We show exact signal recovery in generative compressed sensing for feedforward neural networks with piecewise-linear activations and a final sigmoid activation, with no assumption of expansivity. Consider a ReLU network of depth $d$ with a sigmoid final activation, whose input layer, of dimension $k$, is the narrowest, and let $K$ be the geometric mean of the widths of its layers. We show that, with high probability, all signals of the range whose entries lie in $[\delta, 1 - \delta]$ are uniformly recovered from $m \gtrsim k(d \log(K/k) + \log(1+\log(1/\delta)))$ measurements.
Training on large datasets motivates selecting which examples receive expensive training computation. Fixed quotas determine the amount of pruning in advance, while absolute confidence cutoffs depend on the score scale. We study a population-relative alternative that adapts the threshold and retained count together. We introduce LACS (Label-Agnostic Confidence Selector), a training-time rule based on maximum predictive confidence, requiring neither labels at selection time nor per-example gradients. LACS removes examples at or above a log-mean-exp reference plus a margin. We characterize its sensitivity to score gaps and mass, and derive an exact no-pruning condition. For $0<\rho<1$, the removed fraction is at most $\rho$ and at least one example remains. The unsigned score permits both confidently correct and confidently wrong examples to be removed. Seven evaluation settings include neural training, language modeling, and boosting. Neural execution studies reduce backward-active rows by 10-35%; on CIFAR-10, 29% fewer backward rows accompany accuracy means of 89.77% versus 89.82% for full CE. At a matched backward budget and number of optimizer updates, LACS is 1.0 pp above small-batch training and 1.2 pp above random selection over five seeds. Across 14 tabular datasets, quadratic boosting reduces mean test error from 23.36% to 22.96%. Its full training replays use about 30% fewer fitting rows; these counters span a different horizon from validation-selected accuracy. In one RBF-SVM case study, elapsed time falls by 37% with slightly lower error. These findings separate selective-work reduction from realized acceleration: the current exact-current-score neural implementation uses an additional scoring forward, so its backward sparsity does not yet translate into wall-clock acceleration. This overhead is an execution choice rather than a requirement of the selection rule.
Multi-teacher on-policy distillation (MOPD) extends on-policy distillation (OPD) to consolidate capabilities from multiple teachers. Although prior work reports near-perfect in-distribution (IND) distillation performance, we find that this can mask incomplete capability consolidation beyond the supervised distribution. A preliminary study on synthetic reasoning shows that OPD does not fully transfer teachers' generalization capabilities. We then extend this investigation to real-world tasks, evaluating extrapolative and compositional generalization across mathematical reasoning, coding, agent, and long-context capabilities using Qwen3 models. Across extrapolative settings, OPD and MOPD reach 96.04-102.73% of their teachers' IND scores. Yet their out-of-distribution (OOD) scores remain 2.45-16.25% below the base model's. Across three compositional settings, MOPD improves IND performance over the base model, but its performance on composed tasks as OOD remains 18.19-24.34% below that of the best teachers. To understand these gaps, we examine teacher capability and data coverage as two potential bottlenecks. Jointly alleviating these bottlenecks eliminates OOD forgetting in our experiments and suggests the potential for generalization beyond the teacher. (M)OPD also reduces OOD forgetting on average relative to its reinforcement learning (RL) baselines. Analysis of the associated update geometry reveals smaller and more spectrally concentrated updates on average than those of the RL baselines, along with greater shared update structure under MOPD. Motivated by these findings, we propose geometry-based patching and base-preserving replay to improve OOD performance while largely preserving IND performance. Our findings identify a blind spot in capability consolidation via (M)OPD, highlighting the need for algorithms that explicitly support OOD retention and generalization.
Variational approaches to learning transport maps for multimodal sampling can suffer from mode collapse. One way to mitigate this problem is to learn transport along a prescribed path of intermediate distributions defined by geometric annealing. We show that annealing can take a substantial detour in Wasserstein space: when mass moves between modes, the path can be much longer than the distance between its endpoints. We further link path geometry to transport learning, showing that long annealing paths can slow population optimization, while sufficiently regular paths permit rapid convergence. Motivated by these findings, we propose componentwise annealed transport (CAT) to reduce the need for mass transfer between modes. Starting from a mode-informed mixture reference, CAT constructs an exact mixture decomposition of the target and learns separate transport maps along componentwise annealing paths. We estimate the component weights separately and show that these estimates remain consistent even when the learned flows are inexact. We also develop adaptive annealing schedules guided by transport effort. Experiments on multimodal synthetic distributions and Bayesian inference problems demonstrate improvements in transport learning and sampling efficiency.
A deep RL agent trained on a fixed persona reward plays one way: an Explorer keeps exploring while starving, a Survivor hoards on an empty map. Since each persona fixes its own reward, automated playtesting trains N agents for N styles. Changing reward composition mid-episode runs into the switching-stability dilemma: PPO degrades under non-stationary rewards. We present RIDGE, a single-agent architecture that sidesteps the dilemma with two design choices: a state-conditioned blender that computes softmax-normalized sigmoid weights from slow-moving state variables, and a multi-head critic that trains one value head per persona on its own stationary signal. On Crafter with four personas, one \ridge agent acquires both branches of the wood-tier crafting/combat split that no fixed-persona baseline solves jointly, and a sharpness ablation confirms state-conditioning drives the gain. We position these results as demonstrations of the blending mechanism's dynamics across n=3 seeds per condition; larger multi-seed sweeps remain future work.
Grounding a text query in an image needs a matrix scoring every term against every patch. Vision-language models compute it with an aligned text encoder, which many image encoders, such as histopathology-related models trained on slides alone, lack. The Caption-Anchored Patch-Term Occurrence Representation (CAPTOR) estimates the matrix without a text encoder. It clusters a frozen image encoder's patch tokens into visual words and describes each captioned image by its share of patches in each word. A term signature is the standardized difference of those shares between images whose captions mention the term and the rest. The matrix factorizes into a term-by-word table and a query-independent image factor, so an image is encoded once and a new term adds one table row. Against every training-free baseline, CAPTOR leads on Common Objects in Context (COCO), Pascal Visual Object Classes (VOC) and Cityscapes. On COCO, scored on the cells one object owns, it reaches 0.565 mean intersection over union (mIoU) against 0.502. On all three, it is also level with a text alignment trained on the same backbone, which needs gradients, a text encoder and several times the captions. On held-out breast-cancer slides, it exceeds the strongest baseline scoring patches with a pathology text encoder by 0.160 mIoU. Read through CAPTOR, UNI, an encoder trained on slides alone, exceeds such baselines by 0.132 on a common grid. Counting can therefore give a text interface to any image encoder trained without language, wherever captioned images exist.
Different retrievers exist for retrieval-augmented generation, yet none performs best for all queries. Therefore, the key question is: given a query and retrieved passages from different retrievers, which one is most suited for that query? Existing routers usually make predictions from query-specific features and a fixed pool of retrievers; such routers have no output for an unseen retriever and are incapable by design on a new corpus. To fill these gaps, RAGCompass reads the retrieved passage sets themselves: passages from an unfamiliar retriever or corpus stay directly comparable to other passages, so routing stays well designed where a query-only map would fail. We contrastively fine-tune a reranker against the answer each evidence set produces, which accounts for most of the gain; a small comparison head adds a moderate improvement, and neither component has retriever-specific parameters. RAGCompass improves answer F1 by 28% relative to the best fixed retriever, and outperforms fusion baselines, three query-only routers, and a prompted 8B language model. Generalization experiments show that RAGCompass keeps most of its gain under an unseen corpus, retriever, or reader, while query-only baselines fall to the best fixed retriever or below.
Relevance propagation, which redistributes a network's prediction backward through the layers, is usually treated as a heuristic. We show that it is the solution of a single variational problem: Layer-wise Relevance Propagation (LRP) is the divergence-free flow on the active computational graph of zero relative entropy to the backward transition law of the forward activation flow; for a non-negative reference, this is a relative entropy between Markov measures on backward paths. The rules LRP$_0$, LRP$_{z^+}$ and LRP$_{\alpha,\beta}$ follow from this principle by the choice of a reference, and are instances of one rule, LRPN$_{\alpha,\beta}$, whose weights may vary from node to node, for instance with the node's activity, in the spirit of Hebb's principle. The principle rests on a measure-theoretic formalization of feed-forward networks as directed acyclic graphs of Filippov operators, which covers piecewise-smooth and discontinuous architectures (e.g. hard-routed Mixture-of-Experts) with a Clarke-Filippov generalized-gradient calculus; it also characterizes backpropagation as the adjoint equation of a Lagrangian relaxation. On networks of ReLU and max-pooling nodes, LRP$_0$ is Gradient$\times$Input, with an explicit jump across a kink where LRP$_{z^+}$ stays continuous. Taking the entropy edge by edge yields a new rule, Sinkhorn Relevance Propagation. Aggregated over an input region, the flow of a non-negative reference tells, with exact accounting, which hidden units carry its evidence, the conditioned flow being a Schr\"odinger bridge. On VGG-16, the conservation identities hold to round-off, and the rules are told apart by the sign of their maps, the share of the output left to the bias nodes, faithfulness, localization and class dependence; on ConvNeXt-Tiny, LRPN$_{\alpha,\beta}$ conserves the output and its Hebbian weights give the highest ABPC of the rules tested.
Can a neural network learn to render 3D Gaussian splats on its own, with no rasterizer in the renderer? We present GaussianRenderFormer, a transformer that takes a set of Gaussians and a camera and outputs the image directly. Each Gaussian becomes one token in a view-independent encoder, and a view-dependent decoder turns small patches of camera rays into pixels by attending to those tokens. Trained on 26,820 Objaverse objects pruned to 20,000 Gaussians each, GaussianRenderFormer renders 300 unseen objects at 38.08 dB PSNR against the full, unpruned splats, while rasterizing the same pruned input reaches 44.02 dB. The gap is not one of generalization, and our diagnostics are consistent with a bottleneck in the view decoder: telling it the projected footprint of each Gaussian, which it could in principle infer, adds 1.4 dB. Under-training is not ruled out. Code, trained models and videos: https://svlwoof.github.io/gaussianrenderformer
Reasoning-oriented language model training increasingly relies on multiple on-policy rollouts, but standard Group Relative Policy Optimization (GRPO) uses a fixed group size, wasting compute on easy prompts and under-exploring hard ones. Existing adaptive-budget methods alleviate this inefficiency, yet their stopping rules are heuristic and do not provide a finite-sample calibration guarantee. We propose CGRPO, which replaces the fixed rollout budget of standard GPRO with calibrated conformal thresholds over task-specific nonconformity scores, combining Adaptive Prediction Set (APS)-style calibration for closed-form reasoning, a first-success score for execution-based code, automatic miscoverage selection, and periodic re-calibration. We prove finite-sample coverage for the resulting stopping rule and show that, under monotone policy improvement, recalibration induces a non-increasing expected training budget. Across six benchmarks spanning mathematics, graduate-level science, and code generation, CGRPO reduces mean training rollouts by 49-70% relative to fixed-budget GRPO while retaining strong fixed-budget accuracy, achieves the strongest pass@1 among evaluated static and dynamic baselines on different datasets, and transfers across model families with up to 82% rollout savings.
Data tends to be most useful when it is shared, and the data most worth sharing is often the data that privacy obligations restrict. Synthetic data from a generator trained under differential privacy (DP) is one way to reconcile the two, but its utility depends on the privacy budget $\varepsilon$ and on how much data the generator sees - a parameter that most data sharing applications of synthetic data control. We present SAVANT, a self-attentive variational autoencoder for tabular data trained with DP-SGD that, to our knowledge, is the first to combine attention, a VAE, and DP in one tabular model. We evaluate it against two marginal mechanisms, PrivBayes and MST, jointly over the privacy budget and the dataset size on two public datasets, measuring the utility a downstream classifier loses by training on synthetic rather than real data. We find that, under the same DP guarantee, more data generally means more utility. For the marginal mechanisms it absorbs most of what a strict budget costs, and a noise-scaling argument suggests DP-SGD exchanges data fo budget at a worse rate. Two independent membership-inference attacks, judged against a measured null, found no aggregate membership signal.
We study repeated general games over $T$ rounds between a utility-maximizing optimizer and a learning agent, where budget constraints couple the players' decisions across rounds. Without budgets, low swap regret simultaneously guarantees the learner sublinear regret and caps the optimizer at its Stackelberg value. Budgets break this picture: adversarial no-regret becomes impossible, so the natural goals are non-manipulability---no adaptive optimizer earns more than $\mathtt{BSE} \cdot T + o(T)$, where $\mathtt{BSE}$ is its Budgeted Stackelberg Equilibrium value---and low stochastic regret when the optimizer plays a fixed mixed strategy. We give an efficient, budget-feasible learning algorithm that achieves both goals in every budgeted game: any adaptive optimizer, budgeted or not, earns at most $\mathtt{BSE} \cdot T + O(\sqrt{T})$ in expectation, and if the optimizer plays a fixed mixed strategy on each of a constant number of intervals, the learner has $O(\sqrt{T})$ regret on each such interval. Our learner best responds to forecasts of the optimizer's play that are calibrated in two novel senses, since under budgets the calibration that prevents manipulation no longer yields low stochastic regret. Response calibration, which conditions on the learner's own mixed response by embedding its continuous mixing weight as a step function in $L^2([-1,1])$, prevents manipulation without discretizing responses; dual-weighted Lagrangian calibration ensures low regret. Although low calibration error requires randomized forecasts---distributions over a continuous space that need not have a succinct representation or be directly computable---we construct such forecasts in polynomial time by simulating online gradient ascent.
Multimodal AI fusion of histology and clinical data is an active research area in computational pathology, but fusion methods are rarely compared under identical conditions and across different clinical endpoints. We present a modular framework in which slide-level embeddings, obtained by training an attention-based aggregator over patch embeddings from a frozen foundation model, and per-visit clinical embeddings are fed into 15 interchangeable fusion methods with a shared output dimension and prediction head (except decision-level methods). Using a prospective cohort of newly diagnosed inflammatory bowel disease patients, we predicted STRIDE-II-aligned remission outcomes from baseline or follow-up visits. The best fusion method reached macro-F1 of 0.57-0.75, outperforming both single-modality models in four of six scenarios, and differed by outcome and input. Image reliance, the drop in balanced accuracy when the slide embedding is removed, tended to follow the remission type - highest for histological and lowest for endoscopic remission; methods with similar accuracy relied on the slide to very different degrees.
Self-supervised world models are often said to organize their latents around physical quantities. We test this on 2D decaying Navier-Stokes flow in three pre-registered studies, each tested once on held-out data. In Study 1, a joint-embedding predictive architecture (JEPA) trained to predict the flow fails the pre-registered criteria for organizing its latent around fluid quantities better than an untrained network of the same architecture (not supported); a JEPA trained on non-physical linear dynamics and a static invariance model that never sees time do as well or better. In Study 2, plain autoencoders that render the field make enstrophy prominent in their slowest unsupervised latent directions (best r 0.80-0.86 vs 0.38 untrained; fresh seeds, sealed data). Adding VICReg lowers this prominence in every comparison, though not by the pre-registered margin for vorticity, while raising the autoencoders' linear decodability (partially supported). In Study 3, we build steerable directions ("dials") by routing: the decoder gets overall scale only through a small conditioning code, while its other pathway carries a degraded copy of spectral shape. Label-free rules find a loudness dial in all 8 trained models and a shape dial in 3 of 5 fresh seeds on sealed data, against a pre-registered bar of 4; neither control shows a shape dial (partially supported). Across the studies, what a latent contains and what can be controlled through it come apart. In exploration, a supervised direction reading spectral shape almost perfectly moved nothing when steered. During the build, the loudness dial had to be found by its effect on the output (closest to a pure rescaling), not by data variance; for shape, a variance rule and an effect-constrained rule each worked in 3 of 5 seeds. The program used about 70 logged hours on one consumer GPU; all pre-registrations, deviations, code and outputs are public.
Remote sensing change detection is typically formulated as Siamese feature comparison, where two images are encoded independently and a decoder determines which discrepancies correspond to meaningful change. This is challenging because the encoders preserve both true scene changes and nuisance variations such as illumination, seasonal, radiometric, and geometric differences. We introduce ChangeJEPA, a bitemporal joint-embedding predictive architecture that reformulates change detection as latent post-state prediction. Given a pre-change image and partial post-change evidence, ChangeJEPA predicts the full post-change representation under two complementary conditions: a prior hypothesis derived from the pre-change scene and an updated hypothesis conditioned on observed post-change context. A joint-embedding predictive loss aligns both hypotheses with the target latent representation, encouraging semantic consistency while suppressing transient appearance variations. Change is then inferred from the disagreement among the pre-change representation, the observed post-change target, and the predicted latent states, capturing complementary evidence of change, prior surprise, contextual update, and prediction reliability. Across SYSU-CD, LEVIR-CD+, and CDD, ChangeJEPA consistently outperforms its decoder matched Siamese baseline by up to 1.51 IoU points and achieves the best IoU and F1-score on all three benchmarks, demonstrating the advantage of latent post-state prediction over direct feature comparison. Page: https://buddhi19.github.io/ChangeJEPA
Point cloud segmentation is a fundamental task in 3D data analysis, with applications across robotics, augmented and virtual reality, and remote sensing. However, the inherent sparsity, noise, and unordered nature of point clouds introduce significant challenges, particularly in achieving reliable classification. While the state-of-the-art rely on training models for explicit modeling of uncertainty, there is a gap in exploring different roles that estimated uncertainty using generative approaches can play. This work investigates the impact of deterministic uncertainty in point cloud segmentation by explicitly modeling epistemic uncertainty and using it in different roles, namely, as a feature and in loss function. We evaluate a baseline artificial neural network (ANN) and two uncertainty-aware variants (uncertainty as an additional input feature and as a per-sample confidence weight in the training loss) on four large-scale LiDAR benchmarks spanning airborne and mobile acquisition, namely ISPRS Vaihingen, AHN3, Paris-Lille-3D, and Toronto-3D, together with PointNet++ and RandLA-Net references and their uncertainty-augmented counterparts under canonical spatially disjoint train and test splits. Under this protocol, uncertainty-aware modeling improves mean Intersection over Union, recovers minority classes that the baseline ANN never predicts at all, and in half of the configurations improves the alignment between prediction error and estimated uncertainty as measured by Patch Accuracy versus Patch Uncertainty. Our findings indicate that uncertainty can play multiple roles in the segmentation pipeline rather than as a device for improving accuracy alone.
Contrastive vision-language models score left/right relations near chance, but the symptom has two causes that call for opposite repairs: the relation may survive in the pooled embedding yet be unreachable by native image-text similarity, or pooling may discard it while the dense features retain it. On clean two object photographs, across ten encoders, the lateral polarity is linearly decodable from the flip odd component of the pooled vector at 0.77-0.99, yet a search over a million captions leaves nine of the ten at chance under native scoring: an access failure. On controlled C4 orbits, where the relation is the only thing that varies, pooling keeps the arrangement axis at 0.998-1.000 and loses its polarity, which the patch grid retains: a pooling failure. Both sit downstream of the dense features, so one readout covers both. Our zero parameter, query conditioned spatial moment recomputes the geometry from the patch grid and stores it as a vector beside the native embedding. On a 118,287 image index it raises left/right pairwise accuracy against reversed geometry negatives from 0.51-0.53 to 0.81-0.96 and conditional R@1 among 21 category matched candidates from 0.05 to 0.37-0.75, at under a millisecond per batched query, while reproducing the native ranking exactly for queries that name no relation.
Metabolic tumor volume (MTV) summarizes whole-body PET disease burden for prognostic stratification and treatment assessment. Automated MTV is useful only if its estimate is reliable for the individual patient, yet segmentation models are usually evaluated with overlap metrics such as Dice rather than burden error. We examined whether overlap metrics identify unreliable MTV estimates using 1,014 autoPET FDG PET/CT studies. We evaluated LesionTracer predictions held out from fine-tuning, compared five selected released models on 489 tumor-positive patients, and examined the findings in 597 PSMA studies. Across the five FDG models, ranking by higher mean Dice gave the same order as ranking by lower mean relative MTV error ($\rho=-1.00$, exact $p=0.017$). At the patient level, however, out-of-fold Dice tracked relative MTV error strongly ($\rho=-0.75$), yet 82 of 489 patients (16.8\%) fell on opposite sides of the Dice 0.80 and MTV-error 25\% thresholds. Within the training-exposed released models, correlations ranged from $-0.47$ to $-0.74$ over narrower, high-Dice ranges. Out of fold, opposing false-positive and false-negative volumes cancelled about one third of spatial error in net MTV. In 236 of 489 patients, each of the five models missed at least one lesion, though not necessarily the same lesion. Thus, agreement in average model rankings did not establish the reliability of individual MTV estimates or reveal shared lesion misses. PET segmentation benchmarks should report patient-level MTV error and missed-lesion counts alongside Dice and other overlap metrics.
In multi-objective reinforcement learning, preferences may be specified only after data collection, requiring one dataset to support a continuum of reward trade-offs. We study finite discounted Markov decision processes with unknown transitions, $S$ states, $A$ actions, and $m$ known reward functions taking values in $[0,1]$, given access to a generative model. The learner must provide an $\varepsilon$-optimal policy for every convex combination of these rewards without further sampling, with a guarantee that holds simultaneously over all preferences and initial states. Writing $H=(1-\gamma)^{-1}$, we establish the constant-confidence minimax sample complexity $\widetilde\Theta\!\left( \frac{SAH^3}{\varepsilon^2}\min\{S,\log(em)\} \right)$, where the suppressed factors are logarithmic in $S,A,H,\varepsilon^{-1}$ and the dependence on $m$ is explicit. Thus, up to these logarithmic factors, the worst-case statistical cost of supporting all preferences grows logarithmically with the number of objectives and saturates at the number of states. The upper bound uses uniform state--action sampling and empirical planning, with reward rounding in the logarithmic-objective regime. Its analysis chains optimal-value increments over covers of the reward hull. A one-sided Bellman variance bound controls signed value differences under a second transition kernel, preserving the $H^3\varepsilon^{-2}$ dependence. A complementary uniform guarantee for all bounded rewards yields the state-dimension saturation. We establish lower bounds matching up to logarithmic factors through a rare-exit construction with subset rewards and action encoding. These bounds apply to adaptive simulator queries and arbitrary randomized history-dependent output policies, while the upper bounds are attained by deterministic stationary policies.
In biodiversity monitoring, a single image raises several questions: which taxon, how many, which trait, and whether the individual has been seen before. Vision-language models (VLMs) could answer all of these from the image, but their ability is largely untested in biology, as existing benchmarks focus mostly on species identification within a few taxonomic groups. We present Bio-VQA, the largest biological visual question answering (VQA) benchmark and the only one to span four aspects of biological visual understanding: hierarchical taxonomic reasoning, fine-grained attribute grounding, quantitative spatial reasoning, and individual re-identification. Bio-VQA contains 1.86M questions over 217K images from 11 public datasets, and extends right-or-wrong scoring to graded scoring, so a near-miss counts for more than a distant one. Evaluating 18 open-weight models from 3.75B to 38.4B parameters, we find that general visual capabilities do not transfer to biological organisms: the best model counts animals exactly on 41.7% of images with 2 to 10 individuals, less than half its accuracy on everyday objects at the same counts. Taxonomic accuracy falls to one sixth, from 0.72 at phylum to 0.12 at scientific name, and scaling helps mostly at coarse ranks. The gap to a domain specialist widens accordingly, from 0.731 against BioCLIP-2's 0.874 at class (1.2x) to 0.269 against 0.544 at scientific name (2x). On trait size measurement, 15 of the 18 models do no better than an image-blind baseline that predicts one fixed number for every specimen, and on re-identification, 17 of the 18 models fall below chance under adversarial conditions. These failures persist across a tenfold range of model size. We hope Bio-VQA becomes a standard test of how well VLM capabilities transfer to the biological domain.