Likelihood-based methods are widely used to learn the parameters of partially observable Markov decision processes (POMDPs). While finite-sample guarantees for policy suboptimality exist for certain POMDP subclasses, analogous guarantees for parameter recovery remain less well understood. For any candidate POMDP parameter, we bound its parameter recovery error in terms of its likelihood gap to the ground truth, the structural conditioning of the observation and transition kernels, and the behavior policy used to collect data. The bound applies in both the undercomplete and overcomplete regimes, with the latter following naturally from the undercomplete analysis. Because the result applies to any sufficiently good candidate, it yields finite-sample guarantees for the maximum-likelihood estimator and any procedure producing a parameter with controlled likelihood gap. This is particularly useful for near-optimal learning in infinite-horizon POMDPs, where existing approaches largely rely on model-based planning with accurate transition and observation kernel estimates. The bound separates the contributions of the behavior policy and model, showing how data-collection choices, including action coverage, and their effects on mixing can accelerate parameter learning.
As vision-language models (VLMs) move toward interpreting medical images directly, their outputs increasingly include not just a label but a stated confidence and a written justification. Yet the field still predominantly evaluates VLMs using just accuracy. We evaluate the coherence between answer, confidence, and explanation across four VLMs on breast density and BI-RADS categorization, using two mammography datasets. We perform controlled perturbations of the question, multiple choice option orders, and the images and assess changes in the three outputs. Our results suggest confidence does not decrease even as the mammography image is changed. Shuffling the options also changes most VLMs' answers. Notably, there are cases where a model's own stated reasoning contradicts its final answer, a distinction accuracy alone cannot capture but one that governs whether an AI-generated answer can be trusted.
Many tasks compose transformations by a known rule: chained rotations in robotics and navigation. Learned models relearn it, compounding error; classical estimators hardcode it and miss hardware deviations. We impose the known structure and learn only the remainder: an exact non-commutative group law sits in the node of a self-contained layer, its edges learn local increments, and a zero-initialized correction psi learns unknown deviations of the law. A pre-training criterion predicts exact representability. On fixed-angle SU(2) composition, a 54-parameter dense edge is the most accurate dense model, at 27x fewer parameters than the strongest fixed-matrix baseline, and lies 0.05 deg above the Bayes floor. On real robot arms, an isolated 306-parameter psi reaches 1.12-1.14 mm versus 1.29-1.33 mm for Gauss-Newton at N >= 60, and the recipe transfers across three vendors and two protocols. The recipe matches classical fits where the data obey the law and loses where the imposed law is not the task's.
Vision language models (VLMs) encode substantial expert knowledge within their weights, yet their direct application to specialised classification domains such as benthic imagery remains limited. Meanwhile, vision transformers trained on these tasks rely predominantly on low-level visual latents (colours, textures, and shapes) that lack the higher-level reasoning humans employ for identification. We propose Guided Hierarchical Multi-Label (GHiML) learning, a pipeline that extracts conceptual latent variables from a VLM conditioned on training data, generates sample-relevant captions grounded in those variables, and fuses the resulting textual embeddings with image embeddings via cross-attention and a per-node gated router. We also introduce the inverted HML score ($d_{\mathrm{inv}}$), a similarity score that evaluates hierarchical multi-label predictions through graph traversal distance and a cardinality penalty, providing an interpretable alternative to existing measures, which are insensitive to realistic degradations in sparse, deep hierarchies. Experiments on BenthicNet, FathomNet and iNaturalist show that the router improves macro $F_1$ over the vision-only baseline by up to 6.7 pp and $d_{\mathrm{inv}}$ by up to 21.1 pp on in-domain benthic data, with gains persisting under distribution shift. Ablations locate the contribution of the text itself: on the multi-label biotic hierarchies, removing the per-sample text signal degrades $d_{\mathrm{inv}}$ while leaving macro $F_1$ unchanged, consistent with text acting through prediction cardinality rather than node discrimination, whereas iNaturalist, whose annotations are single-label, loses 18.7% of macro $F_1$. Our results support the hypothesis that VLMs contain exploitable specialised knowledge that, when appropriately elicited, can augment classification in domains where direct visual learning is insufficient.
Video world models conditioned on an input video make it possible to explore real, recorded dynamic scenes, enabling re-shooting of footage, free-viewpoint replay, and real-to-sim environments for embodied agents. Such uses require the generated rollout to stay faithful to the recorded world, yet existing benchmarks start from a single image or a text prompt, and video-conditioned generation is evaluated mainly on perceptual quality and camera following. These evaluations do not directly assess whether exploration preserves subject identity, event progression, and scene consistency upon revisiting a location. We introduce WanderBench, a benchmark for exploring dynamic worlds from video. It comprises 745 single-shot source videos of real and AI-generated content in three difficulty tiers (3-25 s), paired with camera routes given as keyboard actions, and evaluates six dimensions with 24 human-validated sub-metrics. Benchmarking eight models, we find that as clips grow longer, models keep the static scene and visual quality but lose camera control and the source dynamics, and that none keeps revisited places consistent. We further propose WanderLand, a training-free framework that conditions each frame of a frozen model on the same-instant source and, on revisits, on its own earlier static content. WanderLand improves five models on every dimension, by up to 8.3 points in source-dynamics fidelity.
We propose SciMARS, a framework for recursive task-local adaptation of multi-agent harnesses in closed-loop scientific discovery. Multi-agent systems coordinate analysis, experimental design, and verification as scientific evidence evolves. However, two problems limit their effectiveness: (i) without shared memory, agents lose observations and repeat work; (ii) fixed role procedures, handoffs, and memory policies perpetuate mistakes despite retained evidence. We address these problems with a shared evidence graph and separate component editors. SciMARS adapts roles and coordination; SciMARS + Memory adds a third editor for how agents record, retrieve, and receive evidence. Candidate probes guide composition and selection of the next harness. We evaluate scientific discovery through experimental-data replay and a digital twin. SciMARS + Memory resolves 12/18 GPT-6-Astra and 2/18 GLM-5.3 attempts on Autonomous Science Bench, compared with 8/18 and 1/18 for the base multi-agent team. On ScienceAgentBench and AppWorld, it solves 14/38 and 11/20 tasks, compared with 11/38 and 5/20 for the base team. These results support combining shared memory with harness adaptation across scientific and interactive tasks.
We prove $F_{\mathrm{weighted}}(2,2,3)<11/6$ for minimum-Frobenius-norm least-squares regression with two input and two output coordinates, improving the upper bound $15/8$. More precisely, a universal $\varepsilon_0>0$ satisfies $\mathcal T_3\le(5/6-\varepsilon_0)E$ for every finite residual frame. The proof combines an exact scalar certificate, rigidity of equality in planar selection, and a compactness argument that controls disappearing weights. Together with the known lower bound, this gives $13/8\le F_{\mathrm{weighted}}(2,2,3)<11/6$. No numerical value of $\varepsilon_0$ is claimed.
Alternating minimization is a standard tool for optimization problems with two coupled variable blocks. Its cost is dominated by the inner subproblem solvers, which rarely have a closed-form solution. In this paper, we propose to learn an inexact alternating scheme in which neural networks approximate the subproblem solution at each iteration. For convex problems with quadratic couplings between the variable blocks that satisfy a quadratic growth condition, we derive a worst-case bound on the optimality gap that exposes the per-iteration error that each network is trained to minimize in a greedy fashion. Combining the bound with a PAC-Bayes argument, we obtain a fixed-horizon, high-probability bound on the expected optimality gap over the distribution of problem instances. We apply the learned scheme to robust matrix completion and low-dose CT reconstruction with dictionary regularization where we achieve a 5x-13x speedup over the inexact first-order baseline at matched accuracy. A brief end-to-end fine-tuning stage extends this up to 30x and outperforms other learned schemes. The code is available at https://github.com/paulhausner/learned-alternating-minimization.
Multilingual generation requires large language models (LLMs) to determine both what to say and in what language to say it, yet these two commitments are difficult to disentangle from model outputs alone. We use translation as a controlled probe: for multi-parallel segments that a model translates well in one language direction but poorly in another, content is held fixed while language varies. Using activation patching, we write residual-stream states from successful runs into failing ones and compare content-matched states with content-independent controls, allowing us to trace content-specific and language-related signals across layers. Across three LLMs of different scales and architectures, we find a consistent three-stage pattern of multilingual generation. Intermediate layers first express a default English signal; content information becomes causally available one layer later; and the final target language is resolved only at later layers. The ordering of the English default before content also appears in a reading-comprehension task. Our sparse autoencoder analyses reveal a further dissociation: the intermediate language signal is only weakly represented by sparse features, whereas a later language distinction is clearly sparse, but resists intervention at the patched position. These results provide a mechanistic account of how language selection and content information emerge and interact during multilingual generation.
Self-recurrence represents an important component for temporal processing in neural networks, both artificial and biological, where it appears naturally in the form of leaky integration. However, due to the complex temporal dependencies of the gradient, an efficient and effective method of online training is still lacking. Here we propose K-prop, a novel online learning method for leaky neural networks. It dramatically reduces the computational and memory cost of real time recurrent learning (RTRL) by propagating kernel gains. K-prop also outperforms existing RTRL approximations and achieves performance comparable to backpropagation through time in sequence learning, classification, and reinforcement learning tasks. We further analyze how its sole approximation affects gradient quality and identify conditions under which K-prop remains accurate and effective.
This work proposes WaveGen, a fully end-to-end diffusion-based text-to-waveform framework based on Spectral Forcing, which jointly learns an internal spectral trajectory and the target waveform trajectory within a single model, enabling direct end-to-end text-to-waveform generation. Unlike recent TTS pipelines, WaveGen does not rely on separately trained or externally decoded intermediate representations, such as neural audio codecs, Mel-spectrogram vocoders, or VAE-based latent models. In addition, WaveGen performs internal text-speech alignment within a single model, eliminating the need for external alignment modules such as duration predictors. To preserve a truly end-to-end formulation, our framework further avoids dependence on self-supervised representations, such as BERT-based text embeddings and Wav2Vec 2.0-based semantic representations, as well as semantic distillation methods for training optimization. To achieve this, we carefully design Spectral Forcing, a model architecture with disentangled diffusion heads, and a training objective based on a multi-scale spectral $v$-loss. We further validate the scalability of end-to-end TTS models by scaling the model size from 0.1B to 0.6B parameters. Using a single architecture, open-source data, and one-stage training, without external modules such as duration predictor or REPA, WaveGen demonstrates fully end-to-end waveform generation achieving state-of-the-art performance on the Emilia training dataset. In particular, WaveGen achieves a WER of 1.47 and a SIM of 0.75 on the Seed-en benchmark. Source code and checkpoints are available at https://github.com/sh-lee-prml/WaveGen
Vision-language models (VLMs) are increasingly used as automated judges for image-editing benchmarks, yet their reliability is poorly understood. Auditing these judges is difficult because real edited images lack controlled ground truth and entangle the requested changes with model artefacts, unintended side effects, and ambiguous degrees of success. In this paper, we introduce EditJudge-Bench, a diagnostic benchmark of 1,382 synthetic before/after image pairs rendered in Blender, with edits parameterised across nine categories derived from real user requests and exact ground truth from known 3D scene parameters. Because every edit is generated by a known intervention on a controlled scene, EditJudge-Bench evaluates VLM judges as binary verifiers of whether an instruction was correctly executed, isolating judge perception from editor quality. Across 25 judge configurations, we find substantial reliability gaps: (a) the widely adopted VIEScore template reaches only 69.0% AUROC with GPT-4o; (b) fine-tuned judges score 51.7-66.3%; spatial edits such as rotation and shape remain challenging; (c) prompt-formatting choices like image-text ordering shift performance by up to 14.1 points. Despite using only synthetic data, judge rankings on EditJudge-Bench correlate strongly with two real-image human-preference benchmarks ($\rho$=0.95 and $\rho$=0.71), indicating that controlled audits surface failure modes relevant to real evaluation.
"System 1" decision models (pioneered by TypeSafe AI's Jev) replace autoregressive text generation with a single forward pass returning typed probability distributions over declared schemas (Choice, Score, Noul). While deployed for low-latency routing and automated triage, their statistical calibration and post-hoc explainability remain uncharacterized. We introduce an open, deterministic audit benchmark evaluating System 1 decision models across calibration under label-preserving perturbations, cross-question interference, and post-hoc attribution faithfulness. Evaluating five open-weight models (0.8B to 27B) alongside generative System 2 baselines -- both single-token logprob scoring and structured JSON/CoT decoding -- across NLI (e-SNLI), adversarial syntactic shortcuts (HANS), synthetic clinical triage (N=900), prompt injection defense (Tensor Trust, N=250), and an 86-case logic-gate trigger suite yields five findings: (1) provider confidence is an exact closed-form dispersion transform (1 - H(p)/ln K on sub-10B models; normalized margin on Clef-27B) rather than empirical accuracy; (2) bundling multiple questions in one request induces directional cross-question interference (11.7% to 22.4% binary flips on Tev1-0.8B, vs. 0.1% to 0.2% on Tev1-4B); (3) System 1 models with >= 4B parameters are faster and better calibrated under syntactic shifts (ECE_w = 0.030 to 0.087) than logprob (0.144) and JSON (0.378) baselines; (4) for prompt-injection defense (N=250), >= 4B models achieve 0.988 to 0.998 AUROC with up to 95.2% recall and 0.0% false alarms at 48 to 60 ms; and (5) single-pass inference makes 512-coalition XAI interactive (1.3 to 3.2 s/input), while KernelSHAP with a logit link (S1-SHAP) mitigates softmax saturation on disjunctive OR gates (89.3% to 100.0% gate recovery vs. 57.1% to 82.1% in prob space; p_adj = 0.004).
Accurate prediction of electronic Hamiltonians would enable broad property inference while avoiding the high computational cost of Density Functional Theory (DFT). However, progress toward general-purpose materials foundation models is limited by a data bottleneck: existing Hamiltonian datasets are typically small, lack structural diversity, and often omit essential relativistic physics such as spin--orbit coupling (SOC). We therefore construct UniHam, a large-scale Hamiltonian dataset and benchmark suite comprising 100,000+ DFT-computed complex-valued Hermitian Hamiltonians with full SOC, covering 72 elements and a wide range of crystal geometries and symmetries (spanning diverse lattice types and space-group families). Building on UniHam, we benchmark two representative state-of-the-art models under a standardized protocol and introduce complementary evaluation metrics that jointly assess three dimensions: (i) Hamiltonian reconstruction accuracy, (ii) out-of-distribution (OOD) generalization across composition/symmetry shifts, and (iii) the ability to support downstream property prediction from the predicted Hamiltonians. Experiments on UniHam demonstrate that the proposed benchmark and metrics effectively differentiate model capabilities, revealing intrinsic SOC- and element-dependent failure modes, large variations in compositional OOD robustness, and the necessity of spectral-level evaluation to assess whether Hamiltonian predictions reliably support downstream electronic-structure properties. Overall, UniHam provides a reproducible, SOC-complete benchmark that can sharpen model comparisons and accelerate the development of next-generation foundation models for quantum materials.
Does grouping translation updates before speech synthesis change the measured completion gap between two translators? We compare StreamAtt with its fixed-audio-history baseline on ten English-German talks. For each translation, we keep the intended text and its availability times fixed and vary how updates are grouped for synthesis. Four-second grouping increases StreamAtt's mean completion lead from 106.6 to 229.1 seconds with XTTS-v2, but reduces it from 169.5 to 146.5 seconds with Piper. Both produce shorter audio; different changes in playback idle account for the opposite effects. These directions recur on a separately selected twenty-talk panel. In continuing CosyVoice 3 sessions, grouping instead leaves paired audio unchanged and delays completion on nineteen of twenty blocks. A speech-timing comparison therefore needs to specify how text enters synthesis, not just which synthesizer is used. Recognition and automatic word timing describe the outputs, but do not establish translation accuracy or intelligibility.
AI-text detectors reach near-perfect scores in evaluations that contrast fully human-written texts with texts generated from researcher-designed prompts. In authentic student writing, however, AI use is voluntary and often partial. We study this setting with the PISA 2025 FLA dataset, in which 15-year-olds in Germany wrote English argumentative essays with and without access to a gpt-4o-mini-based chatbot. Keystroke, clipboard and chat logs let us trace every character of the final essay to its origin. From this provenance, we derive text- and sentence-level labels for human-written, AI-generated and mixed text, and for text written after chatbot use without traceable AI material. We evaluate a feature-based detector and a classifier on frozen Llama-3.1-8B embeddings, both adapted from top-ranked shared-task systems. Both separate AI from human text reliably (macro-F1 up to 0.96 for texts and 0.95 for sentences), independently of spelling errors. The feature-based detector, however, relies on lexical complexity, which a simple student-style rewriting prompt removes (AI recall 0.92 $\to$ 0.60 with gpt-4o-mini, 0.41 with Qwen3). The embedding-based detector stays robust for whole texts (AI recall $\geq$ 0.94) but not for single sentences (AI recall as low as 0.59). Mixed text falls between the two classes, and text without traceable AI material is classified as human-written in about 95\% of texts.
Answer accuracy alone does not distinguish failures in acoustic perception, rule knowledge, and rule application. AuDiR (Audio Diagnostic Reasoning) examines these distinctions through controlled input interventions. Each item contains an audio clip, a multiple-choice question, a task-general rule, and a textual description of the acoustic observations needed to answer. Three conditions progressively supply the rule and observations; a fourth removes the audio while retaining the question and rule. The first correct diagnostic condition assigns an operational response label, without identifying a unique internal cause. The current benchmark contains 201 items in 18 families drawn from 9 sources, with answer references documented separately from percept extraction. In a preliminary evaluation of Gemini-3.8-flash, accuracy is 29.3% without supplied information, 31.3% with the rule, and 98.5% with the observations also supplied. No-audio accuracy is 23.9%, with a confidence interval containing the 25.0% chance rate. A family-level knowledge probe associates incomplete rule recall with larger gains from rule provision. These results show the benefit of supplying acoustic observations as text, while incomplete validation, possible recognition shortcuts, and non-monotonic responses limit stronger failure-attribution claims.
Multi-view anomaly detection is crucial for industrial inspection to identify defects hidden from single viewpoints. Yet, despite their strong success in single-view applications, reconstruction-based methods remain largely underexplored in multi-view settings. A key challenge in extending these methods lies in generating view-consistent pseudo-anomalies. To address this, we propose PAR, a reconstruction-based framework for enhanced multi-view anomaly detection. First, we introduce a feature perturbation strategy coupled with a consensus aggregation module that integrates these perturbed features across viewpoints to mitigate view-inconsistent anomalies and construct highly aligned pseudo-anomaly representations. Second, we employ a view-specialized multi-decoder architecture; by assigning a dedicated decoder to each view, we increase the model's representational complexity, enabling the generation of highly feasible, localized anomaly maps rather than less precise, view-agnostic reconstructions. Last, we incorporate Fourier attention into the decoders to emphasize global frequency-domain structures over local token evidence, effectively preventing identity mapping. Extensive experiments on the Real-IAD dataset demonstrate that PAR consistently outperforms state-of-the-art methods in sample-level, image-level, and pixel-level anomaly detection. Our code is available at https://github.com/SpatialAILab/PAR_mvad.
We investigate convolutional-neural-network (CNN) based automatic solver selection for sparse linear systems, a problem of practical importance where solver performance varies dramatically with matrix sparsity structure. A CNN is trained to regress log timing ratios between candidate solvers using spy plot images of matrix nonzero structure as input, achieving strong prediction accuracy on both held-out SuiteSparse matrices and real-world matrices from high-performance simulation codes.
Foundation models can supply predictions for a downstream task that draw on knowledge from extensive pretraining, but they differ in task-specific accuracy and query cost. We develop a framework that improves prediction for the task under a budget on the average query cost per input, using training data to learn which models to query and how to combine their outputs with features available without querying. The first method queries the same models for every input; the second lets them vary across inputs and uses Q-learning to decide which model to query next and when to stop. For the first method, we show that several individually inaccurate outputs can jointly outperform the best single one. It attains the minimax rate when the regression function depends on the outputs only through models within the budget. Otherwise, any method within the budget can be forced to a slower rate, which the first method attains. For the second method, we show when varying the queried models across inputs strictly improves on querying the same models for every input, and bound its prediction risk and average cost in finite samples. In simulations and a multimodal laptop price application, both methods outperform existing methods.