Papers

Diffusion language models (DLMs) promise fast parallel generation, yet high-quality samples often require large number of refinement steps, which diminishes their advantage in practice. This has led to massive interest in and rapid development of new methods for effective few-step generation. We show that much of the supposed quality gap at few steps can instead arise from a suboptimally configured sampler. Modest sampler sharpening, without any model retraining, enables a couple years old masked DLM to rival supposedly far improved successors. This differently sampled DLM in fact achieves lower generative perplexity in just 16 steps than what its standard sampler obtains with 1024, while improving both judged quality and semantic diversity. We further show that conventional per-output metrics can fundamentally obscure these gains, since any optimal trade-off between two such metrics can be attained by a generator supported on at most two outputs. We subsequently introduce GroupEval, which separately evaluates quality and across-output semantic diversity, and offers fresh insights including uncovering how 1.5-4.7x perplexity gains of a distilled model yield no corresponding quality gain. Finally, we explain why sharpening helps: parallel unmasking destroys dependencies among simultaneously generated tokens, creating a gap between prediction and generation. We prove that pervasive temperature choice of one is generically suboptimal under parallel sampling even for an exact denoiser, and that worse predictions can yield better samples. Through these results, we argue for a broader evaluation principle of treating the deployed generator as the object of comparison, benchmarking it against tuned baselines, and assessing quality and diversity jointly and with more human-aligned measures.

Detecting failures in World Model (WM)-based planning requires monitoring whether the model is behaviorally aligned with the current task, which in turn requires studying its internal representations. Here, we show that centroids—sub-component Jacobian row-sums—effectively identify the behavioral properties of WMs, complementing traditional activation-based knowledge signals. The centroids of a model are easily computed through Jacobian vector products and characterize how the model organizes the geometry of its input space, yielding an efficient perspective on internal representations, including the generation of task-relevant saliency maps. Evaluated on continuous control tasks using JEPA WMs, this behavioral view reveals a structural dissociation, where the encoder correctly represents the goal while the predictor remains behaviorally unresponsive. This failure mode directly predicts planning failure before any action is taken, allowing for goal resampling to recapture out-of-distribution success. Moreover, centroid-based methods outperform baseline methods as distribution-shift detectors. Together, these tools yield a behavioral monitoring stack that is operational and consequential under distribution shifts.

Recent research argues that language models exhibit pronounced homogeneity in open-ended generation, framing such behavior as an Artificial Hivemind that poses a long-term threat to human creativity. We examine three of its central results. First, the flagship example is that model responses to "Write a metaphor involving time" collapse into two clusters. Visualization, spectral analysis, clustering, and language model labels all contradict this description. The labels record each response's vehicle, what it compares time to. Our responses and the original authors' own show one dominant vehicle plus a heavy tail of distinct minority vehicles. "Time" is one of our least diverse topics, so the example is a favorable case, not a representative one. Second, the paper measures homogeneity against an undemanding null: responses to unrelated prompts. Under a more demanding null (same-prompt responses expressing genuinely different ideas), 20%-32% of such pairs already exceed the paper's 0.8 convergence threshold. A residual effect survives this null. The paper's same-prompt pairs exceed 0.8 roughly two to three times as often as our different-idea pairs. Much of what the paper calls homogeneity is the shared geometry of answering the same prompt. The remaining measurements lack any null: no human baseline is collected, and the model-indistinguishability statistic has no null. Third, the paper concludes that inference-time interventions are inadequate for combating the Artificial Hivemind, writing that "more generalizable solutions are needed at the model training level." We show that this conclusion is unsupported in three ways, and that an inference-time intervention (prompting) reliably raises measured response diversity. We do not resolve whether the Artificial Hivemind is real. We show that the published evidence does not establish it.

Extending forecast lead times while maintaining predictive skill remains a major challenge in environmental forecasting. We investigate diffusion-based rollouts as a stabilization mechanism for recursive forecasting using low-dimensional water-level time series and high-dimensional precipitation fields. Across both modalities, diffusion suppresses recursive error growth, with the largest stabilization occurring where deterministic rollouts are most unstable. However, stabilization does not guarantee forecast fidelity. In the water-level experiments, forecasts progressively lose event-level fidelity as the rollout loses access to external predictive information, and trajectory-level comparisons show that diffusion can remain numerically stable while contracting toward central values and exhibiting reduced variability. In the precipitation experiments, which retain conditioning from numerical weather prediction throughout the rollout, diffusion better preserves spatial organization and event-detection skill. Together, these contrasting experiments indicate that diffusion can control recursive error amplification, while its practical benefit also depends on the predictive information available to constrain future evolution.

This study explores the optimization of the Phi-2 Small Language Models (SLMs) for real-time chatbot applications through Parameter-Efficient Fine-Tuning (PEFT) and Quantized Low-Rank Adaptation (QLoRA). QLoRA specifically refers to the integration of PEFT with LoRA alongside a 4-bit quantization process, aimed at enhancing computational efficiency. These models, initially designed for high performance with minimal computational overhead, are further refined to address the constraints of mobile and edge computing environments. By integrating PEFT with QLoRA, the research aims to reduce memory usage significantly while maintaining, or potentially improving, the accuracy of model responses in real-time interactions. The effectiveness of these techniques was evaluated using the ROUGE metric system, which showed notable improvements in the summarization tasks performed by the models. This approach not only confirms the feasibility of using SLMs in resource-restricted environments but also opens up new avenues for deploying advanced AI-driven applications in real-time settings. The study's findings have significant implications for the development of efficient, scalable, and accessible AI technologies, paving the way for broader adoption in various industries.

A single random Gaussian probe gives an unbiased estimate of the squared Frobenius norm of a layer's quantization error. The estimator is well-behaved because round-to-nearest error is spectrally flat. Across 1,683 tensors from a 35B MoE and a 9B dense model, effective dimensionality is 0.93 to 0.96 times the i.i.d. noise value of the same shape, and on the MoE the median is unchanged from 2-bit to 8-bit. The probe coefficient of variation is predictable from tensor shape. One probe measures per-tensor sensitivity to within 4 to 7%; twenty probes reach 1.3 to 1.4%.RAM applies the propagated form of this estimator to budget-targeted mixed-precision quantization with no calibration data. Gaussian probes carrying the network's own input statistics score every tensor at six bit-widths. A knapsack solver allocates bits under an exact byte budget, with guardrails against catastrophic 2-bit assignments. One probe pass serves any budget. Isolated and propagated scores rank tensors independently on Qwen3.5-35B-A3B (Spearman -0.01), yet the propagated probe rank-correlates 0.81 to 0.83 with the GPTQ layer objective from real activations, while the isolated estimator is uncorrelated with it. That objective is the wrong allocation target: at matched bytes on Qwen3.8-27B, a block-output probe beats a vendor IQ3_M mix and an oracle that allocates from the real-activation objective. On Qwen3-8B the propagated probe ties HAWQ-V2 at matched bytes. Across seven architectures from 8B to 122B, with probe timing up to a 400B model in nine minutes on one workstation, RAM reaches 3.5 to 13.6% lower median WikiText-2 perplexity than size-comparable uniform 4-bit builds on the tested MoE models. (Black Sheep Ai baa.ai)

Progress in machine learning cannot outpace our ability to verify it. With an explosion in papers today, every scientific claim rests initially on trust in the trainer, leading to uneven evaluation, baselines, and forestalling of reliable progress. Traditionally, the burden of verification falls on the reader, who must reproduce expensive training runs. This strategy is impractical due to an explosion in slop contributions, diversity of methods, and the sheer compute required. We put the burden of proof where it belongs, on the trainer, and in the process also cut the overall cost of verification significantly. We introduce Witnesses, a method for certifying training, data usage and evaluation in a neural network training run. Our key insight is that fast behavioral fingerprints with occasional replay challenges are sufficient for auditing neural network training. Our method is applicable at scale with minimal overhead to the trainer, is cheap for the verifier, rejects bad training runs with amplifiable probability, and allows for exact queries of both data inclusion and exclusion. We test our method on language model training runs from 100M to 2B scales, across DDP and FSDP, and demonstrate this minimal overhead. We also introduce a self-regulating leaderboard of "auto-certified" training runs that enables shared baselines and progress. We invite the community to participate in the leaderboard to improve reproducibility in machine learning.

Generative models often suffer from mode collapse and limited sample diversity. While prior works attempt to mitigate this by jointly generating a batch of samples and repelling their trajectories, these heuristics do not explicitly maximize the diversity of the resulting endpoints. We introduce JIVE, a training-free framework that enhances generative diversity by injecting velocity perturbations aligned with the leading right singular subspace of the generator's endpoint Jacobian. By leveraging this local geometric structure, JIVE provably maximizes endpoint diversity while preserving sample quality. To maintain practical efficiency, we compute these perturbation directions via matrix-free iterations rooted in classical numerical linear algebra, requiring only a small computational overhead. Across different benchmarks, JIVE boosts both pixel and feature-level diversity in few-step and one-step generation.

We consider the behavior of deterministic, full-batch, bias-corrected Adam in separable linear classification with softmax parametrization under log-loss. In this setting, under a wide range of conditions Adam is known to approach max-norm-margin optimality when its stability constant $\epsilon$ is zero, while with a positive $\epsilon$, it is known to approach Euclidean-margin optimality. Our main contribution is the quantitative description of Adam's behavior for small fixed positive $\epsilon$. We give sufficient conditions under which an Adam-trained classifier nearly maximizes the max-norm margin before the updates become gradient-like. We also show that the classifier reaches a fixed target Euclidean margin only much later. Specifically, we show that for polynomially decreasing stepsizes with exponent \(a\), where \(1/3<a<1\), the updates become approximately proportional to the negative gradient after $\Theta(\log(1/\epsilon)^{1/(1-a)})$ iterations. At that time, the classifier still nearly maximizes the max-norm margin. Reaching a fixed target Euclidean margin above that of every max-norm-optimal classifier, but below the optimum, is shown to require $\epsilon^{-\Theta(1)/(1-a)}$ iterations. Under inverse-linear stepsize decay (\(a=1\)), the update transition takes polynomially many iterations, whereas reaching the target margin takes exponentially many. Experiments support these predictions. The later change in the classifier can improve or worsen generalization after training error reaches zero, connecting the analysis to grokking and its reverse.

Learning properties of neural networks has recently attracted growing interest, with existing approaches operating either directly on network parameters or through probe-based representations of network behavior. While probing methods have shown strong empirical performance, their theoretical foundations remain limited. In this work, we study when finite probe-based representations are sufficient for learning neural functionals. We establish general identification and universality results for probing, and show that using intermediate hidden representations can provide significantly more informative representations than relying only on final outputs. Motivated by these results, we introduce HIDDENPROBE, a simple architecture for learning from hidden probe responses. Across a range of neural functional benchmarks, including both MLPs and Transformers, HIDDENPROBE consistently improves over existing probing methods and achieves state-of-the-art performance. Our code is publicly available on GitHub.

Training efficiency has become the central driver of recent progress in foundation models. To overcome the massive computational and data requirements of large-scale training, researchers increasingly adopt strategies such as selective data sampling, efficient pre-training, and simplified reinforcement learning pipelines. While these strategies drastically reduce overhead, they prompt a critical, yet neglected question: Is efficiency achieved at the expense of model robustness and security? To our knowledge, we present the first systematic cross-domain investigation of the efficiency-vulnerability trade-off. Across vision and language models, we show that efficiency-oriented training increases susceptibility to adversarial and privacy attacks. We characterize this vulnerability by analyzing the models' internal geometry and functional representations, demonstrating that the evaluated efficient variants consistently exhibit sharper loss geometry together with systematic changes in representational structure. We further extend our analysis to "zero RL training", finding that models trained using simplified RL recipes exhibit substantially greater susceptibility to catastrophic forgetting and more pronounced overconfidence than those trained through conventional alignment pipelines. Our findings suggest that training efficiency is rarely a "free lunch"; rather, the mechanisms that minimize computation can inadvertently compromise safety. We conclude by calling for a paradigm shift toward multi-objective training that jointly optimizes for performance, cost, and security.

In diffusion-based generation, a neural network can be trained to predict the clean data, the noise, or the velocity from a noisy input. These prediction targets are interconvertible and describe the same generative process, yet plain Diffusion Transformers operating on large pixel patches succeed with clean prediction and fail with noise or velocity prediction. We argue that this asymmetry arises because noisy targets require the residual stream to preserve noise-dependent input variation through depth for the final readout, forcing subsequent layers to compute on noisy representations. A spectrally concentrated clean target imposes a lighter demand, leaving greater freedom to organize hidden representations for subsequent computation. We call this preservation requirement *residual-stream burden* and show how it shapes representation learning in Diffusion Transformers. Controlled experiments indicate that the exploitable structure is spectral concentration in patch space and that the bandwidth of the persistent residual state is a key resource for noisy prediction. We further show that this account is consistent with recent decoupled pixel-space architectures, whose diverse designs all reduce the residual-stream burden on the main pathway. To examine this understanding from a complementary direction, we expand and reorganize the residual-stream bandwidth directly, introducing Spatially Indexed Hyper-Connections (SiHC) that reach FID 1.71 on ImageNet $256^2$. Together, these results identify residual-stream burden as a mechanism through which prediction targets and architecture jointly shape representation learning in Diffusion Transformers.

Large audio-language models (LALMs) produce fluent responses about audio but often hallucinate by making plausible yet ungrounded claims. Existing audio hallucination benchmarks mainly measure response correctness, leaving it unclear whether an LALM hallucinates or simply fails to understand the audio. We challenge correctness-based evaluation by defining two hallucination categories: (i) context, where claims are not grounded in the audio; and (ii) knowledge, where claims about audio-related topics lack support from externally verifiable facts. We introduce MISHAP-Bench, a comprehensive benchmark with 12,000 challenging open-ended question-audio pairs and a rigorous evaluation pipeline covering both categories. To evaluate open-ended responses, we propose a groundedness judge that uses reference rubrics and judge prompts guided by human annotations. We extensively evaluate ten state-of-the-art LALMs and show that hallucination remains substantial. Even a frontier model such as Gemini 3.7 Flash reaches a hallucination rate of 36.5%. We further adapt and benchmark four mitigation methods from multiple domains for LALMs. Despite some improvements, effective hallucination mitigation remains an open challenge. Finally, we call on the community to evaluate hallucination and benchmark mitigation methods with MISHAP-Bench.

Multicalibration requires a predictor's residuals to be unbiased not only globally, but also after conditioning on the predictor's own level sets and reweighting by a rich class of test functions. Standard boosting approaches in the distributional setting achieve this by repeatedly discretizing the predictor's range then auditing and repairing the resulting level sets. One consequence is that in practice, the algorithm's guarantees are sensitive to this parametrization of the rounding parameter. A natural theoretical question, then, is how to do discretization-free boosting which avoids this rounding within the boosting process itself. Here, we analyze an alternative feature-augmentation boosting paradigm inspired by Tax et al. (2026): at each round, a squared-loss oracle is called on hypotheses that receive the previous predictor's output as an additional feature, and only the final predictor is rounded to have a finite set of level sets to provide the multicalibration guarantee with respect to. We give a theoretical analysis of this procedure through the expressivity of the augmented hypothesis class, and show how the expressivity of this class yields a hierarchy of guarantees, including multiaccuracy, multicalibration, and the stronger notion of level-set multicalibration.

At each step, decoding one sequence with a large language model rereads the projection weights, whose traffic is fixed, and the key-value (KV) cache, whose traffic grows with context. Activation sparsity trims the first term and KV-cache sparsity the second, yet their reported speedups are hard to compare because each depends on context length and on the dense attention kernel it is measured against. We derive a byte crossover, the context length at which the two savings are equal, together with ideal speedup bounds for each branch and for their composition, from model dimensions and keep ratios alone. We then time both branches and their composition from 2K to 128K tokens on two GPUs after a dense prefill of real text, with dense and sparse modes reading the cache through the same split-K attention kernel. The projection branch leads at short context and the KV branch at long context, with speedups that follow their byte bounds up to fixed kernel costs. Adding these costs, measured in separate sweeps, lets the byte account predict the measured crossings of three keep-ratio pairs, a second model, and a second GPU to within 4.1K tokens. Timing the dense baseline with masked instead of split-K attention inflates the apparent speedup of the same KV policy about fivefold. An attention-scored KV selection answers the same passkey and multi-key placements as dense decoding up to 127K tokens, whereas a KV window misses most of them. Under matched perplexity budgets, activation sparsity composed with this selection decodes 14 to 26% faster than the best single branch on both GPUs. Code is available at https://github.com/js-lee-AI/ByteCross.

Can vanilla policy optimization explore enough to achieve near-optimal regret in stochastic contextual bandits? We show that standard exponential policy updates driven by offline regression do so under realizability, without exploration bonuses or importance weighting. For $A$ actions, $T$ rounds, and a finite prediction class $F$, vanilla PO achieves $\widetilde O(\sqrt{AT\log(|F|)})$ regret with high probability. Our analysis reveals an implicit exploration mechanism of independent interest: gradual policy updates prevent actions from losing probability too quickly, allowing the regression oracle to learn their expected losses. We further develop a batched version using only $O(\log T)$ regression calls and policy switches, and show how private regression oracles yield differentially private contextual bandit algorithms without composition across batches. For a finite class, this gives pure $\varepsilon_{\rm priv}$-DP and regret $\widetilde O\left( \sqrt{AT \log(|F|/\delta)}(1+\varepsilon_{\rm priv}^{-1/2}) \right)$. Finally, experiments across oracle-based contextual bandit algorithms, with and without privacy, demonstrate the practical effectiveness of policy optimization and the value of explicit exploration under stronger privacy constraints.

Block-diffusion language models are served at hand-picked operating points, such as acceptance thresholds, buffer depth, schedule, checkpoint and precision, and each point is chosen by its mean benchmark accuracy. However, a mean does not tell an operator how often a faster configuration fails on prompts that the slower one answers correctly. On the serving engine and its decode traces, the default commit rule already commits every fully resolved block, a static skip rule captures nearly all of the compute that allocation can save, and self-distillation on engine-decoded targets adds speed at unchanged accuracy. Larger speedups come from lower thresholds, which commit tokens that are still uncertain. We therefore present Redline, a finite-sample procedure that selects operating points, hand-picked or learned, from the correctness of their answers on calibration prompts. Redline keeps the reference-relative risk, the joint probability that the reference answers correctly and a candidate configuration does not, within a user-chosen budget with high probability, and deploys the fastest configuration that passes. It speeds up math at a smaller risk budget than code in both model families, and at a budget of ten percent it deploys a LLaDA2 math configuration that commits over a third more tokens in each forward. It also applies without modification to the acceptance rule of speculative decoding and to weight quantization. On the same calibration data, Redline stays within its stated failure probability, whereas each tolerance of a mean-accuracy rule either gains less speed for some model and task or exceeds the risk budget far more often for another. Code is available at https://github.com/js-lee-AI/Redline.

Large language model (LLM) agentic systems increasingly rely on models communicating with one another, yet existing uncertainty and multi-agent methods rarely estimate how a particular receiver will interpret a message before it is sent. This matters in heterogeneous systems, where capable receivers can reconstruct different tasks from the same message. We model this as a sender-receiver problem with a latent receiver type and define prospective interpretation risk (PIR): the probability that a receiver reconstructs a task other than intended. Rather than model an LLM's full input-output behaviour, we use black-box probes relating messages, intended tasks, and receiver-specific reconstructions, yielding scalable supervision while separating interpretation from downstream capability failure. Offline, heterogeneous frozen receivers provide supervision for receiver-conditioned risk and the effects of predefined mutable message features. At deployment, history induces a posterior over receiver types, guiding message revision and selection. We introduce value of interpretation information (VoII), querying for receiver information only when its expected communication benefit exceeds its cost. Our theory characterises when receiver information has decision value and bounds such queries. Empirically, interpretation-failure rates vary by 4-13x across receivers. Receiver information reduces PIR calibration error by 68% relative to a receiver-agnostic predictor, largely by correcting receiver-specific risk levels. PIR-guided revision reduces interpretation failure by 44% relative to the original message and 40% relative to a generic rewrite, mostly through a repair that helps every receiver. VoII outperforms information-gain and random querying at matched cost on the interpretation objective it optimises, lowering interpretation failure from 3.84% to 3.79% while querying 18.2% of episodes.

Convolutional EEG decoders are trained at a fixed width, usually set by their authors on other data. Growing methods add neurons during training where the loss could decrease the most, but whether they improve compared to a reference width is untested on EEG. Here, we grow three convolutional backbones on 12 motor-imagery datasets under three protocols and compare each with its reference model per subject. The growing ShallowFBCSPNet scores 2.9 points above its reference model with only half the parameters (0.57x), SCCNet changes by at most 1.2 points. Deep4Net growing models show decreased accuracy, but they require adaptation that prevent to compare faithfully the results. These differences follow the selection step, which keeps a candidate neuron relying on a dynamic threshold from singular values decomposition. Overall, these results suggest that growth helps when its criterion can rank the candidate neurons, and that the rate of skipped neuron addition tells where a decoder can be grown small from scratch.

Neural surrogates can accelerate large-scale optimization by replacing expensive or intractable model components with efficient learned approximations, but solving the resulting embedded problems can remain prohibitively costly. For instance, standard exact encodings of neural networks with ReLU activations allow the problem to be solved by mixed-integer solvers, but add large numbers of binary variables to accommodate the nonlinearity of the activations, which can render the problem computationally prohibitive. To address this, we propose DCEmbed, a heuristic for optimization problems with embedded neural surrogates that leverages the difference-of-convex (DC) representation of the network and avoids adding activation binaries. Exploiting shared structure within the DC representation of a ReLU network, we derive a reduced-size, exact formulation for its convex components that can be embedded in optimization problems using just two linear inequalities and one continuous auxiliary variable per hidden neuron. Using this, the problem is solved via an iterative penalty convex-concave procedure, where only the concave portions of the neural terms are approximated at each stage. The original objective, constraints, and any discrete decisions are retained, allowing standard convex or mixed-integer optimization solvers to optimize the host and surrogate jointly at each iteration. In experiments on quadratic programs, mixed-integer resource allocation, and neural two-stage stochastic programming, our method demonstrates much faster progress toward high-quality feasible solutions than approaches using exact mixed-integer embeddings. In particular, DCEmbed achieves $4\times$ lower normalized primal integral than the best exact baseline on the resource allocation problem, while in two-stage stochastic programming it reaches the global surrogate optimum $\sim 5\times$ faster than Gurobi ML.

Merchant matching resolves a noisy payment descriptor to a retrieved merchant entity or returns no match. A key challenge in curating training labels is distinguishing teacher abstention from evidence that no acceptable entity exists: false no-match labels contaminate pseudo-labeled data, while conservative labeling reduces coverage. We investigate whether agreement between two local large language model judges improves pseudo-label reliability. A label is retained only when the judges agree, with separate ordered thresholds for selections and abstentions that guarantee disjoint positive and negative label sets. Retrospective replay on 2,000 expert-annotated queries shows that higher selection thresholds can improve positive-label purity, whereas higher abstention thresholds increase false no-match labels. At thresholds (0.86, 0.80), Muse Glimmer 30B and Gemma 4 31B jointly label 1,633 queries (81.7% coverage) at 96.88% purity; positive and negative purities are 99.47% and 93.38%. This exceeds either constituent model at the same thresholds by more than two percentage points, with lower coverage. A split-half check finds only 0.14 percentage points of threshold-selection optimism. A symmetric threshold of 0.86 adds 40 erroneous no-match labels, while 46 false abstentions persist even with no confidence threshold. Across five matched within-model comparisons, higher reasoning effort yields no clear F0.5 gain and increases median latency by 1.8-5.0 times. These results motivate separate thresholding and auditing for positive and negative pseudo-labels. The study establishes label purity, not student utility; fresh-data curation and student fine-tuning remain necessary to demonstrate downstream value.

Outcome-only reinforcement learning gives software engineering (SWE) agents a terminal success signal but little direct guidance about intermediate decisions. We introduce Counterfactual Rollout Replay (CRR), a training-time procedure that uses forkable executable environments to obtain step-level return contrasts. CRR selects a small set of decision points, restores each state, samples an alternative action, and rolls the branch forward under the policy. It retains the realised training trajectory and replaces the advantage at selected steps with the difference between its terminal return and the sampled counterfactual return. The method needs no human process labels or learned process reward model; free refers to those supervision costs, not replay compute. With a 14B policy, CRR improves pass@1 on SWE-bench Verified, SWE-bench Live, and SWE-rebench, and combines with process-reward and trajectory-search methods. On SWE-bench Verified, an equal-wall-clock comparison on the same hardware yields 41.7% versus 36.7% for extended outcome-only GRPO, a 5.0-point gain with fork overhead included. These results apply to environments with affordable, reliable state restoration; stochastic continuations and expensive or imperfect replay remain limitations.

JET uses pretrained language and vision-language models to select among a finite set of answers without additional training. It evaluates candidate likelihoods directly and shares computation across candidates. Experiments on desktop CPUs and consumer GPUs assess decision accuracy and execution cost. Qwen3.6-35B-A3B achieves 87.48% accuracy on the full MMLU test set and 3.69 requests per second on a separately timed MMLU subset. The accuracy-throughput comparison covers model, hardware, and reasoning choices, with Jev as an external reference. Controlled execution experiments show 2.18-2.23-fold speedups from prefix reuse and cache management, and a 30.8% reduction in process time from input preparation optimizations, with unchanged outputs. Optional reasoning has a task-dependent accuracy-throughput trade-off. These results support local decision inference from existing models.

Robotic manipulation policies are advancing rapidly with increasing reliance on vision-language models for end-to-end decision making. However, reliable deployment remains challenging because many policies lack explicit mechanisms for predicting task outcomes and evaluating whether generated actions will achieve desired final states, causing execution errors to accumulate during long-horizon manipulation. We present Robot-GST, a geometry-aware spatio-temporal behaviour representation and evaluation framework that constructs a Gaussian-SAM robotic environment for real-to-sim policy verification and improves the reliability of real-world manipulation deployment. Our approach constructs a high-fidelity robotic environment from RGB-D observations using 3D Gaussian Splatting and SAM3D, enabling “simulation and evaluation before acting”. It integrates visual observations and language instructions with spatio-temporal reasoning for long-horizon task planning using large vision-language models. To bridge high-level planning and real-world execution, we introduce Gaussian-aware final-state estimation through geometric sampling and state-based trajectory planning. Before execution, candidate action sequences are simulated and evaluated in the Gaussian-SAM environment to filter infeasible behaviours. We validate our approach on representative manipulation tasks involving rigid, soft, and deformable objects, including cube placing, toy packing, and duck rearrangement, demonstrating that geometry-aware spatio-temporal reasoning and state-aware execution improve manipulation reliability across different object categories. Our results suggest that combining geometry-aware reconstruction with high-quality rendering and simulation provides a scalable approach for evaluating robotic manipulation behaviours. Website: https://robot-gst.github.io

Recovering hidden transport mechanisms from sparse spatiotemporal observations is a fundamental inverse problem in scientific machine learning. In brain tracer imaging, dynamic contrast-enhanced MRI (DCE-MRI) provides time-resolved measurements of tracer concentration, while the underlying velocity and source mechanisms governing tracer propagation remain unobserved. We formulate this problem as physics-informed latent-state inference, in which the transport field itself is the primary object of inference rather than an auxiliary variable used only to reconstruct observed densities. We propose Physics-Informed Neural Optimal Mass Transport (PI-NOMT), a framework that represents density, velocity, and source as continuous neural fields and combines a continuous neural density teacher, recursive differentiable advection–diffusion–source rollout, unbalanced optimal-transport regularization, and governing-equation supervision. Physical laws act as structural priors that constrain the space of admissible transport mechanisms, while observed tracer dynamics provide evidence for estimating the latent transport state. We evaluate PI-NOMT on a synthetic benchmark with known ground-truth transport and on DCE-MRI sequences from nine control rats. On the synthetic benchmark, PI-NOMT accurately recovers the prescribed velocity field, including its magnitude, direction, and integrated trajectories, rather than merely reconstructing endpoint densities. Across the nine rat datasets, the framework yields sub-percent local endpoint error, consistent physical speed scales, and low post-training PDE and incompressibility residuals. These results support physics-informed latent-state inference as a general framework for recovering hidden transport mechanisms from observed dynamic scalar fields.

Self-evolving vision-language models train on questions they generate from unlabeled images. Since these questions have no gold answers, prior methods label them by majority vote over sampled answers or by a model judge. In a human evaluation, we find that 24% of majority-vote labels and 18% of model-judge labels produced during self-evolution are wrong. To address this problem, we present Verifiable QA Generation for Self-Evolving Models (VQS), which changes how the model judges answers. Instead of voting on an answer, the model parses each image into a structured record, such as a scene graph, a chart table, or a diagram graph. Fixed programs then write a question from the record and compute its answer. The model still acts as a visual checker, but it only confirms the individual facts the program reads, one short claim at a time. These claim-level checks select the parser's training targets, so the parser also improves without labels. Human raters find 94% of VQS answers correct, against 76% for majority voting. Across ten benchmarks, VQS improves Qwen3-VL by up to 3.18 points at the 2B, 4B, and 8B scales and outperforms the strongest self-evolving baseline at each. Gains keep growing over three training rounds, reaching 3.84 points at 2B. Code is released at https://github.com/ahmedheakl/VQS

Contextualization, the core operation of language modeling, transmits information across words to build sentence-specific word representations. Prior works mainly study contextualization, focusing on individual words and attention heads as a growing discrete dictionary, lacking a global view of their general behavior. Therefore, we propose a general principle: Globally, we find and estimate that different words carry different amounts of information, and less-informative words tend to absorb more contextual information. Specifically, these low-information words do not absorb contextual words uniformly, and finer-grained selectivity enables more precise routing to promote information transmission between matched words. Moreover, to find what mechanism causes such processing, we reinterpret attention heads as channels gated by their singular vectors and find that: (1) these singular vectors point to the hidden states of more informative words, allowing such words to write their information to others more strongly to act as information sources, and vice versa; and (2) these singular vectors can be viewed equally as hidden state features, enabling automated interpretation of attention heads beyond prior heuristic head discovery, also embedding heads into a continuous space rather than treating them as discrete, independent dictionary entries.

Large language model (LLM) agents increasingly undertake extreme-long (xlong) horizon tasks, where a single execution can span hours, hundreds of model–environment interactions, and nearly 1M tokens per rollout. Applying online reinforcement learning (RL) to such executions poses two fundamental challenges: (1) severe execution variance and prolonged rollout delays cause massive GPU idling; and (2) complex non-linear branching generates massive trajectory redundancy, crippling training efficiency. To address these, we presents QwenGyre, an end-to-end framework for xlong-horizon online RL. QwenGyre elastically reallocates GPUs between rollout and training without interrupting live executions, while its trajectory processor reconstructs branching histories, scores partial progress, and deduplicates redundant paths to bound training costs. Scaled to our flagship model, Qwen~3.8 2.4T, with 700K tokens per rollout, QwenGyre yields a 6.0% absolute gain on NL2RepoBench (52.5% $\to$ 58.5%) in 48 steps. Across our evaluations on diverse domains of training datasets, QwenGyre delivers up to $1.85\times$ and $1.78\times$ speedups over Colocate and Async, respectively.

Modeling temporal dependence and uncertainty is central to forecasting with world models. The Visual Bayesian Regression World Model combines visual features, physical histories and known covariates through interpretable regression, within a modular architecture supporting trend, seasonal and cycle dynamics. Visual compression reduces representation dimension, while Bayesian variable selection reduces active regression dimension. Posterior prediction combines forecasts across predictor subsets using their posterior probabilities as weights and accounts for parameter uncertainty and future disturbances. The model forecasts joint visual–physical states recursively and physical targets directly. Across four forecasting tasks spanning object motion, vegetation greenness and solar power, ViBR-WM achieves lower mean overall physical-target error than Temporal Straightening, ConvLSTM, PredRNN and SimVP on every task. Repeated fitting and resampling support these overall gains.

The shipped Laya Typed-Decisions checkpoint, a 421M-parameter ModernBERT-large assessor that answers typed choice/noul/score questions over workflow state, is uniformly under-confident. The signed confidence-accuracy gap is $-0.214$, every occupied reliability bin's accuracy exceeds its confidence, and that sign uniformity collapses every binned ECE variant to the same value, $0.214$. The card frames the risk as over-confidence; the measured direction is the opposite, and the direction decides which way a confidence-gated cascade fails. A single disjointly fitted temperature ($T=0.469$, sharpening) removes most of the miscalibration (held-out ECE $0.204$ to $0.037$) and outperforms the shipped per-option-count table. The frozen selection rule instead chose isotonic regression, which overfit and failed its held-out NLL contrast on both tracks, so hypothesis H2 is not supported. Re-running the released checkpoint on its full official test split reproduces the card's headline accuracy ($0.767$ vs. $0.766$). The retrospective E1 reproduction preceded the analysis freeze; E2-E8 were prospectively preregistered, and 20 of 22 executed confirmatory tests reject under Benjamini-Hochberg FDR at $q=0.05$ (two descoped). The frozen gate beats random escalation but misses its 10% accepted-set error target on both tracks, an exploratory out-of-distribution probe finds no zero-shot transfer (accuracy $0.617$), and every score measures agreement with a synthetic teacher whose self-agreement ceiling ($0.735$) the specialist exceeds. Per-decision predictions, run manifests, and the frozen preregistration are in the ancillary files. The author has no affiliation with the model's publisher, the dataset's publisher, or TypeSafe.