forecast with its weight moved to the cutoffs the decision maker cares about. We also show that weighting pairs by $|\pi(s) - \pi(s')|$ is, up to scale, the only pair weight fixed in advance that makes a pair-weighted AUC an average of the Youden curve for every score distribution. As a ranking measure, AUC weights every model the same way. Hand (2009) shows that it can also be read as an average over cost thresholds, and in that reading each model weights the thresholds differently. The numerator of AUC$_\pi$, which we call the policy Youden index, is the expected share of class 1 a fixed policy acts on minus the expected share of class 0. It judges every model by the same policy and so avoids this; the full ratio AUC$_\pi$ does not.
Cross-validation estimates predictive performance by repeatedly splitting the data into training and validation sets, fitting the model on the first and evaluating it on the second. Every observation influences the estimate in two roles, through its own evaluation when it is held out and through the models it helps fit when it trains. The random splits determine its weight in each role. From this representation, we prove an exact finite-sample identity for the variance of the cross-validated estimate. For a given splitting scheme with a common training set size, the variance is a function of these weights, with constants of the model, the metric, and the population. It covers metrics averaged over observations, such as mean squared error, and over case--control pairs, such as the AUC. The identity shows which schemes to use. The variance decomposes into sampling variance, which depends only on the training size, and variance due to the random fold splitting, which arises when the weights change from one draw of the splits to the next. K-fold cross-validation, whose weights are equal across observations by design, is therefore variance optimal at first order; Monte Carlo and bootstrap cross-validation are not. The first-order variance of K-fold does not depend on K, and under reasonable conditions on the model leave-one-out has the smallest variance of any K-fold scheme. For the AUC the same holds for leave-pair-out, which evaluates every case--control pair. Finally, we show the cross-validated estimate is asymptotically normal with the identity's full variance. Together, these results provide a general theoretical foundation for inference on cross-validated performance and establish principles for selecting variance-optimal cross-validation designs.
We use TabPFN, a recently developed tabular foundation model, to predict three properties of energetic materials: impact sensitivity, decomposition temperature and crystal density. Without any hyperparameter tuning, TabPFN outperforms the best published models for all three properties. It also substantially outperforms a wide range of established machine-learning models, including tuned gradient boosting (XGBoost), random forests and Gaussian processes. This holds both when the models are given hand-crafted molecular descriptors and when they are given learned representations from Chemprop's message-passing neural network. In particular, for all three properties, TabPFN applied to Chemprop's own embedding is more accurate than Chemprop's predictions. TabPFN is also among the fastest models we consider, more than an order of magnitude faster than XGBoost. As TabPFN predicts the entire distribution of the target rather than just a single value, it offers built-in uncertainty quantification, which we compare with the empirical distributions of compounds with repeated measurements. Finally, we interpret the predictions using Shapley values, which rank the features by their importance.
UTF-8's role as the de facto encoding standard in language models introduces cross-lingual inequities due to its variable-length design. However, the effect that this encoding choice has on model performance remains largely unexamined. In this work, we investigate how text encoding impacts the performance of byte-level language models, comparing UTF-8 with a wide range of other encoding strategies, including other Unicode-based approaches, language-specific encodings, and our own customized encodings. To facilitate our analysis, we propose Aether: a hybrid encoding scheme that uses script-specific encodings for our target setting while maintaining full Unicode coverage through UTF-8 fallbacks, using this framework to model four orthographically diverse languages. Surprisingly, we find that while languages with information-dense characters (Chinese, Korean) significantly benefit from compressive encodings, languages whose characters carry less information individually (Georgian, Tamil) are largely insensitive to encoding choice. In light of these findings, we argue that text encoding should be a language-specific design decision, rather than a universal default.
Inverse rendering for skin typically relies on offline path tracing to estimate physical properties, which fail to map directly to the approximate shaders utilized in interactive applications, such as game engines. We bridge this gap by directly estimating subsurface scattering parameters tailored for real-time rendering pipelines. We introduce a differentiable Screen-Space Subsurface Scattering (SSSS) framework that enables the joint optimization of spatially varying base textures and scattering distances. To account for local variations in skin thickness, we estimate these parameters across distinct facial regions, generating a blended profile map ready for production runtime engines. Our pipeline utilizes a low-cost, single-light, single-view capture setup designed for accessible and practical appearance acquisition. Optimization is driven by a masked loss function targeting regions where subsurface scattering effects are dominant. We validate our approach on a real-world dataset of $45$ individuals across the Monk Skin Tone Scale, ensuring demographic generalization. Furthermore, we complement this evaluation with synthetic data, a multi-light setup validating our single-light assumption, and a tilted-view configuration confirming angular consistency. Results demonstrate that our method effectively recovers scattering parameters from lightweight inputs, yielding realistic skin appearance directly compatible with standard real-time shaders.
Mixtures of logistic regressions provide a flexible framework for modeling binary responses, extending standard logistic regression to accommodate unobserved heterogeneity while retaining an interpretable regression structure. In this paper, we study such models with an unspecified mixing distribution. We analyze a sieve nonparametric maximum likelihood estimator (NPMLE) and derive nonasymptotic bounds on the estimation error of the resulting mixture regression function under both fixed and random designs. Our results reveal an explicit tradeoff between statistical estimation error and the approximation error induced by the sieve radius, from which we derive convergence rates under several tail conditions on the mixing distribution. For computation, we propose a finite-particle algorithm that approximates the sieve NPMLE by optimizing the locations and weights of a finitely supported mixing distribution. The method offers numerical advantages over existing algorithms and comes with rigorous convergence guarantees. For any fixed dataset, we establish global convergence of the likelihood to the restricted NPMLE value when the number of particles exceeds a finite threshold. When the true mixing distribution is finitely supported, we further establish uniform global convergence guarantees and derive error bounds that separate algorithmic and statistical errors. Our analysis also yields a denoising bound for the mixing distribution and an identifiability result that complements existing results in the literature.
Understanding 4D scenes is a core challenge in visual intelligence, requiring models to jointly capture scene geometry, appearance, and dynamics from partial observations. However, existing approaches are subdivided into feedforward models for direct geometric prediction, and generative models for visual synthesis. We introduce UGG, a unified framework for 4D scene understanding that jointly models RGB video and geometric representations such as depth and camera pose. By assigning different noise levels across modalities and time, UGG formulates visual generation and geometric prediction as conditional completion within a single generative model. This flexible formulation enables the model to learn from diverse, heterogeneous datasets with different subsets of geometric and visual annotations. Furthermore, our generative formulation enables inference-time search over multiple plausible scene hypotheses, using geometric evidence extracted from observed RGB frames to improve geometric prediction without ground truth geometry. We demonstrate UGG's effectiveness over non-unified baselines across camera-controlled video generation, depth estimation, and camera pose prediction, including in challenging sparse observation settings. Code and visualizations are available at https://ugg-model.github.io/
Deepfake detectors are commonly evaluated on real/fake classification alone, but real-world applications require explanations of the visual evidence behind each prediction. Joint vision-language models struggle with authenticity assessment, producing "real" narratives for manipulated images. To address this, we propose a decoupled predict-then-explain pipeline as our submission to the ACM Multimedia 2026 Explainable Deepfake Detection Challenge, with methods trained and evaluated on the XPlainVerse benchmark. A vision-foundation-model ensemble first classifies the image, then a Qwen3-VL-8B model generates a decision-conditioned explanation of the visual evidence. For fake images, a second Qwen3-VL-8B compresses that text into a short non-expert summary without using the image. We further apply metric-aligned Group Relative Policy Optimization (GRPO) to optimize both explainers against the challenge criteria. Under matched conditions, GRPO raises complex BERTScore-F1 from 0.686 to 0.712 and also improves grounding (entity and evidence F1); our submission ranked first on complex BERTScore-F1 and seventh overall.
We describe our solution for the 2026 BEHAVIOR Challenge: 100 long-horizon household tasks in the OmniGibson simulator. We placed 9th on the final leaderboard, the only independent team in the top 10. The core of the solution is a single vision-language-action (VLA) model, one checkpoint for all 100 tasks, built on our 1st-place 2025 solution, which is based on $\pi_{0.5}$. On the 50 tasks shared with 2025, its Q rises from 0.260 to 0.371; across all 100 tasks it scores 0.392. Our main conclusion from 2025 was that the model was undertrained, so most of the 2026 work went into training the model for longer, optimizing some training parameters, making each training step slightly faster and reducing the gradient noise: a migration from 50 to 100 tasks (semantic stages built from the demonstrations' skill annotations, task and stage embeddings of new tasks warm-started from similar old ones, and a 21-dimensional action space matching the teleoperation controls); variance-reduced multi-sample flow matching with a simple noise-vs-compute model that picks the number of action-expert samples per observation without ablations; and a 200k-step training run, about one pass over the 211M frames of the demonstrations. A cloud evaluation infrastructure keeps a shared record with provenance, episode diagnostics and restorable simulator states, and runs six simulators on one GPU together faster than real time. On top of the learned policy, 14 tasks are routed to a GPT-6 Astra coding agent that controls the robot through observation and motion tools, starting every episode from a frozen copy of its own practice session on training instances. Across these 14 tasks, Astra's mean Q is 0.639, against 0.108 for the learned policy. The public submission included only part of the Astra results and scored 0.446 Q with 27.3% full-task success.
Branch-and-Bound (B&B) and $A^*$ are popular exact search methods originating from Operations Research and Artificial Intelligence, respectively. Although the two methods share fundamental search principles, their different ways of exploring the search space can lead to substantially different computational performance. We investigate these differences using the Resource-Constrained Project Scheduling Problem (RCPSP) as an exemplary problem setting. We first establish the relationships between mixed-integer linear programming, state-action, and timed Petri-net representations. Since B&B and $A^*$ admit numerous algorithmic enhancements, we focus on relatively basic implementations. This allows us to examine the fundamental differences between the methods rather than compare the performance of state-of-the-art solvers. By using the same lower bound and best-first node selection, we isolate the effects of the search mechanisms, particularly the identification of equivalent states reached through different scheduling decision sequences. Experiments on 480 benchmark RCPSP instances show that graph-based $A^*$ substantially outperforms the B&B variants considered. With additional pruning rules, $A^*$ proves optimality for all 480 instances, compared with 445 for best-first B&B under the same time and memory limits. Depth-first B&B uses less memory and can find feasible solutions even when the search does not finish within the time limit. Stronger linear programming relaxations reduce the search space but introduce considerable computational overhead. Our findings highlight the importance of duplicate-state detection and clarify the computational trade-offs between the search approaches. These insights can help researchers select an appropriate method and identify opportunities to incorporate effective mechanisms from one approach into the other.
Time-Series Foundation Models (TSFMs) are increasingly used in sensor applications, where sampling frequency (SF) varies across sensors, devices, and operating modes (e.g., power saving). Ideally, TSFMs must remain effective during such SF changes, provided the SF retains task-critical bandwidth. In this paper, we stress-test 4 prominent TSFMs: MOMENT, Chronos-2, Moirai-MoE, and TimesFM, to quantify their SF-invariance and investigate whether their behavior tracks the resulting loss of information bandwidth rather than the SF value itself. Our stress-test probes 3 dimensions: (i) the shift in embeddings, (ii) the degradation in forecasting performance, and (iii) the degradation in downstream performance, quantified by three novel metrics: FSI, FDI, and CDI, respectively. To our knowledge, this is the first study to place SF-invariance at the core of TSFM evaluations. Our results characterize: (i) how SF-invariance depends on the margin between signal's native-SF and the bandwidth its task requires, (ii) how much of that invariance the choice of TSFM explains, and (iii) how much of TSFM's performance degradation under a lower SF reflects information loss rather than a limitation of their abilities. These findings reframe SF-invariance as a measurable, signal-aware property of the (signal X TSFM) pair, and guide practitioners on when TSFMs display SF-invariance.
Neural Processes (NPs) are flexible meta-learning models capable of uncertainty quantification across diverse applications. However, the standard NP framework assumes access to noise-free inputs. This is unrealistic in many practical settings, where sensor noise or measurement errors are unavoidable. This discrepancy between model assumptions and real-world data can significantly degrade prediction quality and uncertainty estimates. We introduce Denoising Neural Processes (DnNPs), a principled extension of the NP framework that explicitly models input uncertainty. We train DnNPs via variational inference. DnNPs jointly learn to denoise corrupted inputs and perform meta-learning, making them robust to noisy training data. We demonstrate DnNPs' effectiveness on both synthetic and real-world datasets, showing consistent improvements in prediction and uncertainty quantification compared to standard NPs. On PM2.5 spatial interpolation from real-world air quality sensors, DnNP reduces RMSE by up to 30% and NLL by up to 73% relative to standard NPs. We also provide theoretical grounding: we show that the Bayes-optimal predictor under noisy inputs belongs to the DnNP model class, and that plug-in models with fixed observation variance incur an irreducible KL divergence gap that DnNP provably avoids.
Streaming speech recognizers output a word only after the speaker has finished it, typically 90-400 ms later. For a voice agent every millisecond counts: the sooner the words are outputted, the sooner the agent can start a search or plan its reply. We present Earlybird, a 29M-parameter streaming recognizer (16.8 MB at 4 bits). Earlybird learns from audio cut off at random points, including partway through a word, and is asked to predict every word that is due by then in full, plus a guess of the next two words. Its raw token probabilities give a confidence for every word including the one still being spoken. On clean audiobook speech (LibriSpeech test-clean), the median word is predicted correctly 90 ms before the speaker finishes it, while released streaming models get it 90 to 400 ms after it ends. Earlybird is the earliest on all six test sets we time, covering audiobooks, TED talks, public speech, earnings calls, and synthetic speech. On the same audiobooks, 69% of words are predicted correctly before they end, against 3-25% for the other models. For words still being spoken, confidence is calibrated, with an expected calibration error (ECE) of 3.7 to 5.5. On the eight English Open ASR Leaderboard sets, it is more accurate than Moonshine Tiny on every set and roughly ties Whistle, a model of the same size.
Video world models promise general-purpose interactive simulators, but their training is limited by action supervision: action labels are scarce, fragmented across incompatible specifications, and do not scale with the unlabeled video pretraining corpus. A common workaround jointly trains an action encoder to compress consecutive frames into a latent action and a decoder to reconstruct the next frame from the previous frame and that latent, forcing the latent to capture only new information. This has two limitations: learning the action space from scratch can limit generalization to unseen domains, and the bottleneck trades expressiveness for separation, with tighter settings losing nuance and looser ones risking scene leakage under domain shifts. We argue for a different starting point: rather than learning actions from scratch, we start from a pretrained foundation model (V-JEPA 2), whose embeddings of a frame pair already carry substantial action information after large-scale, diverse video pretraining. The remaining challenge is to extract the action while discarding the scene. We propose a label-free contrastive adaptation framework that pulls together frame pairs sharing the same motion under a different scene and pushes apart pairs from the same scene with different motion, reshaping the geometry of the pretrained space without imposing a reconstruction bottleneck. We introduce a unified evaluation protocol along three axes (action expressiveness, scene invariance, and action-conditioned video generation quality) across robot manipulation, camera navigation, game, and driving benchmarks, most of which are unseen during adaptation. Compared with annotation-free latent-action methods, our representations are substantially more expressive while matching their scene invariance, and they yield world models that transfer demonstrated actions to new scenes more faithfully.
Formal world models can support planning, verification, and safety reasoning, but constructing them from open-ended natural-language situations is difficult, particularly in multi-agent settings where several formalizations may be plausible. We study whether five LLMs can generate such world models as partially observable stochastic games and whether mechanical validity is sufficient to establish that the resulting model faithfully represents the intended decision problem. We introduce a benchmark of 44 scenarios spanning game, situational, and cultural settings and evaluate generated world models along three dimensions: mechanical validity, reference-relative structural agreement, and semantic criteria. Across three empirical tracks, we analyze 410 first-pass generations. In the reference-backed game track, 134/135 generated world models pass a strict mechanical audit, yet mean permutation-invariant structural F1 is only 0.366, with just 28/134 mechanically valid models exactly matching the authored reference. In the cultural track, explicitly naming the country changes the generated structural representation for most LLMs while producing only small changes in automated semantic scores. These results show that executability is a necessary but insufficient criterion for LLM-generated world models: models that are mechanically valid can still differ substantially in the structure used by downstream neuro-symbolic reasoning and planning.
Local decoding usually fixes one weight precision for the whole generation. We study a different allocation: store 8-bit weights, decode with their top 4 bits, and redo a step at 8 bits only when the low-precision output distribution looks uncertain (entropy above a threshold or a small top-two margin). Quantization is simulated, the model is Qwen3-0.6B, and the benchmark is GSM8K. Cost is counted in weight bytes read per token. On 36724 replayed steps the entropy of the low pass ranks argmax flips with AUC 0.93. At a 15% escalation budget it catches 78% of flips and the margin gate 94%, against the 15% expected for a random gate of equal rate (a computed expectation, not a run). In 150 live problems all-high scores 64.0% and all-low 55.3% (McNemar p=0.053). An entropy gate with a 16% escalation rate scores 62.7% at 0.87 of the all-high weight bytes. Its paired difference is -1.3 (-8.0 to +5.3) points to all-high and +7.3 (+0.0 to +14.7) points to all-low (p=0.080), so the live data do not separate the gate from either end. Verified speculation reaches zero KL at similar bytes in the same cost model. A live random control, a thinking-mode phase rule, a second model and live KL were not run.
Audio foundation models like CLAP, MuQ, and MERT excel at capturing general musical similarity. On the other hand, their entangled latent spaces limit interpretability and fine-grained attribute control. Conditional Similarity Networks (CSN) address this by pairing a backbone representation model with learned masks that gate the embedding space into attribute-specific subspaces. However, existing methods train the backbone from scratch, rather than leveraging pretrained foundation models. We instead present a systematic study of post-hoc disentanglement, adapting CSN to gate frozen pretrained embeddings. We evaluate adding trainable parameters to adapt the embedding under different masking and pooling regimes across two datasets. Our findings show that minimal adaptation capacity is sufficient for post-hoc gating to improve both attribute-specific disentanglement and overall retrieval performance, while pooling, masking, and depth choices yield limited additional gains.