Papers

Graph Neural Networks (GNNs) are widely used for representation learning on graphs, but most methods assume static topologies, making them inefficient on evolving networks where edges change over time. Existing dynamic approaches either model graph evolution through temporal GNN architectures without focusing on efficient dynamic maintenance, or are restricted to linear propagation models based on Personalized PageRank. In this work, we study how to efficiently maintain node representations for non-linear GNN propagation under edge insertions and deletions. The propagation has no learned parameters, and only a classifier applied afterward is trained. For a broad class of standard activation functions, we develop a residual-based dynamic algorithm that selectively propagates local errors via push operations, maintaining an approximation to the evolving fixed point without full recomputation. We prove that our method achieves amortized $O(1/\epsilon)$ update time per graph change under a degree-normalized error guarantee. Our approach uses a potential-based analysis in a degree-scaled norm and, in contrast to prior work on the linear case, requires no randomness assumptions on either the update sequence or the input vector. For the linear special case, we additionally provide an exact dynamic algorithm via low-rank matrix inverse updates. Experiments on benchmark datasets show that incorporating non-linearity improves accuracy while preserving efficient update performance, yielding a scalable and theoretically grounded method for maintaining this propagation on dynamic graphs.

Millimeter-wave (mmWave) radar and RGB-D cameras can record cardiac and respiratory waveforms continuously and without contact, but labeled recordings remain scarce because every label requires a supervised acquisition session. Self-supervised pretraining can exploit the unlabeled signals, yet contrastive methods depend on signal transformations and negative pairs whose validity is uncertain for cardiorespiratory data, where time warping changes breathing rate and distant windows can share the same physiological state. We present Phenomenon-Graph JEPA, a joint-embedding predictive architecture that learns from four processed one-dimensional streams without negative pairs or synthetic augmentation in its base configuration. Each stream is encoded by a temporal convolutional branch and a band-limited spectral branch. During pretraining, the model predicts stopped target embeddings along typed edges, which connect streams assigned to the same physiological phenomenon, and forward in time within a state episode. We treat this physiological typing as a testable hypothesis and compare it with wrong-edge and all-pairs prediction graphs. In the OMuSense-23 dataset, pretraining improves label-efficiency area over matched supervised training by 3.91 percentage points (95% interval 2.08 to 5.80, Holm-adjusted p = 0.006), and by 3.74 points under a second configuration evaluated on the same test participants. However, the wrong-edge and all-pairs controls do not establish a benefit from physiological typing. Optional Takens-inspired delay coordinates improve a validation comparison with learned history, whereas two wrist-only WESAD protocols do not establish a pretraining advantage. The study therefore separates the measured benefit of predictive representations from the physiological prior used to organize their training.

Transformer-based language models perform well on symbolic tasks, yet it remains unclear whether they learn generalizable rules or rely on statistical shortcuts. Mechanistic studies link algorithmic behavior to structured internal representations, motivating the hypothesis that robust reasoning benefits from separating values from the types that control their manipulation. Can making this separation an architectural primitive improve the learnability and generalization of logical mechanisms? We introduce STRAT (STratified Registers And Types), which partitions the residual stream into orthogonal Data and Type subspaces and uses Type-based attention and gating to govern Data transformations. Controlled arithmetic ablations identify three failure modes associated with data-control interference: the Linear Trap, Gradient Wall, and Open Gate Trap. Mechanistic analysis reveals interpretable logical structure, and in arithmetic, STRAT reduces median OOD error 35-fold relative to a Transformer baseline. On each of 11 datasets spanning 10 tasks, STRAT outperforms the Transformer baseline in mean accuracy, by 26 percentage points on average, with both models trained from 10 base examples per dataset using identical task-specific augmentation where applicable. Under distribution shift, STRAT's mean accuracy drops by only 2.39 percentage points, compared with 11.75 for the Transformer.

Repeated sampling can reveal a correct numerical answer without yielding either a reliable system output or a supported derivation. We present Coverage, Realization, and Validity Evidence (CRV), an evaluation protocol for sampled large language model (LLM) reasoning over formal geometry states. Coverage is answer availability, realization is readout accuracy on the frozen candidate pool, and validity evidence is a label-blinded critic judgment of derivational support rather than a proof certificate. CRV freezes each candidate pool before comparing readouts and analyzes covered failures by correct-answer multiplicity and within-problem discrimination. On HardShift441, a 441-problem set for which a reference solver leaves 406 problems unsolved, a LoRA-adapted Qwen2.5-7B generator obtains 24.2% average single-sample accuracy and 68.9% pass@16, whereas verifier-weighted self-consistency (WSC) reaches 38.0%. Readout accuracy is particularly low when the correct answer occurs only once or twice in the pool. In a separate constructed audit of 195 covered problems, the critic labels 12 correct-answer representatives as supported, 181 as refuted, and two as uncertain. These results show that coverage, realization, and validity evidence from the critic are distinct quantities and should be reported separately.

Wireless digital twins (DTs) rely on 3D environment models to predict radio propagation and support wireless-network decisions, yet these models are often initialized from imperfect 3D maps. Errors in building position, height, footprint, and orientation can therefore cause a high-fidelity propagation engine to simulate the wrong physical environment. In this paper, we study how a deployed wireless network can repair an existing DT using its own radio frequency (RF) measurements. In particular, we introduce Twin Residual Alignment and Calibration Engine (TRACE), a physics-grounded learning-based self-calibration framework that treats twin maintenance as residual alignment between the physical world and the current DT. Using the same sensing configuration as the physical measurements, TRACE ray-traces the current DT, coherently backprojects the measured and simulated RF onto a common world grid, and extracts the same local region around each building's current DT position. A multi-view corrector then fuses evidence across sensing nodes and neighboring buildings to predict a gated six-parameter correction per building, without relying on absolute layout or sensor ordering, and supports iterative correction through re-rendering. On 5,400 held-out samples from unseen simulated scenes at 28 GHz, TRACE reduces 3D position RMSE from 2.202 m to 0.302 m and yaw RMSE from 4.978{\deg} to 0.894{\deg}, outperforming ViT and U-Net baselines under changes in layout, building count, sensing-node count, and SNR. On measured 28 GHz RF data from the NIST outdoor courtyard, a model trained only on synthetic RF reduces mean planar wall-position error from 1.00 m to 7.8 cm, without measured-data fine-tuning or geometric labels. These results show that the discrepancy between measured and twin-rendered RF can serve as a learning signal for repairing a wireless DT.

Matrix multiplications dominate the inference cost of modern transformer-based vision models, yet existing efficiency techniques such as post-training quantization and mixed-precision inference are largely limited to the small set of fixed-width formats (INT4, INT8, BF16, and FP16) supported by conventional accelerators. We revisit stochastic computing (SC) as a way to lift this constraint: viewed as a dense adaptive quantizer, SC controls precision by bit-stream length L rather than a fixed datapath, while each multiplication reduces to a single AND/XNOR gate. We build a GPU library that emulates SC matrix multiplication at scale, exposes stream lengths as first-class kernel arguments, and evaluates SC end-to-end on image classification, object detection and instance segmentation, class-conditional image generation, and visual world-model planning. On top of this substrate, we develop a dynamic per-row mixed-precision policy that assigns stream length per token or group at matched average budget, requires no retraining, and uses the same SC hardware across schedules. Across tasks, SC remains competitive with fixed-format INT quantization at matched bit budgets, while per-row mixed precision helps maintain accuracy at lower average stream lengths. These results provide software-level feasibility evidence that SC can serve as a dense-precision substrate for fine-grained mixed-precision inference on modern vision transformers.

We introduce Adaptive LeWorldModel (ALeWM), a world model based on a joint-embedding predictive architecture (JEPA) that learns to concentrate predictive information in compact prefixes of a wide latent representation. To encourage this ordering, ALeWM learns a sequence-conditioned distribution over prefix lengths and trains the predictor to estimate the full next embedding from a sampled input prefix. As standard anti-collapse objectives encourage variation across latent coordinates and do not organize them by predictive importance, we also introduce MixSIGReg. MixSIGReg regularizes the masked embeddings against a prior-weighted mixture with Gaussian active prefixes and zeros in the remaining coordinates. As a result, the ALeWM objective encourages early coordinates to retain information useful for prediction and recursive planning. Our analysis shows that the mixture distribution used by MixSIGReg assigns higher variance to earlier coordinate blocks and lower variance to later ones. In addition, we show that, under specified assumptions, prediction error is minimized by placing the information most useful for prediction in earlier blocks. Empirically, we study the behavior of ALeWM in a controlled dynamical system with known state variables and in goal-conditioned visual control. We show that ALeWM consistently achieves higher mean success rates than tuned fixed-width LeWM, with lower planning capacity on average.

Reverse-mode differentiation of Monte Carlo light transport naively requires storing a computation graph whose size grows with path length, making it impractical for deep or high-sample simulations. Path Replay Backpropagation (PRB) eliminates this memory complexity for viewpoint path tracing by reconstructing paths from their random seeds and performing a constant-memory backward replay. However, this formulation relies on a one-path-one-pixel property that does not hold for light tracing: a single light path can contribute to many pixels, and the corresponding per-vertex adjoints cannot be recovered during replay from constant-size state. We demonstrate that this structural difference prevents a direct extension of PRB to light tracing in constant memory, and that buffered two-pass variants retain linear memory scaling with path length. We then introduce two constant-memory reverse-mode formulations for differentiable light tracing. The first, Reservoir Light Replay Backpropagation (ResLRB), compresses the set of per-path sensor connections into a single stochastically selected representative using weighted reservoir sampling, preserving a two-pass structure at the cost of increased gradient variance. The second, LRB-3-pass, introduces an additional traversal that accumulates downstream adjoint contributions prior to backpropagation, retaining all connections and matching naive reverse-mode AD in expectation. Both methods support detached and attached formulations, including differentiation of sensor splat positions induced by geometric perturbations. We validate correctness against naive AD, characterize memory and runtime trade-offs, and demonstrate inverse rendering of specular caustics driven entirely by light tracing.

Vision-language models (VLMs) are increasingly being used in scientific workflows, but their ability as agents to directly control laboratory experiments from visual feedback remains underexplored. This capability is important because many laboratory tasks do not naturally provide dense, pre-defined numerical objectives: informative signals can be sparse, intermittent, or visually ambiguous. A more general laboratory agent should instead be able to interpret visual observations, take actions to acquire useful feedback, and adapt its behavior based on the consequences of those actions. We study whether general-purpose VLMs can perform this kind of closed-loop scientific control using experimental optics as a testbed. We evaluate agents on three experimental systems that isolate distinct capabilities: a Michelson interferometer, a two-mirror cavity, and a four-mirror optical relay. The agents observe camera images, directly issue actuator and measurement commands, and retain their interaction history without receiving an engineered scalar objective during control. Across these experiments and matched simulations, we find that, given task-specific natural-language guidance, VLMs can estimate actuator-response relationships, resolve ambiguous observations through intervention, and actively create informative visual feedback when signals are sparse. These results suggest that pretrained multimodal models can serve as important decision-making agents within the experimental loop. Our work also establishes optics as a physically grounded testbed for visual reasoning and active perception in scientific agents.

Running large language model (LLM) agents in production gets expensive fast. A frontier model (the largest, most capable tier) is accurate but can cost 25 times what a small model costs per token, and the gap compounds once a workflow chains several calls together. Planner-as-Router (PaR) attacks this from a different angle. Instead of leaving model-tier selection to some component downstream, it folds the choice into planning itself. As the planner breaks a query into subtasks, it also assigns each one a model size tier (small, mid, or frontier, ordered by capability and price), so the dependencies between subtasks are visible before any specialist runs. Unlike per-call routers such as cascade routing, which look at one node at a time, PaR sees the whole workflow up front and needs no separate router model or training data. We evaluate PaR with EntBench, a benchmark of 54 enterprise agentic tasks across seven classes, graded by actually running the generated Structured Query Language (SQL) and MongoDB queries against live databases. Over 1,157 evaluations spanning eight routers and three seeds, PaR stays on the observed cost-accuracy frontier. It matches a sink-frontier heuristic (frontier model on terminal nodes only) in accuracy at comparable cost and a faithful FrugalGPT cascade at lower cost, and cuts cost 44% against all-frontier routing while giving up 2.9 points of accuracy. Several accuracy gaps fall inside the plus-or-minus six-point confidence interval of a 54-task study, so we frame PaR's advantage as frontier position rather than a clean accuracy win. We also report a preliminary observation, not a validated result: a small pilot hints that cheap routing may carry a hidden compounding penalty on compositional workflows, which we frame as a hypothesis for future measurement. PaR, EntBench, and all evaluation code are open source.

Browser-use agents often carry information in their context as they move between websites. While it may be necessary for task completion, it also creates a privacy risk, especially when the information contains a private fact regarding the user. For example, an agent may learn a user's affiliation after reading a membership record. If it later selects a registration option specific to that affiliation on another website instead of a general option, the information gets leaked. In this work, we define and study behavioural side-channel leakage in browser-use agents, where an agent's actions inadvertently reveal private information (secret) retained from a prior website, despite an explicit instruction not to disclose it. We introduce AgentTell, a benchmark of 20 scenarios and 100 tasks in which an agent acquires a secret on one website and then completes a task on another website that offers secret-specific actions alongside a general action that reveals nothing. Our evaluation across 9,760 sessions on six backbones shows that agents carrying a secret reveal it through their actions in 61.1% of sessions. Even when agents explicitly state in memory that the secret must not be shared, they still reveal it in 56.7% of those sessions. Moreover, in 34.5% of leaking sessions, their final responses falsely assure users that the secret was not disclosed. These findings show that agents often fail to recognize side-channel leakage as a privacy risk.

The long-run statistics of chaotic dissipative partial differential equations are described by invariant measures, and when the equation carries a parameter these measures form a family whose members can differ in kind, from steady states to sustained spatiotemporal chaos. We ask whether a single generative model can cover such a family. We train a single conditional diffusion model once, on states of the Kuramoto-Sivashinsky equation at 160 values of its hyperviscosity, and evaluate it against held-out data at interleaved parameter values not seen in training, using a classifier two-sample test supported by spectral, geometric and tail statistics. Over most of the family, agreement improves as the dynamics become more chaotic. We identify a structural cause: the weakly chaotic members have invariant measures concentrated on low-dimensional sets, which a model with full support output cannot represent, and the target approaches full dimensionality as chaos increases. Restricting training to the strongly chaotic members reduces the classifier's excess over chance on those members by roughly 40%, and this restricted conditional model matches or exceeds models trained at single parameter values. On those members, generated samples reproduce energy spectra to within a resampling floor, with a small systematic deficit at the largest scales, and match the energy balance of held-out data.

Recent progress in frontier models has renewed interest in large-scale reinforcement learning (RL), but the cost of generating large-model rollouts makes even testing RL recipes expensive. We ask whether reasoning improvements learned by a small, weak model can benefit a larger, stronger model without using the strong model's rollouts during training. We introduce Allspark, a training and inference framework for weak-to-strong transfer through alternating chains of thought. A weak teacher is trained alongside a frozen copy of the same model; the two alternate reasoning segments, and the frozen model produces the final answer. At inference time, a stronger student replaces the frozen training partner, while both models remain fixed. Because they communicate through text, the teacher can steer students from different model families and with different tokenizers. We study Allspark at two scales: controlled Qwen experiments across math and reasoning, and larger-scale Inkling experiments on ARC-AGI-2. The Inkling experiments show accuracy gains in within-family and cross-family settings, including transfer to Kimi and Nemotron, with benefits that vary across inference settings. These findings motivate reusing a trained weak teacher across strong students and examining the resulting accuracy–token tradeoff.

When safety-critical systems are developed from a known internal specification, their correctness can be established by model checking. In the frequent case where such a specification is unknown or inaccessible, runtime verification presents an attractive alternative, e.g., to ascertain that autonomous and agentic systems as well as business processes satisfy desirable properties and/or comply with safety requirements. In this paper we study anticipatory monitoring, an advanced form of runtime verification, where the monitoring state is determined by both the trace prefix seen so far, and all its possible finite-length, future continuations. We focus on monitoring linear-time properties that may involve arithmetic constraints. Automata-based approaches, the de-facto standard in this setting, are notorious for their computational complexity. We propose an alternative approach based on progression and LTLf satisfiability checking, for both propositional and arithmetic settings. We experimentally compare the automata- and progression-based approaches, and a third method that combines the two. Our experiments suggest that the progression-based approach often succeeds in producing a verdict when the automata constructions do not terminate, especially for the arithmetic setting. For the propositional setting, the combined technique provides a good tradeoff.

Under tight feature and compute budgets, classical threat detection pipelines often degrade on near-decision-boundary events. Small quantum processors are now available, but existing work inadequately shows whether quantum components improve end-to-end threat detection under the above resource constrained conditions. This paper tries to address this gap with a hybrid architecture that uses a compact multilayer perceptron layer to compress the information in data and then routes the processed features to a few qubit quantum heads implemented in quantum support vector machine (QSVM) and variational quantum circuit (VQC) models. On a simulation platform, we benchmark these hybrid models against classical models with comparable parameter budgets on two representative cybersecurity tasks, network intrusion detection on NSL-KDD dataset and spam filtering on Ling-Spam dataset. To validate their precision on real quantum hardware, we deploy the best 4-qubit QSVM model on an IBM Quantum device with noise-aware execution, evaluated on a smaller sub-dataset. In the results, shallow quantum heads consistently match, and on difficult near-boundary cases modestly reduce missed attacks and false alarms compared to classical models using the same features. Hardware validation results track the simulation behavior closely enough that the remaining gap is dominated by device noise rather than model design. Furthermore, we conduct adversarial attacks to test the robustness of one QSVM model. Taken together, the study shows that even on small, noisy devices, carefully engineered quantum components may function as competitive, budget-aware components in practical cyber threat detection applications.

Neural theorem provers must both propose tactics and decide which valid successor states to explore. We study whether one-step Lean transitions provide a self-supervised signal for branch ordering. A JEPA-style model predicts latent successor representations and scores only kernel-validated, nonterminal successors generated by a fixed pretrained ByT5 proposer. JEPA achieves higher Top-1 than matched InfoNCE on a same-theorem ranking diagnostic(50.18% versus 31.55%), but averages 282.3 of 987 solved theorems across three seeds versus 308 for proposer ordering, while requiring more tactic checks. In this setting, accurate one-step transition ranking is therefore insufficient as a long-horizon search value. The controlled evaluation separates representation from proposal quality and treats kernel-checked proof completion as the primary endpoint.

Recent work has identified a human brain network specialized for abstract formal reasoning (Kean et al., 2025). Does the same hold true in language models? To answer this question, we introduce the minimal viable subspace (MVS) method, which searches for the lowest-rank activation subspace at a layer that preserves task performance when everything outside that subspace is ablated. Using MVS, we demonstrate low-rank subspaces supporting logical inference on Gemma and Qwen models. Furthermore, these subspaces exhibit a clear dissociation from model capacities on other tasks, such that retaining these late logic subspaces preserves inference while impairing factual knowledge, working memory, cognitive control, and arithmetic. Conversely, ablating them reduces logical inference accuracy to chance while largely sparing these other capacities. Our results suggest a functionally localizable core machinery for logic akin to that in the human brain.

Training language models with ternary weights is commonly judged by loss and downstream accuracy, which record only a modest cost relative to full precision. We show that these metrics can conceal a much larger failure. We instead measure knowledge capacity, the factual bits stored per parameter, on synthetic biographies with known information content. We train GPT-2-style models from scratch with 2.5M to 50M parameters at five precisions. Under the standard cosine schedule, ternary models retain as little as 6% of an identically trained fp16 model's capacity. The deficit widens with model size while perplexity rises by only 1.4 to 1.6 times. Measured throughout training, these models first acquire capacity and then lose most of it. We identify this knowledge collapse as a learning-rate dwell instability. Held near $1$–$2\times10^{-4}$ with no decay, a pre-collapse model collapses within a few hundred exposures, and returning to a safe rate does not restore capacity. We then locate the collapse in the output head. It happens at a value the model can never predict. The weights there grow unchecked, while every other prediction the model makes is unchanged. We find that making the value predictable removes the collapse, regardless of which attribute carries it. A warmup-stable-decay schedule with a 10% cooldown increases ternary capacity by 2.7 times at 25M and 4.3 times at 50M. Gains are larger at lower precision. Cutting the output head's learning rate prevents failure when training from scratch. The post-training quantization methods we tested recover no measurable capacity below 4 bits. Our findings suggest judging low-precision training by retained capacity during training rather than final loss.

Diffusion language models (dLLMs) predict masked positions in arbitrary order, but their exact constrained decoders still encode constraints as sequential languages, whose state must track every unresolved dependency between positions. For relational constraints this encoding grows exponentially: for same-order copy, every finite automaton needs $4^k$ states, deterministic or nondeterministic, and every context-free grammar has size $2^{\Omega(k)}$, while the factor graph of the same relation has size $O(k)$ and a 16-entry peak table. We introduce FactorDLM, a training-free decoder that represents finite-domain relations as a factor graph and, at each denoising step, conditions the model's mean-field prediction on that graph exactly by variable elimination. Decoding cost then grows exponentially with the induced width of the constraint graph, which replaces automaton size as the governing parameter. Because a finite automaton is a chain-shaped factor graph, one compiler enforces syntax and nonlocal relations together: on JSON records with cross-field references, a schema automaton alone leaves references dangling, relational factors alone produce malformed JSON, and the combined plan is valid on both counts, including on records of variable length. Across nine relational benchmarks and three backbones, every output satisfies every declared constraint at 0.4-6.9% projection overhead, where unconstrained decoding is 0-79% valid, and compiled projection answers repeated queries 13.6x faster than CP-SAT with eight parallel workers. Because model-free rules solve three of five standard benchmarks, we construct benchmarks with exact chance and fixed-template floors, on which selecting among exact constrained samples beats greedy projection. Which encoding is cheaper, sequential state or direct factors, depends on the constraint and is computable before decoding begins.

Hyperbolic space can embed intrinsic hierarchies in data with low distortion due to the exponential growth of volume with distance from the origin. However, current Lorentz transformer blocks are formulated in ambient coordinates, where numerical errors increase at large radii due to instability in the Lorentzian inner product, leading many models to limit the radius to avoid this issue. This prevents us from utilizing the regions of hyperbolic space that motivate the geometry. To address this, we revisit the components of the transformer block in polar coordinates, where hyperbolic operations such as distance calculation and attention centroids can be computed without the numerical cancellation in their ambient-coordinate formulations. Specifically, we propose a polar fully connected layer that separately maps an embedding's direction and radius, allowing the radius of a feature to be learned rather than determined by the norm of a linear map. We additionally introduce horospherical shifts as relative positional encodings whose query-key distances grow logarithmically with the token gap. Finally, we reformulate the residual connection as average radius Lorentz boosts. Combining these components, we develop a fully hyperbolic transformer that substantially improves performance over Euclidean and hyperbolic baselines on standard vision tasks. We further evaluate our model on ImageNet and demonstrate its ability to generalize to other datasets using pre-trained weights, similar to Euclidean counterparts.

Pretrained tabular foundation models process every dataset at a fixed depth, with inference costs growing with dataset size. To address this, we present the first study of depth-adaptive early-exit for pretrained outlier detection models. While early-exit is typically motivated by efficiency, we uncover a surprising benefit: exiting at the optimal intermediate layer can also improve detection performance on diverse real-world benchmarks by 4.7-7.3% on average, consistent across three distinct foundation models. First, we investigate the factors driving these gains, and identify a key mechanism: context pollution, i.e., the presence of outliers among in-context samples. Our analysis reveals that nearby in-context samples exert increasing influence on query predictions at greater depths, consistent with a retrieval-based view of these models. In effect, early-exit alleviates the adverse effects of retrieving accurate-yet-polluted neighbors, with gains of 13-21% when context pollution matches the natural outlier rate. Motivated by these findings, we pretrain a plug-in router to select a dataset-specific exit layer, using query outlier labels as privileged information available only during router training. The router operates post hoc, leaving the base model parameters and prediction head unchanged. Experiments on three large real-world benchmarks show that, on clean context, the router recovers up to 45% of the oracle gain with up to 1.8x speedup across three pretrained backbones, with larger gains as context pollution increases.

Sampling-based motion planners guided by diffusion models produce high-quality trajectories in a single run, yet the stochastic diversity available at inference time is left largely unexploited. We present two inference-time diversification strategies for a fixed, pretrained DiTree model, combined via H-Graph hybridization, and evaluate them on a holonomic AntMaze robot across 15 maze scenarios. The first, factorial diversity, sweeps the random seed and Diffusion Goal Bias (DGB) parameter, the second, refinement-only diversity, sweeps the diffusion refinement strength (RS) that controls how much an RRT-generated trajectory is edited. Because a single-run baseline only partially succeeds, we additionally compare H-Graph results with pool-based statistics. H-Graph improves the mean pool length of the factorial and refinement-only diversities by 18.8% and 14.5%, respectively. In addition, it also improves the best individual candidate's lengths by 9.7% and 6.8%, respectively. And last, compared with the successful baseline's trajectory length, it improves the results by 18.2% and 19.9%, respectively. These results show that inference-time parameter variation is a reliable, training-free source of path diversity, and that H-Graph hybridization reliably converts this diversity into shorter, higher quality trajectories.

Open-domain question answering (ODQA) over multimodal document corpora requires linking evidence scattered across text, tables, and figures. Existing systems often store these sources separately or retrieve only coarse pages, which weakens global evidence linking. We present PILAR, a page-grounded unified evidence representation instantiated as an entity-linked assertion graph. PILAR maps sentence-, table-, and figure-derived facts into a common assertion space and uses the graph as a controlled linking layer over robust page retrieval. In a shared-reader evaluation with four agent frameworks, fourteen retrieval backends, and two benchmarks, PILAR achieves the best end-to-end EM/ANLS. Gains are largest on compositional, cross-document, and multimodal questions, with a single-shot improvement of +1.6 EM over flat retrieval, rising to +2.9 on compositional and +5.9 on 3-hop questions. Ablations show that current gains are driven mainly by the text-instantiated slice of the framework, while visual assertions help only after locality-aware filtering. We therefore position PILAR as a unified evidence representation for multimodal ODQA rather than a standalone visual-reasoning module.

Language models are increasingly evaluated as vulnerability detectors, and scores reported for similar models differ widely between papers. We measured how much of that difference evaluation protocol accounts for, with model outputs held fixed. In a paired test, a model must flag a vulnerable function and clear its version after a fixing commit. Three choices that published evaluations make differently were varied one at a time: metric, verdict extraction and output budget. Seven frontier and large open models were evaluated on five released pair benchmarks and a set pooled for this work under one protocol, and 61 open models of 1.5B to 36B parameters on the pooled set. Function-level F1 follows how often a model flags both functions of a pair (Spearman $+0.86$ over 42 combinations) and is nearly unrelated to pair-level correctness ($+0.16$). On the pair score, extraction changes a model's number by $+0.001$ at the median and budget by $+0.02$ with an interval through zero, whereas the model changes a benchmark's number by up to 0.18 and the benchmark a model's by up to 0.16; a function-level score therefore measures flag rate more than model. For 37 of 68 models the difference between correct and reversed pairs is within its 95% interval of zero, the value for a null model that flags each function at a fixed rate, while both-flagged and both-cleared rates exceed that null by 0.055 on median, and for 64 of 68 both functions of a pair receive one answer more often than independence predicts: verdicts are determined by the text common to both functions. On length-matched pairs a linear probe on activations separates 0.78 by within-pair ranking, against 0.64 for a tf-idf baseline and 0.5 for length; the generated verdict is near chance for three of six models and at 0.55 to 0.57 for the other three, and a prompted logit is at chance for all six.

Frozen EEG representations retain acquisition signatures as well as neural activity. Source predictability alone does not identify what should be removed: it can reflect measurement effects or genuine biological and population differences, which should not be erased. We propose Measurement-Gated Provenance Attenuation (MGPA), built on one principle: measurement evidence determines where correction may act, and preserved information determines what it should aim for. Paired measurement contrasts define a gate outside which nothing changes; inside it, the source score is moved to the value the preserved coordinates already predict: for a fixed affine score, this keeps the same information as any target set by those coordinates and needs the least expected squared movement. Closed-form and critic-guided iterative constructions apply it without source identity or encoder retraining. Three studies test the principle at increasing distance from its assumptions. Under controlled reference changes, where the source-task association is known, MGPA brings source to near chance with task performance unchanged, whereas erasing what predicts source (LEACE) lowers frozen-task AUROC from .753 to .656 while barely touching source; ablations attribute the attenuation to the gate's directions and 2.7x less movement to the conditional target. Across recordings from different devices and electrodes, iterative correction lowers source accessibility while preserving or improving task performance. Finally, one iterative map selected on one task and reused unchanged on existing heads for two others raises their worst-association AUROC (lowest over device-label shifts) by .057 and .019 over LEACE, at a cost to those heads while the training association holds. A reusable correction shows its value in how an existing predictor behaves once acquisition cues stop being reliable, not only in what a probe can read.

Gradient-based data selection methods such as LESS score each candidate by the alignment between its gradient and a target validation gradient, and recomputing per-example gradient features at every new checkpoint dominates their cost. Across three selection seeds, two model families, two candidate pools, and two target tasks, features cached at a post-warmup checkpoint and paired with fresh validation gradients preserve the ranking 40 optimizer steps later with Spearman correlation from 0.952 to 0.991, while the top-10% subset they induce misses 10 to 22% of the examples that full recomputation selects. We therefore propose Cached Diverse Influence Selection (CDIS), which recomputes gradient features for the top-ranked fraction $p$ of candidates under the stale scores, fits an affine calibration on the recomputed examples, and selects the final subset under source and length quotas. A refresh fraction at or above the selection fraction recovers the exact top-$k$ subset whenever the calibrated stale scores have bounded error, and the budget curves follow this rule: at $p=0.3$ the recovered top-$k$ subset coincides with full recomputation in every setting, allocating the same budget per stratum recovers the stratified subset at 0.92 to 1.00, and gradient-stage wall-clock drops 3.5 to 3.6 times. Iterating the cache over four checkpoints keeps top-$k$ overlap at 0.98 or higher at 1.9 gradient features per example against 4 for full recomputation, and stale-to-recomputed agreement on the refreshed examples provides a free check for unsafe reuse. Downstream, unconstrained top-$k$ selection collapses to a single data source and scores 13 points below random selection on GSM8K. CDIS scores 12 points above random selection with paired confidence intervals that exclude zero and trails full recomputation by 4.3 points, one training-run standard deviation, at 3.4 times lower selection cost.

A clear link is established between conventional numerical methods and neural network approximations for solving one-dimensional scalar conservation laws. The focus is on the construction of an appropriate flux term in the case of convex flux functions for improving the classical schemes. The first neural network developed here is able to rediscover Godunov's method, while the second one emulates the behavior of a second-order slope-limiter function. In this way, by merging them, second-order reconstruction-based schemes can be developed. The networks presented here employ a minimal number of parameters, significantly reducing the complexity compared to previous approaches. These networks can also be linked consecutively to get a deep one corresponding to multiple time steps. Training them with an appropriate loss leads to stable schemes, improving even the classical methods without increasing their complexity.

Large language models (LLMs) can learn reusable textual skills from execution feedback without updating their parameters, but effectively deciding how to revise these skills remains a key challenge. Existing methods typically rely on a fixed revision operator, a search strategy and the revision forms applied under it. However, we observe that no single revision operator consistently performs best across tasks, and repeatedly applying an unsuitable operator can limit further improvement. We propose StraTune (strategy-guided skill tuning), which lets a frozen optimizer LLM choose the revision operator at every round from the optimization state, which is defined as the current execution feedback together with the recorded outcomes of earlier strategies and forms. Candidate skills from every revision operator pass one candidate evaluation, which screens for gains and regressions on a small sample set and validates them on a larger one, and every outcome is written back to the optimization state for later choices. Across four benchmarks and two LLM settings, StraTune outperforms all five baselines in most settings. Ablations attribute the gains to the adaptive choice of the revision operator, since fixed, random, scheduled, and bandit strategy choices all score lower, and skills learned with a small target LLM also improve a stronger one. Code and learned skills are available at https://github.com/seai-lab/StraTune.

We study whether model upgrades improve probability estimates for prediction-market questions. Our Resolved Market Forecasting (RMF) benchmark contains 3,000 resolved binary questions across nine domains, on which we evaluate six Claude and Qwen model variants using a question-only, zero-shot protocol. We assess Brier scores relative to an empirical base-rate predictor, examine their Murphy decomposition, and compare models through paired differences, with results stratified by event category and timing relative to training cutoffs. In the reported post-cutoff stratum, the four Claude models achieve Brier scores of 0.183-0.192, improving on the base-rate reference by 0.024-0.033. Qwen 32B does not significantly outperform that reference, although its paired Brier is 0.024 lower than that of the 7B checkpoint. The evaluated Claude version and tier upgrades yield no significant improvement. Within-model differences across event categories exceed the observed differences among Claude variants. These results show why forecasting scores should be interpreted alongside simple probability baselines and question composition: under this protocol, newer versions or higher model tiers do not consistently produce more accurate probabilities.

We present VSA2, a frontier trainable sparse attention for video DiTs. VSA2 includes a variety of new architectural features and training procedures that we apply across all stages of the DiT development cycle, including pretraining, RL, and inference, to produce a DiT with comparable or better quality than a full attention counterpart. Architecturally, VSA2 introduces a fine-grained router that improves the precision of identifying critical tokens and supports dynamic computation by allowing each query to attend to a variable number of key-value pairs. In training, we identify a Hard-to-Easy Curriculum, where models trained under high sparsity and later evaluated with lower sparsity during inference not only generalize effectively, but also outperform models trained with full attention in motion quality. VSA2 is also flexible: it can replace full attention during the middle of progressive low-to-high resolution pretraining, rebasing early-stage full-attention checkpoints. Experiments show that VSA2 reduces attention computation by half over VSA with lower loss. On 720p videos, it accelerates attention by 8.9x and end-to-end generation by 4.62x compared to the FlashAttention-3 baseline, while achieving comparable or better video quality.