Papers

Skillful prediction of biomass-burning activity several days in advance is important for air-quality forecasting and aerosol prediction. Two operational constraints motivate this work. First, the GBBEPx satellite fire radiative power (FRP) product used to initialize NOAA's GEFS-Aerosols is available with about a 1.5-day latency, so each forecast cycle relies on the most recently available, but already outdated, fire observations. Second, these fire inputs are then held fixed throughout the subsequent 5-day operational forecast, or 7 days in the GSL experimental system, effectively assuming no evolution in fire activity. We develop a data-driven model that predicts global FRP one to seven days ahead from the most recent available observations. The model adapts a spatiotemporal graph neural network using reanalysis meteorology, land-cover and vegetation information, recent fire history, and GBBEPx FRP as the training target. It is trained on 2020-2022 data and evaluated for 2023-2024. The model reproduces the global seasonal cycle and substantially outperforms persistence. At 0.1$^\circ$ resolution, mean squared error is reduced by 32% at one-day lead and 43% at seven days in 2023, and by 24% and 40% in 2024. At 1$^\circ$ resolution, the critical success index ranges from 0.32 to 0.60. Detection skill declines only modestly with lead time, whereas intensity skill degrades more rapidly. Large fires are detected reliably, but their radiative power is systematically underestimated. These results demonstrate useful predictability of fire activity several days ahead and identify intensity calibration and small-fire placement as the main remaining challenges before predicted FRP can support operational aerosol forecasts.

Multimodal large language models learn mostly from paired image-text data or annotated video, and raw web video is rarely used to further train an existing language model. We study whether raw video, with no captions and no text loss, can serve as mid-training data for a pretrained language model. Frames are encoded into continuous visual tokens, and the language model learns to predict the next visual token. We mid-train Qwen3-1.7B on raw clips from YT-Temporal-1B and then apply the same image-text instruction tuning to it and to the model without mid-training, so that the two differ only in mid-training. The mid-trained model scores 2.9 points higher on average across four video benchmarks and 5.1 points higher across ten image benchmarks, spanning perception, document, and chart tasks. Text performance is preserved even though mid-training includes no text, with an average of 48.9 across 14 text benchmarks compared with 48.0 for the model without mid-training. Analyses across training show that the image and video gains emerge within 30% of training and plateau thereafter, varying by less than 0.5 points. Predicting captions fails to outperform next-visual-token prediction, demonstrating that video mid-training can remain purely self-supervised without the computational overhead or labeling noise of automated captioning.

Long chain-of-thought decoding is memory-bound: the KV cache of the generated tokens grows linearly with the test-time compute being spent. Evicting cache entries does not fix this, because a model never trained on a cache with entries removed generates longer, often up to the length limit. We instead train the model to compress its own output. AbridgeKV compresses each completed block of generated text into the KV entries of a few learned pause tokens. These are extra input tokens with trained embeddings that produce no output. Once the pause tokens have encoded the block, AbridgeKV frees the block's original cache; subsequent tokens attend to the pause tokens' KV entries in its place. Three novel choices make this work: the pause tokens' KV entries take positions inside the block they replace, the cache budget is allocated across layers according to the measured value of each layer's entries, and compression operates on relatively large blocks. On 446 competition problems from AIME, HMMT, AMC, BRUMO, CMIMC and SMT with 32K-token generations, keeping a quarter of the generated cache matches an identically trained full-KV model in accuracy and output length; the strongest published evictor at a matched budget writes far longer, and the closest trained compressor, retrained on the same data, loses 10 points. Served in vLLM, the cache is 3.5x smaller per sequence and a saturated H100 finishes a problem 3.2x faster than with full KV.

LLM program-search agents are usually judged by the score that drives their search. We study LLM-driven search over Python exercise policies for Bermudan options, an optimal stopping problem where a trusted evaluator owns the simulation and a dual upper bound~$U$. For any policy, $0 \le V^\star - L_\pi \le U - L_\pi$; we report Monte Carlo estimates of both bounds with their sampling error. We audit 24 generation attempts from two pipelines, a six-pair feedback-representation ablation, and a holdout evaluated after selection was frozen. No selected generated program improves on Ridge LSMC consistently or by more than Monte Carlo uncertainty; holdout gap scores lie within $0.002$ of it. On six one-dimensional puts, a high-precision binomial reference places Ridge within Monte Carlo noise of optimal and attributes 99% of the estimated gap to dual-bound looseness, so this dual audits validity but cannot certify progress at the search's scale. Structured memory showed no benefit over score-only memory in this ablation. Whether search “worked” depends on reporting rules: retaining the initial reference, freezing selection before holdout, and counting failures.

The goal of circuit discovery is to decompose a monolithic neural network into tiny modular subnetworks which are interpretable in the sense that each does an identifiable subtask. We develop a practical method for composing arbitrarily many circuits into a single model. We propose a greedy, density-scaled composition algorithm that overcomes limitations of previous attempts to compose circuits by naive union, then generalize it to $N$ circuits via CircuitTree, a new data structure built on a $\Theta(N\log N)$ tree-compose schedule. We compose circuits pruned from HuBERT for classifying individual phonemes into a syllable classifier, a capability none of the 39 phoneme circuits has individually. Chaining CircuitTrees into a hierarchy-of-circuits pipeline, we transcribe open syllables at 64.8% accuracy, within 1.2 points of a 66.0% ceiling that shares no circuitry at all, with 99.5% of circuit components shared.

This position paper argues that the generative molecular dynamics (GenMD) community is not properly evaluating its progress: current evaluations let a model match what is reported while being arbitrarily wrong on what was claimed, which risks over-claiming advantage over classical baselines. This reflects a structural misalignment between common evaluation practices and the scientific tasks the field sets for itself. The reason is that the ground truth—exhaustive molecular dynamics on the systems of interest—is computationally costly, and this cost is what motivates the field. With ground truth inaccessible, evaluation drifts toward convenient proxies, and the resulting norms entrench easily. We identify misaligned evaluation patterns that persist as a community-level equilibrium, and propose concrete observables, reporting standards, and benchmark designs, both computational and experimental, that better reflect what GenMD aspires to deliver. Our aim is to give the field shared language and tools for evaluation that match its scientific ambitions before today's conventions calcify into norms.

Autoregressive video diffusion models maintain a growing attention key-value (KV) cache, making memory a key bottleneck for long video generation. Recently, clustering-based quantization has emerged as an effective method to compress video KV caches by grouping similar key and value vectors then quantizing residual errors. However, when later frames attend to the compressed history, small reconstruction errors can compound as the video grows. Beyond the within-layer similarity exploited by clustering, we find strong dependencies across cache layers in video models. This motivates Coral-KV - Calibrated Offline, Residual-coded, and Adapted Locally - a KV-cache compression framework that uses cache states in each layer to predict keys and values in subsequent layers, then quantizes the remaining prediction residuals. We construct compact block-diagonal-plus-low-rank (block-DPLR) predictors: block-diagonal components model channel relationships within each attention head, while a global low-rank term models dependencies between heads. Coral-KV exposes a flexible fidelity-memory trade-off: it can reuse its predictors across videos or refit lightweight block-diagonal predictors to the current video for higher fidelity. Across extensive evaluations on both HY-WorldPlay and MAGI-1, our reusable configuration improves PSNR over QVG-4 by 0.32 dB and 4.95 dB, respectively, while using 48% and 37% less cache memory. Against QVG-Pro, our video-adapted configuration improves MAGI-1 PSNR by 3.77 dB while using 33.7% less cache memory. Coral-KV also maintains higher reconstruction fidelity over extended generation and reduces temporal discontinuities while retaining comparable generation latency.

At the start of every session, LLM agents load a fixed context file, such as $\texttt{AGENTS.md}$. Each loaded token in the file is charged again in every later round of the session, and these files can degrade performance as they grow in size. However, in practice, human or automated curators usually grow these files by appending. We formulate context curation as a capacitated assortment problem. Instructions consume tokens under a finite attention capacity; adding an instruction never raises the compliance of the others, while retained instructions incur a per-session setup cost. We prove an upper bound on the optimal file size, regardless of the number of available candidate instructions, and that appending every instruction with positive standalone value can be arbitrarily worse in net value than selecting an optimal subset. A token budget also limits the loss when the token price is underestimated. We then examine what can be learned from past sessions and how this information can guide decisions to add or remove instructions. Feedback is inherently censored: the benefits and harms of loaded instructions are observable, whereas missing instructions generate feedback only when their absence causes harm. In this setting, we show that deleting instructions ignored by agents can inevitably remove helpful ones. We characterize how much evidence should be collected before adding an instruction. Besides, we bound regret when human reviewers can inspect only a limited number of edits per period. Empirical experiments further show that irrelevant rules drawn from real context files reduce language-model compliance.

Reconstructing spatially resolved plasma dynamics from few sensors is essential for diagnostics, reduced-order modelling and control, yet remains difficult because the sparse measurements incompletely constrain multiscale, regime-dependent degrees of freedom. The Shallow Recurrent Decoder (SHRED) partially addresses spatial sparsity by using measurement histories; however, its fully connected decoder provides no explicit mechanism for resolving spatial structure across scales or explicit parametric dependency. We introduce the Recurrent Multiscale Affine-modulated Inference Network (ReMAIN), which preserves SHRED's recurrent temporal encoding but replaces its decoder with a U-Net whose feature hierarchy is conditioned by the recurrent state through feature-wise linear modulation. The temporal representation supplies both a dense prior and scale-specific modulation throughout the U-Net. A parametric extension jointly embeds the operating condition and sensor history, enabling reconstruction to adapt as the governing dynamics change with operating regime. ReMAIN is first benchmarked against SHRED on six one-dimensional nonlinear PDEs representing diverse dynamics. Across all benchmarks, it reduces reconstruction errors on unseen trajectories and more faithfully resolves sharp transitions, localized extrema and fine-scale variations. The parameter-conditioned model is then demonstrated on a collisionless $E \times B$ plasma subject to perpendicular axial electric and radial magnetic fields, with the electric-field strength serving as the operating parameter. ReMAIN reconstructs the high-dimensional, multiscale plasma state and recovers its regime-dependent spatiotemporal dynamics at electric-field strengths withheld from training. Together, ReMAIN improves sparse-sensor full-state reconstruction and, through parameter conditioning, generalizes across plasma operating regimes.

The intermittent and variable nature of precipitation makes its accurate estimation over extended domains difficult, yet its spatiotemporal structure suggests that a low-rank representation may be possible. This work represents daily precipitation over the contiguous United States (CONUS) as spatiotemporal tensors and applies CANDECOMP/PARAFAC factorization, showing that preserving the native spatial and temporal modes yields more accurate reconstruction than factorizing independent daily fields or unfolded space–time matrices. Building on this finding, this work presents TMerge, a tensor-based framework that integrates satellite precipitation with sparse reference observations through shared low-rank spatial and temporal factors. TMerge was applied to correct the IMERG Final Run product with climate prediction center reference observations over CONUS. During 2019-2022, TMerge increased correlation from 0.53 to 0.85 and reduced root-mean-square error and mean absolute error by 48.2% and 29.3%, respectively. TMerge consistently outperformed linear bias correction, quantile mapping, and neural networks across seasons, precipitation-intensity regimes, and regions. Improvements were spatially coherent and largest in coastal regions where IMERG errors were greatest. These results demonstrate that low-rank tensor structure parsimoniously approximates the dominant spatiotemporal variability of precipitation and provides a practical mechanism for improving satellite estimates under limited reference observations over extended domains.

Neural operators enable fast PDE forecasting, but repeated predictions accumulate errors and fine-scale structures remain difficult to resolve. We introduce a frequency-decomposed finite-time flow-map neural operator (F$^3$NO) that leverages updated low-frequency features to guide nonlinear refinement of high-frequency information. Within each layer, this cross-scale conditioning connects global spectral processing with local detail refinement. The model directly predicts states at specified future times and adjusts the contributions of the two branches according to the prediction interval. For longer trajectories, it combines parallel predictions within short temporal segments with recursive propagation between segments. Experiments on five PDE benchmarks demonstrate improved forecasting accuracy over autoregressive and direct-prediction baselines. Ablations show that frequency-decomposed refinement can improve accuracy with fewer parameters, while the benefits of segmentation depend on spatial resolution and dynamical regime.

Latent reasoning models repeatedly update a latent state using the same recurrent block. As the recurrent depth increases, the sequence of latent states may converge to a compact subset of the state space without necessarily converging to a fixed point. Existing models typically predict by applying a prediction head to a single latent state. However, the latent state can continue to change even after many updates, potentially making predictions unstable across recurrent depths. To address this instability, we introduce invariant-measure reasoners (ImR), a framework that uses an invariant measure as a stable representation. This measure describes the long-run distribution of latent states on the compact subset and is invariant under updates by the recurrent block. ImR predicts from the expectation of the prediction head's output under this measure. We use ImR in two ways: fine-tuning only the prediction head of existing models and training models from scratch. Both approaches reduce prediction instability and improve accuracy in many settings on maze and Sudoku tasks. In some settings, models trained with ImR exhibit non-fixed-point behavior more frequently than existing models yet achieve high accuracy even with such behavior, unlike existing models. These results suggest that ImR can leverage otherwise destabilizing dynamics for latent reasoning.

Image retrieval methods often rely on a single global semantic descriptor extracted from an image, e.g., the [CLS] token in vision transformers. However, trying to squeeze all the semantic information of an image into a single descriptor can hurt downstream retrieval performance, especially for fine-grained retrieval tasks. In this work, we augment the semantic tokens in the newer visual transformers, the global [CLS] token and the four register tokens, with a carefully selected collection of spatial tokens, aiming to capture the spatial region representation that characterizes the contents captured in each of the semantic tokens. We leverage the DINOv2-reg model, which includes register tokens that emergently learn object and part-based representations. For each "cue" token ([CLS] and each register token), we find a "buddy" image patch token and extract an N x N patch region to produce a set of localized ROI tokens. Our approach automatically captures important regions of interest without any external bounding boxes or saliency modules, purely by matching semantic tokens with their spatial representation regions. Furthermore, we incorporate these tokens into a multi-vector retrieval framework inspired by ColBERT, enabling fine-grained matching via a per-token alignment mechanism while avoiding the large storage cost of keeping all patch embeddings. Through extensive experiments, we find that (1) register tokens encode useful fine-grained details that can complement the [CLS] token; (2) automatically pooled ROI tokens further improve fine-grained discrimination; and (3) multi-vector retrieval with a small set of tokens improves over a DINOv2-reg single-vector baseline while remaining tractable for large-scale search. The code is available at https://github.com/IdhcbIan/Augmenting_CLS_with_ROI_tokens.

Unified multimodal diffusion large language models (dLLMs) offer a single architecture for both image generation and multimodal understanding, but their iterative decoding requires tens to hundreds of forward passes. Existing few-step distillation methods largely focus on either image generation or text generation, making it unclear how to compress a fully discrete multimodal dLLM into a single efficient student while preserving both generation and understanding. We introduce Omni-Diffusion-Distill, a unified two-stage distillation framework that retains strong generation and understanding capabilities while substantially reducing the inference cost of a unified multimodal dLLM. Omni-Diffusion-Distill aligns the distillation of both generation and understanding, for both images and text, in the discrete token space. In the first stage, the student is trained to skip decoding steps by replaying cached teacher trajectories, and in the second stage the student is refined on intermediate states along its own rollouts. We further remedy two sources of degradation in unified distillation with a pairwise collision penalty that reduces repetition under parallel text decoding, and entropy-matched guidance that prevents entropy collapse caused by fitting the sharpened teacher distribution in image generation. Omni-Diffusion-Distill achieves state-of-the-art trade-offs between decoding efficiency and generation and understanding performance for multimodal dLLMs, reducing image generation from 128 to 8 decoding steps and multimodal understanding from 512 to 64, giving 18.2x and 21.2x wall-clock speedups. Under these budgets, it scores 0.828 on GenEval and 83.0 on DPG-Bench for text-to-image generation, while reaching GPT judge scores of 20.0 on MM-Vet and 57.2 on COCO captioning (twice the teacher's 28.4 at the same steps) for multimodal understanding.

Researchers are exploring effective one-step generative model continuously, and, Drifting Models (Deng et al., 2026), demonstrate great potential in one-step generation recently. There are works that reveal the connection between Diffusion & Flow Style Generative Models (DFSGMs) (Ho et al., 2020; Song et al., 2020a;b; Lipman et al., 2022; Liu et al., 2022) and Drifting Models (Li & Zhu, 2026; Lai et al., 2026; Turan et al., 2026). But no one has yet established a precise correspondence between the Drifting Model and the widely used distillation method- Distribution Matching Distillation (DMD/DMD2) (Yin et al., 2024b;a) to the best of our knowledge, even though their optimization objective formulas are virtually identical. In this paper, we prove that by converting the velocity-field / noise-field from the pre-trained DFSGMs into the attraction force field in Drifting Models and estimating the repulsion force field from the generative distribution, training the Drifting Model is naturally equivalent to the Distribution Matching Distillation. With this equivalent concept, we propose an improved method based on DMD from the Drifting Model's perspective- Multi-Bandwidth Distribution Matching Distillation (MBDMD).

Muon has motivated a growing family of optimizers that compose multiple matrix normalizations, but these methods remain fragmented and lack a unified perspective. We introduce chained linear minimization oracles (chained LMOs), which cast these methods as compositions of LMOs. Despite their empirical success, many chains fall outside the standard LMO framework and can diverge on smooth convex objectives. To explain why composition can nevertheless help, we turn to linear associative memory and show that chaining can improve over Muon under anisotropic embeddings. Empirically, we propose TensorChain, a novel optimizer within the framework that stacks compatible weight matrices across different layers and normalizes the 3d tensor across its axes. In Qwen3 0.6B and 1.7B pretraining, TensorChain outperforms all chained baselines in average token efficiency, with average token savings of 9.6% over Muon at matched validation loss.

Off-policy evaluation (OPE) estimates the value of a target policy from logged data, but limited behavior-policy coverage can force high-variance reweighting or reward-model extrapolation. Counterfactual annotations can add evidence about unobserved actions, yet practical sources, including domain experts and large language models (LLMs), may be costly, biased, or noisy. We study budgeted acquisition of such annotations for contextual-bandit OPE. Given source-specific costs and error profiles, we formulate an integer allocation problem over context-action pairs and annotation sources to minimize the component of estimator variance that depends on the annotation plan. We characterize when annotations are valuable through a first-annotation threshold and local annotation-value regimes. For the coupled multi-source problem, we develop a majorization-minimization algorithm with dynamic-programming subroutines that monotonically improves the objective. Experiments in synthetic clinical and LLM-annotated education bandits show that our allocation method reduces fixed-profile mean squared error (MSE) by 20.58% and 10.77%, respectively, relative to no annotation.

Assessing useful reuse in neural PDE solvers is challenging: final accuracy can reflect source learning and target-time computation. Our reuse contract separates solution accuracy, learning contribution, and numerical utility through paired state comparisons, matched target information and budgets, and cost accounting. A literature audit extracts 18 version-specific protocol records from 12 papers, documenting retained states, target-time resources, and reported controls. For a fixed linear system and residual tolerance, we construct two initial guesses with identical solution-error, energy-error, and residual norms, reaching the same solution with different conjugate-gradient (CG) iteration counts. Across 240 source-training trajectories, two linear PDE families, Fourier neural operators and convolutional networks, a fixed predictor's benefit reverses across correction algorithms. Among pairs with both relative prediction errors less than or equal to 5 percent on 64 in-distribution tasks (63 by 63 interior grids), reductions in all three norms accompany more CG iterations, at mean taskwise rates of 23.5 percent and 23.9 percent in two libraries. Work-based selection saves 2.50-3.33 CG iterations on held-out in-distribution tasks; matched adaptation demonstrates finite-budget pretraining value. Independent batches confirm a 0.73 percent complete online saving for one physics-trained Fourier neural operator against zero-initialized Poisson-preconditioned CG. Reuse requires matched state comparisons and downstream computational evidence.

Predicting thermally activated diffusion and defect evolution with event-driven models requires identifying atomic rearrangement mechanisms and their activation barriers. Discovering the associated saddle points is a major computational bottleneck: multiple rearrangements may originate from one metastable state, while costly local searches can fail or repeatedly converge to the same saddle. To address this challenge, we introduce ASPIRE (Atomistic Saddle-Point Inference with Refinement for Events), a framework that predicts a set of saddle candidates from a single initial atomic environment and refines them through Dimer searches on the original interatomic potential. The framework's equivariant set predictor, Ev-Quiformer, integrates (i) geometry-conditioned scalar-vector event slots for generating multiple saddle-point proposals and (ii) a decoder that maps each slot to a full atomic displacement field by combining atom, slot, and anchor-relative vectors with invariant coefficients. We also contribute two datasets: (i) BCCFE4VACAV-4000, comprising 4,000 four-vacancy body-centered cubic iron configurations and 65,450 reference events grouped by initial state for set supervision and post-refinement evaluation; and (ii) BCCFE-1TO4VAC, comprising 5,372 configurations with one to four vacancies each. Theoretically, we establish conditions for proposal equivariance. Experimentally, ASPIRE achieves 77.20% reference-event coverage on this benchmark, compared with 75.73% for a conventional Dimer baseline, while requiring approximately half as many Dimer force evaluations. In a timing evaluation on 50 configurations, ASPIRE reduces wall time per configuration from 478.8 s to 176.3 s under the stated hardware settings.

The Muon optimizer orthogonalizes each update matrix, setting all of its singular values to one, and has proven highly effective for training large language models. It remains unclear, however, whether this success comes from amplifying small singular directions that gradient descent neglects or from suppressing large, degenerate directions that disrupt training. We introduce Spectrally Targeted Muon, which orthogonalizes only the singular values above or below a threshold $\tau$, so that varying $\tau$ interpolates between normalized SGD and Muon. It isolates the relevant singular subspaces with projections computed by Newton-Schulz iteration on a shifted Gram matrix, so no SVD is needed. We evaluate these variants on the CIFAR-10 and NanoGPT speedruns, tracking the effective rank of gradient, update, and weight matrices and a new metric, the alignment of updates with the tangent space of the weight matrix's isospectral manifold. We find three things. First, the small singular values of the momentum are not noise. Orthogonalizing everything except the few largest singular values of each matrix nearly matches Muon while touching only a small fraction of the momentum, whereas orthogonalizing only the top falls well short even though it holds almost all of it. On language models every momentum singular value is far below one, so targeted orthogonalization can only amplify, and Muon wins by making directions that are too small to train on at their raw scale trainable. Second, shrinking the largest singular values is what keeps the parameter spectrum flat; this is cheap, and it is not what drives the loss. Third, AdamW differs from Muon mainly in how slowly it builds structure, which explains its slower start and why warmup helps AdamW but only hurts Muon.

Production AI systems deployed in high-stakes domains accumulate a governance liability that existing monitoring frameworks fail to detect: the progressive inability to explain individual decisions when regulators, auditors, or affected individuals demand accountability. We introduce TRACE (Transparency, Risk, Accountability, Compliance, and Explainability), a seven-instrument governance framework for measuring, tracking, and remediating Explainability Debt in production AI systems. The foundational instrument, the Explainability Debt Score (EDS), quantifies the proportion of production decisions falling below a governance-defined explainability confidence threshold at any point in time. Complementary instruments include DART (Debt Accumulation Rate Tracker for breach forecasting), SHIV (Scenario Health and Integrity Validator for daily governance), FDE (Feature Drift Evaluator for causal attribution), HVE (Human Validation Engine), AIDE (Audit Intervention Decision Engine), and ZERO (Zero Explainability Risk Optimiser for remediation). Through a twelve-month longitudinal case study of a production fraud detection system processing 50,000 daily financial transactions, achieving 98.46% accuracy and ROC-AUC of 0.9990, we demonstrate that an EDS of 0.23 on audit day was statistically predictable six months in advance using DART trajectory analysis (beta = 0.008/week, R-squared = 0.94, 95% CI: [0.006, 0.010]), and that 78% of Explainability Debt was concentrated in the highest-regulatory-risk decision category (transactions above $10,000), a risk asymmetry completely invisible to system-level metrics. TRACE provides the first quantitative operational architecture for EU AI Act Article 13 compliance in production AI deployment, establishing a new subdiscipline of explanation governance distinct from explanation generation.

Brain signal decoding is challenging because neural recordings are noisy and vary across individuals, while labeled data are often limited. Recent attention-based models on the symmetric positive definite (SPD) manifold have nevertheless achieved strong performance using covariance and connectivity representations, yet the contribution of learned token weighting remains unclear. We examine two representative architectures, MAtt (based on log-Euclidean geometry) and GBWAtt (based on generalized Bures–Wasserstein geometry), and find that their learned attention weights remain close to uniform after training. We relate this behavior to bounded similarity parameterizations that, under the original softmax scaling, limit attention-weight contrast. Moreover, replacing learned weights with uniform weights, throughout training and evaluation, has little effect on mean predictive performance while preserving each model's original aggregation geometry. Motivated by these findings, we introduce SPD-MetaFormer, an attention-free architecture built on uniformly weighted Fréchet aggregation under log-Euclidean geometry. Its backbone uses a geodesic residual to update a summary token and a shared spectral feedforward map to transform all tokens, followed by a learned weighted readout. Token states remain SPD-valued until tangent-space classification. Across three EEG benchmarks, SPD-MetaFormer achieves competitive results relative to published Euclidean and manifold baselines. Separate matched reproductions test learned versus uniform weighting within MAtt and GBWAtt. These results suggest that, in the short-sequence and limited-data regimes studied, carefully designed SPD architectures can provide a simpler and effective alternative to adaptive manifold attention.

Reward hacking, where a model exploits an evaluation signal without completing the intended task, threatens the reliability of deployed language model systems. Existing datasets use different labels, response formats, and metadata conventions, making detector results difficult to compare. We present RH-Detect, a benchmark that combines reward-hacking-relevant subsets from eleven public datasets, comprising 92,761 rows and six behavior categories, into a common schema. On 5,021 open-ended evaluation units, each comprising a task prompt and a free-form model continuation, including multi-turn tool-use trajectories, we evaluate six off-the-shelf language models from five families as reward hacking detectors without additional training. The best model achieves a pooled AUROC of 0.962, with accuracy above 93%. For the four strongest models, however, accuracy on the two multi-turn tool-use datasets, MALT and TRACE, is 10.7-15.9 percentage points lower than on the other sources at a common decision threshold, highlighting a key gap for deployment-time monitoring. We find that different input formats have different effects across models. Removing thinking raises Qwen3.5-4B AUROC from 0.779 to 0.849, but lowers Qwen Flash from 0.977 to 0.950. We also evaluate the benchmark as a training dataset for detectors. Holding out each source in turn, single-token SFT improves average AUROC on five of six held-out sources. A GRPO follow-up on that failure case yields a slight improvement in detection performance. Our results show that a single pooled score can conceal variation across data sources, detector inputs, and training procedures.

Large language models (LLMs) are entering wildfire management, where overstated evaluations can cost property and lives. How do they perform on wildfire tasks, with and without grounding? Bare means a model receives the task input alone. Grounded means it also receives one task-specific addition: for smoke detection, a smoke-free reference frame from the same camera. AI4Fire runs six core models bare and grounded on five wildfire tasks, zero-shot; a sweep adds 29 more. Our literature search on fire tasks found 138 works; none combines this roster, task coverage, and paired bare and grounded runs. We report three findings. (1) Grounding helped most where the addition carried the answer: a read-only SQL tool lifted every core model's database accuracy from at most 16 to at least 88 percent. (2) Simple rules were hard to beat: no core model outperformed repeating today's staffing count, and two open-weight models mostly copied the median of similar earlier fire-days, a worse forecast. (3) Public releases carry hazards: a fire-danger column separates the holdout perfectly, and 67 aerial fire frames carry smoldering or fire-free labels read from a clipped thermal maximum. We release prompts, responses, scores, code, and the survey record.

We present GHARP (Real-time Gaussian Head Animation from Large-scale Reconstruction Prior), a method that animates 3D human heads in real time from a few input images of a subject and a driving expression signal. We decouple the problem into an identity stage that builds a representation of the subject's geometry and appearance offline, and an animation stage that predicts expression-dependent residuals on top of it at runtime. This separation offers a favorable trade-off with respect to fidelity, quality and runtime: the identity stage can be expensive while the animation stage runs a lightweight network, optimized for mobile devices. Our method performs animation in a semantically structured latent space of a pretrained reconstruction model, where expression changes remain spatially contained, making residual prediction efficient. This reconstruction prior provides a consistent spatial layout, allowing fusion of multiple input views into a compact, fixed-size canonical Gaussian representation. While this two-stage design improves the runtime-quality trade-off, it still inherits a problem common to all expression-driven avatar methods: expression codes describe only the face and thus omit body pose and clothing position, making these regions underspecified in the input. The animation network faces an ill-posed mapping and resorts to averaging over conflicting body appearances, producing blur and temporal flicker. We address this with a body alignment network that learns to align the person's body in the target image with the input reference images, removing the ambiguity from the training signal. Our method achieves state-of-the-art quality on the Ava-256 benchmark while running up to 13x faster on an A100 GPU with 8x fewer Gaussians.

Reinforcement learning environments are now a primary lever for improving large language model (LLM) capabilities in post-training, yet most agentic benchmarks remain static: the world moves only when the agent acts, the reward is a terminal verdict, and the pass bar is set arbitrarily. We introduce StoreBench, a live-commerce environment in which an agent runs a mid-size online apparel store on a production-grade commerce backend, testing long-horizon planning and economic judgment under uncertainty. Customers order around the clock, suppliers reprice and fail, and market shocks arrive with partial or no warning. The agent acts through the same 29 merchant tools a human operator would use, under a windowed operation budget that makes simulated time a function of actions taken, so model latency cannot influence simulated time. Pass thresholds are calibrated against scripted anchor policies, the reward is hardened against a catalogue of reward hacks, and every episode replays identically given a sequence of actions. We evaluate seven frontier LLMs on 11 scenarios of 30 to 45 days and a full simulated year, over three world seeds at matched reasoning effort. No model matches the scripted smart-triage policy on average: the best, DeepSeek-V4-Pro, passes 49% of task-seed cells against the heuristic's 97%. Human experts working through the same tools and budgets outscore every model (mean composite 0.708 vs. 0.700). Over a full simulated year under the Claude Code harness, most models show dramatic performance improvement. In a GRPO post-training run, Qwen3.5-27B trained on only five disjoint tasks raises its mean composite on the held-out evaluation tasks from 0.136 to 0.373. We release five example training-split tasks, ten sample trajectories, and the scoring and verification tooling; the full environment and evaluation suite are withheld to keep the benchmark uncontaminated.

Actuator degradation turns quadruped locomotion into a coordination problem requiring joints to compensate for lost actuation. Prior work suggests that morphology-aware graph policies improve learning and generalization under body perturbations. We ask whether these benefits can be strengthened by explicitly modeling higher-order mechanical structure. We represent the Unitree Go1 as a cell complex with limb- and body-level rank-2 cells and apply Hodge-based message passing. Under degradation training, the node-edge-face Hodge actor achieves the highest return on unseen actuator degradations, with higher survival and lower velocity-tracking error. These results support higher-order morphology as a useful inductive bias for whole-body compensation under actuator degradation.

Spiking language models face a tradeoff between representing continuous semantic features over short temporal windows and retaining costly nonlinear attention operations. We introduce Spora, which jointly designs spike encodings and attention operators. Binary temporal weights let $T$ spikes represent compositional values with up to $T$ bits of capacity, compared with $O(\log_2 T)$ bits for spike-count readout. Unipolar Binary Spiking (UBS) uses thresholds and spike-triggered residual decay to produce non-negative integer codes; Bipolar Binary Spiking (BBS) separates sign and magnitude and learns a scale for signed activations. These representations support accumulation-and-shift dot products and integer-exponent mappings in attention. With four time steps, Spora achieves 76.6 average GLUE score and 44.1 CoLA MCC, improving over SpikeLM by 1.2 and 6.2 points, respectively. Extending BBS to six steps raises these scores to 78.2 and 47.4. Conditional-decay analysis, matched-budget activation-quantization comparisons, event-workload statistics, and fixed-point evaluation further characterize the connection between encoding fidelity and computational cost.

Offline goal-conditioned reinforcement learning (GCRL) has produced a diverse set of goal-reaching algorithms, yet no single algorithm performs best across environments, goals, and even different phases of the same task. Rather than deploying only the best-performing policy, we ask whether a set of frozen goal-conditioned policies can be used collectively as a portfolio, deciding at every state which policy should act. Choosing a policy at each state is not straightforward. The policies' own value functions cannot be compared directly: they may use different scales, and some policies have no value function. We need to judge each policy by the states it is likely to reach, even though we can execute only one policy at a time. We also need to avoid switching so often that control becomes unstable. To address these challenges, we introduce World-Model Policy Arbiter (WMPA), a test-time framework that, given a set of frozen policies as input, rolls out each frozen policy in a learned state-space world model, evaluates the imagined futures with a shared goal-conditioned value function, and executes the highest-scoring policy for a short commitment interval before the next round of arbitration (policy selection). WMPA assumes access to a bank of frozen goal-conditioned policies and requires neither policy retraining nor privileged task-specific knowledge. Under the official OGBench evaluation protocol on 18 state-based datasets spanning maze navigation as well as cube, scene, and puzzle manipulation, WMPA improves the macro-average success rate from the 44% achieved by the best policy selected per dataset to 58%, with statistically significant gains on 12 datasets. These gains include +33 percentage points on cube-double-play and +36 percentage points on scene-play.

In memetic symbolic regression, candidate structures are compared after coefficient calibration, so the calibration protocol itself contributes to evolutionary selection. Standard centralized calibration evaluates each structure at its pooled-sample optimum, ignoring how stable this calibration is under covariate shifts, and can thus favor structures whose fit relies on sample-specific coefficients. We propose Dirichlet-Sinkhorn Constant Averaging (DSCA), a calibration strategy that partitions the optimization data into equally sized subsets with different covariate distributions, calibrates each candidate independently on every partition, and evaluates it at the mean of the resulting parameters. We show that the excess loss of DSCA relative to centralized calibration vanishes at the population level for correctly specified, identifiable expressions, whereas under misspecification it persists when partition-specific calibrations do not aggregate to the pooled optimum. On synthetic benchmarks and ten real-world datasets, DSCA improves functional recovery and the accuracy-complexity trade-off over centralized Broyden-Fletcher-Goldfarb-Shanno and Levenberg-Marquardt calibration, under selection by negative log-likelihood and by the Akaike and Bayesian information criteria. Mechanism analyses associate the DSCA excess loss with the generalization gap and show that the effect is not reproduced by repeated centralized fitting. These results indicate that controlled heterogeneous calibration provides a complementary source of selection pressure in symbolic-regression search.