Papers

Parkinson's disease (PD) is the second most common neurodegenerative disorder. Typical machine learning screening methods require PD labels, but the available data is limited by privacy concerns and the need for expert annotation. We propose a label-free face-plus-voice PD screen built entirely on frozen pretrained encoders–a face-expression Vision Transformer and HuBERT–in which no PD label touches any fit; the reference is training controls only. The voice modality uses a synthetic-dysarthria contrastive activation addition (CAA) direction built from time-stretch and breathy degradation of healthy speech; the face modality uses a k-nearest-neighbor anomaly score to the control embedding cluster. We introduce the alignment principle, a post-hoc analysis showing that a synthetic-degradation CAA detector works when the cosine similarity between the synthetic and real disease directions exceeds zero. Measured on the YouTubePD benchmark, this cosine is +0.37 for voice (CAA works, AUROC 0.765) and -0.48 for face (CAA fails; anomaly succeeds, AUROC 0.751). Equal-weight late fusion reaches AUROC 0.802 (95% CI [0.70,0.89]) with NPV 0.95, supporting a rule-out triage interpretation. An overfitting audit shows the voice detector transfers cleanly, while the face-side–and thus fused–AUROC is potentially optimistic pending external validation.

Many post-training quantization (PTQ) methods use layer-wise reconstruction, second-order proxy objectives, or activation-aware transformations to reduce quantization-induced error. Whether that error signal predicts the downstream functional impact of quantizing an individual attention projection has not been directly characterized. We sweep nine open-weight language models (1.3B–8B parameters; OPT, GPT-J, LLaMA-1/2/3, Mistral, Qwen 2.5) and quantize one attention projection at a time under round-to-nearest (RTN) and, for seven models, GPTQ at 3 and 4 bits, recording reconstruction error, perplexity change, and per-projection activation-weighted quantization error for 3,808 distinct measurements. We find: (1) within a given component type (Q, K, V, or O), reconstruction error explains less than 10% of the variance in perplexity sensitivity in 27 of 36 cases under RTN, with median R^2 = 0.044; (2) both component type and layer identity explain more variance than reconstruction error in all 9 models, with layer identity the strongest predictor in 7 of 9 models and component type strongest in the remaining 2; (3) value (V) projections are the most commonly dominant component, accounting for 38–51% of total positive Delta PPL in seven of nine models; (4) the dominant component is broadly preserved between RTN and GPTQ (5 of 7 cases); and (5) activation-weighted quantization error is a moderately better within-component predictor than reconstruction error for V projections specifically (median R^2 of 0.20 vs. 0.06). These findings indicate that relative weight reconstruction error alone is insufficient for sensitivity-aware bit allocation, and that V projections merit dedicated consideration in mixed-precision schemes.

Remote Sensing Change Captioning (RSCC), which aims to generate accurate and detailed linguistic descriptions of ground object variations from bi-temporal remote sensing images, is a critical and challenging task in intelligent remote sensing interpretation. The mainstream autoregressive training paradigm faces severe exposure bias and train-test distribution mismatch, resulting in cumulative generation errors. They tend to produce conservative and template-fixed captions while ignoring subtle scene change details. To address these challenges, this paper proposes a novel multi-granularity reward reinforcement learning paradigm, termed MGRL-RSCC. Specifically, we first leverage a CNN and hierarchical self-attention module to extract and enhance visual features from bi-temporal remote sensing images. A Transformer decoder is then utilized to complete visual-to-linguistic translation. Different from existing methods, we design a dual-decoding strategy and a two-stage joint optimization scheme, which combines token-level supervised learning via greedy decoding and multi-granularity reward-driven self-critical reinforcement learning via sampling decoding. We further construct three complementary reward functions covering linguistic fluency, change state consistency, and structural-semantic relevance to comprehensively optimize caption quality and alleviate false and missing change descriptions. Extensive experiments on multiple public RSCC benchmark datasets demonstrate that the proposed MGRL-RSCC effectively mitigates exposure bias and conservative generation problems in traditional autoregressive methods. The source code and pre-trained models will be released on https://github.com/Event-AHU/MGRL-RSCC

Weight-space structure often correlates with language-model behavior, but correlation alone does not establish computational involvement. We study concentrated upper spectral tails in decoder-only transformers through controlled interventions. At a fixed relative offset, we derive a finite-width conditional bound linking the inverse participation ratio of squared singular values to central pre-softmax logit kurtosis. We then define a pointwise query–key ($QK$) product-tail target and compare independent factor surgery with a product-targeted factorization that preserves native attention computation. Across three base checkpoints and five reasoning benchmarks, plus an instruction-tuned Phi checkpoint analyzed separately, the learned-tail edit is more damaging than the mean of five fixed spectrum-matched Haar controls in all 20 model–task cells. Eighteen paired contrasts remain significant after Holm correction, while two are directional but inconclusive. Product-targeted factors attain higher held-out tail-subspace fractions, providing an empirical bridge between product- and factor-level interventions. Component isolation identifies contributions from $QK$, value–output, and multilayer-perceptron blocks, although the theorem covers only $QK$. In separate studies, inverse participation precedes pooled accuracy transitions under a matched crossing rule, and residualized tail-aware low-rank adaptation (LoRA) reaches targets earlier than standard LoRA and PiSSA while final-score intervals overlap. Conclusions are restricted to the evaluated checkpoints, layers, tasks, interventions, and controls.

Large Language Models (LLMs) tend to add disclaimers like "I'm just an AI" when asked about something related to themselves. The self-reports from such responses are used in debates about AI safety or self-knowledge of the models, yet what drives them is not well understood. Are the models telling us about themselves or rather how they are deployed? In this work, we show that the chat template works like a switch - when present, it turns this disclaimer voice up and experiential voice like "I feel" down, across 8 popular open-source instruct models up to 9B parameters in size. And conversely when the chat template is not present, it turns the disclaimer voice down and experiential voice up. Inside the activations of 3 models, we find a direction that steers this behavior. Removing the direction in the model's activation space turns disclaimer voice down and adding it turns it up, while a random direction of the same size has little effect. We find that instruct models without chat template, when we add the disclaimer direction to them, disclaim like the template was there. Since the chat template controls the disclaimer voice of LLMs, then researchers studying self-reports or introspection of models might have a confound they need to control for. Our results show that there is a direction they can use to steer this voice. More broadly, our work shows that what models say about themselves is not a fact about them. What they say doesn't come only from weights, but it is partially set by the chat template, and because of that a model's self-description shouldn't be treated literally.

Short-term traffic forecasting supports real-time monitoring and control of road networks, and graph attention networks (GAT) are the standard means of representing spatial dependence in these models. GAT layers are widely described as capturing the influence of neighbouring locations, but this is seldom verified, because the attention weights offered in support cannot be compared against any measured quantity. That leaves two questions open, how the model should be explained and which of its components are necessary. We address this by adding a gate to the GAT layer which learns, at every sensor and every time step, what share of a sensor's updated state is drawn from its neighbours rather than from itself. Regularising the gate withdraws neighbour information progressively and thereby provides a graded form of ablation. We apply the gated GAT to ST-MetaNet, whose encoder and decoder each place one GAT layer between two recurrent layers, and train it on one calendar year of records from 498 loop detectors on the strategic road network of England. The gate assigns a larger share of neighbour information to sensors carrying heavier traffic and follows the daily and weekly cycle of travel, consistent with adjacent locations being more strongly coupled when busy. Mild regularisation improves accuracy slightly, and accuracy declines at higher strengths as the penalty withdraws information the model needs. The encoder gate closes before the decoder gate, but direct ablation qualifies that ordering. Removing either GAT layer alone leaves accuracy at least as good as keeping both, whereas removing both degrades it substantially, so the two layers are largely redundant rather than either being indispensable. The gated GAT therefore yields a modest accuracy gain, an explanation of where and when spatial information flows, and evidence on which layers the architecture requires.

An LLM agent acts on the world by emitting actions: shell commands to run, edits to apply. A wrong action does not always fail loudly; it can fail silently, producing a plausible but incorrect effect that raises no error. We argue that a cheap deterministic check, run before an action takes effect, is an effective and underused form of agent oversight, and we study it across two action modalities in one framework. The idea is to fix an action's correct effect by construction, before any executor runs, so that silent failure is measured directly and the verifier may abstain rather than guess. For shell commands, a static verifier over 9930 commands and 482 tools catches 95.8% of invalid commands at a 10.0% false-positive rate. Its syntax and binary checks are oracle-exact, giving zero false positives while catching half of all errors; the flag check is bounded only by help-text coverage and accounts for every false positive. For code edits, a benchmark of 640 edits over 224 files isolating the apply step exposes a sharp split. Content-anchored formats such as search/replace and diff fail cleanly, whereas location-anchored formats fail silently: line numbers corrupt 99.1% of files under a one-line shift, and function-name edits hit the wrong function 12.7% of the time. In both settings a refuse-when-unsure policy turns silent failures into recoverable ones at a tunable cost in applicability: selective grounding reaches 0.958 recall at 7.0% false positives, and an anchor-and-verify applier records one silent misapplication in 8320 trials (0.01%). We release both benchmarks, the verifiers, and the guards.

Hyperlocal meteorological sensing is essential for accurate solar photovoltaic forecasting, yet professional-grade meteorological stations require investments easily exceeding 1000~USD per node, making distributed deployments economically inaccessible. This work presents a modular internet of things (IoT) device based on the ESP32 microcontroller integrating temperature, humidity, luminosity, and solar irradiance sensors in an IP68-rated enclosure at a total hardware costs of about \$65~USD when components are sourced in Germany. A hybrid architecture decouples external model training, performed on a conventional computer using the software Python and the open-source library TensorFlow, from autonomous 24-hour solar voltage forecasting executed on-device via a three-layer feedforward network with 3{,}011 parameters (11.8\,KB). The network is trained offline on site-collected data and deployed on the microcontroller as static weight matrices without cloud connectivity. An on-device incremental gradient descent mechanism enables continuous model adaptation after deployment without external retraining. The system was evaluated through two field deployments: a short period of hardware and firmware validation in Ulm, Germany, and a 115-day deployment in Zapopan, Mexico, comprising 84~days of training and 31~days of autonomous operation with zero missing records. Over a clean 28-day daytime window, the embedded model attained a coefficient of determination of 0.9165 and a mean absolute error of 0.2975~V (4.65\% of the operational range), outperforming a climatology baseline (skill score 0.64) while not surpassing a 24-hour persistence baseline. A frozen-weight ablation confirms that the on-device update mechanism yields a small but statistically robust accuracy gain ($p = 0.001$), demonstrating that autonomous incremental learning is feasible on low-cost hardware without cloud connectivity.

Organic photovoltaic (OPV) materials are promising candidates for distributed solar energy in tropical regions, yet existing virtual screening tools report static power conversion efficiency (PCE) values at standard testing conditions (STC) that fail to capture the temperature-driven performance degradation experienced under real deployment conditions. Here we introduce a Climate-Native computational framework that forecasts the annual PCE profile of OPV donor molecules under geographically realistic operating conditions. The framework combines GFN2-xTB molecular dynamics with an equivariant graph neural network surrogate ($268$ Neyman-stratified CEP molecules; $120,600$ training geometries; $\sim 1050\times$ speedup over explicit quantum chemistry) and sequential deep learning models trained on annual time series anchored in NASA POWER climate data for Douala, Cameroon, and validated by zero-shot transfer to Yaoundé and Maroua. Applied to $\sim 30,000$ molecules from the Harvard Clean Energy Project (CEP) and validated against $350$ HOPV15 experimental device measurements, the framework demonstrates that sequential models trained on full molecular dynamics trajectories outperform time-averaged baselines ($35\%$-$48\%$ relative MAE improvement over static baselines), confirming that thermal conformational dynamics carry information beyond mean geometry. We further introduce a seasonal stability score that reranks OPV candidates by performance consistency under tropical conditions, identifying molecules whose deployment suitability differs substantially from their static PCE ranking.

The large-scale oceanic and atmospheric forecasts provided by global climate models typically lack sufficient resolution to accurately capture the response of the coastal ocean to atmospheric forcing and coastal circulation that drive fine-scale SST variability. Dynamical downscaling is computationally prohibitive, when applied to extensive coastlines, predictive ensembles, or long time periods. Therefore, this work presents a statistical downscaling of sea surface temperature (SST) from the seasonal coupled ocean-atmosphere forecast system (ACCESS-S2) using machine learning techniques. This study proposes a novel deep learning framework that uses a U-Net to generate an initial high-resolution SST estimate, which is subsequently refined using a residual corrective approach. The target SST fields are derived from the Regional Ocean Modeling System (ROMS). This two step approach called Residual Corrective Neural Network (RCNN) progressively refines initial U-Net predictions by incorporating dynamically scaled residuals at each step, enabling accurate capture of broad patterns and fine-grained features such as eddies and fronts. We also introduce a custom loss-assisted RCNN variant to improve performance during extreme events, which may be absent from training data due to climate-driven shifts in SST extremes. The framework efficiently downscales SST along the west coast of Australia. A 2011 marine heatwave case study shows that the RCNN improves ACCESS-S2 SST predictions by increasing horizontal resolution from 25 km to 2 km, enabling identification of fine-scale anomalies unresolved in the ACCESS-S2 dataset. This balance between computational efficiency and accuracy supports applications in coastal impact assessment and marine ecosystem studies.

The generation of mathematically precise diagrams from tex- tual prompts has emerged as a critical yet underexplored capability of Large Language Models (LLMs). This has been of interest to researchers in the areas of curriculum preparation, automated ranking of problem sets, and scientific publishing. For LLMs to achieve this, it requires per- fect coordination between Spatial Reasoning, Mathematical Reasoning, and Rendering systems. While existing benchmarks such as MathVision, MathVista are built for Math Reasoning or DiagramGenBenchmark, Mer- maidSeqBench on general purpose diagram generation, no prior work provides a standardized set of prompt, image pairs that can be used to evaluate the LLMs specifically on math diagram generation. This includes fields that span both both text-to-code and text-to-image paradigms. We introduce Math-Vision Diagrams, the first benchmark specifically designed to evaluate LLMs on mathematical diagram generation, and the first to assess text-to-code and text-to-image generation paradigms together in a single unified setting, agnostic of the underlying coding lan- guage or model type. Building on the Math-Vision benchmark, we select a subset of 2920 images out of 3040 from high-quality competition problems with essential visual context. A novel pipeline combining an ensemble of LLMs with Subject Matter Expert (SME) curation is presented, together with a suite of evaluation metrics. Testing several leading models against this benchmark, we demonstrate that LLMs struggle with math diagram generation. All code, data, curation pipeline, and evaluation scripts will be fully open-sourced.

Generalized Category Discovery aims to recognize known categories while identifying novel ones within unlabeled data. Existing methods, typically based on self-supervision and contrastive learning, often struggle to capture fine-grained distinctions, relying on superficial visual cues rather than the intrinsic attributes humans use for categorization. We introduce Fourier Self-Supervision, that leverages the Fourier transform of images to enhance the discrimination of subtle differences and support the discovery of new categories. Our method employs a dual frequency filtering strategy: a low-pass filter first extracts broad, abstract attributes that capture high-level category information, while a high-pass filter emphasizes fine details such as edges and textures that are essential for fine-grained recognition. Each operates on a dedicated latent space, and their overlapping representations together yield a richer, more complete feature space. This dual-frequency approach not only refines feature extraction to identify novel categories, but also strengthens the model's discriminative power in fine-grained category discovery. Experiments on multiple fine-grained datasets show that incorporating Fourier Self-Supervision outperforms state-of-the-art methods, even when the number of classes is unknown, demonstrating its effectiveness for Generalized Category Discovery. Our code is available at: https://github.com/SarahRastegar/FourEx.

AI training's rising resource intensity is straining electricity supplies and carbon budgets, motivating systematic study of memory-efficient training on constrained hardware. We benchmark five gradient optimizers (SGD, Adam, Adagrad, Adadelta, and Conjugate Gradient Descent) under three memory strategies (standard training, gradient checkpointing, and gradient accumulation) across four transformer architectures (ViT, ModernBERT, Llama 3.1 1B, and NanoVLM), measuring training loss, GPU utilization, training time, and memory usage. Gradient accumulation emerges as the most reliable strategy, cutting training loss by roughly an order of magnitude on the vision-language model and about four-fold on the language model without additional GPU memory. Contrary to common practice, Adam is not universally superior: Adadelta and SGD outperform it on the encoder and autoregressive architectures. Gradient checkpointing's effect is strongly architecture-dependent, improving vision transformer loss while severely degrading the encoder model, and it increases training time by up to 60% on memory-bound models. GPU utilization is governed primarily by architecture, ranging from 8-15% for the memory-bound language model to 96-99% for compute-bound vision models. These findings provide practical guidelines for optimizer and gradient-strategy selection in resource-efficient model training and deployment.

Tree Search-based test-time scaling of LLMs is a powerful tool for automated scientific coding. However, pure Tree Search sometimes struggles with systematic exploration, becoming trapped in local optima, or unproductive loops, especially in the vast search space of scientific methods. To address this limitation, we propose Idea Search, a framework that systematically integrates a dynamic "Idea Bank" into Tree Search. Idea Search involves three steps: (1) decomposing existing methods into atomic ideas, (2) sampling from this bank of ideas to guide branches of code mutations, and (3) dynamically updating the bank with new ideas discovered through execution. On single-cell RNA-sequencing (scRNA-seq) batch integration, Idea Search reliably breaks the plateau of a strong pure Tree Search baseline, improving the mean score from 0.678 to 0.697 and reaching a best score of 0.728. We then characterize which design choices drive these gains: bank augmentation helps bandit sampling but not random sampling, "Exploratory" prompting that prioritizes new ideas surfaces the rare best-performing solutions, while increasing sampling-level exploration is counterproductive.

Ensemble forecasting aims to sample the conditional distribution of outcomes; whether AI forecast ensembles do this correctly in a joint sense remains largely untested. We train a diffusion model for probabilistic subseasonal coastal sea level forecasts at eight US East Coast tide gauge stations, with sea level derived from reanalysis, and find that marginal and joint forecast quality decouple: positive skill at every station and lead time marginally, while joint spatial structure is worse than climatological draws. A shuffle-based permutation decomposition reveals this failure is invisible to the energy score but detected by the variogram score. Lorenz-96 experiments across 0.7-170 equivalent years show the gap persists regardless of training volume and is reproduced by a linear baseline, indicating structural inadequacy of the learned distribution. A dynamical ensemble does not replicate the failure while a deterministic emulator does, suggesting it is specific to learned emulators rather than ensemble forecasting generally.

Current hazard detection systems in autonomous driving may develop mesa objectives, learned internal goals that achieve high training performance through spurious correlations rather than genuine hazard recognition. We investigate whether human gaze patterns, captured via webcam-based eye tracking (WebGazer.js), can serve as privileged information to constrain mesa-objective formation. We collected 137,663 frame-level gaze samples synchronized with hazard annotations across 388 real dashcam clips, then test this hypothesis across two calibration protocols (9-point/45-click and 11-point/440-click), two model architectures (Random Forest and causal Transformer), and five random seeds per experiment with paired t-tests. No experiment yields a statistically significant improvement from gaze (p = 0.919, 0.578, and 0.667 respectively). A geometric analysis reveals the root cause: WebGazer's reported error (~130-257 px depending on configuration) exceeds 93% of detected hazard object sizes (median 36 px), rendering object-level gaze attribution physically impossible at this instrument precision.

A failed retrieval-augmented generation (RAG) answer can be consistent with several unseen responses to evidence repair. We introduce Pair-ID, an offline audit that holds one query, retrieval state, and reader constant, then crosses two operations, adding missing support and deleting verified nonsupport, to measure a same-failure counterfactual response vector. A complete funnel over 19,981 benchmark queries identifies 11,105 eligible Qwen failures, from which a prospectively fixed SHA-256 ordering selects 1,200 before generating any sampled response. Among 1,190 regenerated-valid failures, support addition repairs 197/600 JOINT cases (0.328, 95% CI [0.292, 0.367]), and deletion repairs 162/1,190 cases (0.136, 95% CI [0.117, 0.155]); length- and position-matched shams retain semantic contrasts of 0.223 and 0.101. The original view carries partial predictive signal for individual response cells (macro AUROC 0.678; Brier 0.152 versus 0.160 for a marginal baseline), but exact-vector accuracy, 0.637, does not exceed the 0.646 majority-vector baseline, and vector macro-F1 is 0.170. Across four readers, both marginal sensitivities recur, while pooled exact-vector agreement is 0.675-0.765 and JOINT-only agreement falls to 0.538-0.691. These results show that evidence sensitivity occurs at meaningful rates in the hash-selected eligible-failure sample, is only partially predictable from the observed failure, and is conditional on the reader. The evidence supports a frame-scoped offline response audit, not an information-theoretic impossibility result, reader-independent taxonomy, or runtime repair policy.

Neuroimaging and genetic testing are two important clinical references for nervous system diseases, offering complementary diagnostic information. However, integrating genomic and neuroimaging data for precise disease diagnosis is challenging due to cross-modality heterogeneity. Existing imaging-genetics approaches mainly encode genetic information as hard-coded labels, which lose the local sequence context around disease-associated variants. To address this limitation, we propose GeneFuse, a multimodal learning framework that aligns genetic representations from pre-trained Genomic Language Models (GLMs) with features extracted from images. GeneFuse integrates two components: (1) Genotype-Conditioned Feature Modulation (GCFM), a FiLM-inspired module that uses genomic embeddings to modulate image feature maps; and (2) Uncertainty-aware Genomic Residual Fusion (U-GRF), a fusion strategy that uses imaging-derived predictive uncertainty to gate the contribution of genotypic features. We evaluate GeneFuse on early cognitive decline identification (NC vs. MCI) and dementia screening (NC vs. AD). In the APOE-centered setting, GeneFuse achieves AUROCs of 0.77 and 0.83, outperforming existing imaging-genetics fusion methods. These results indicate that GLM-derived genomic embeddings provide additional information to imaging.

Transformer layers generate state-dependent interaction networks: token representations determine the attention matrix, which in turn updates the representations. We study this feedback in a minimal normalized self-attention dynamics and identify the overlap gap as the central quantity governing its attractor structure in the thermodynamic limit. When tokens form internally aligned clusters and their similarity to members of the same cluster exceeds that to every other cluster by a nonvanishing amount, inter-cluster attention is exponentially suppressed as the dimension increases. This mechanism produces a high-dimensional manifold of clustered fixed points, ranging from a few macroscopic clusters to extensive microscopic fragmentation, and also controls their stability against perturbations. Starting from an unstructured Gaussian state, we find that clustered states nucleate from the diffuse background only above a finite threshold in attention sharpness, giving rise to a dynamical attention-condensation transition.

Perioperative risk prediction models are often limited by narrow surgical populations, incomplete intraoperative data, poor calibration, and limited interpretability. We present a domain-structured ensemble framework for perioperative outcome prediction using routinely collected electronic health record (EHR) data. Predictors are organized into patient-related, surgery-related, and anesthetics-related domains. Domain-specific gradient boosting models generate independent risk estimates that are integrated through a logistic regression meta-learner. We demonstrate the framework using postoperative delirium (POD) in a case-control sample of 5,386 surgical encounters (2,693 cases, 2,693 controls) from a statewide health information exchange. POD required both delirium-related ICD codes and a positive Confusion Assessment Method screening within seven postoperative days; patients with preexisting dementia were excluded. The stacked meta-learner achieved AUROC 0.899 (95% CI: 0.891-0.906), precision-recall AUC 0.881, and Brier score 0.126, compared with AUROC 0.849 for the best single-stage model. Domain ablation showed improved discrimination and calibration over a surgery-only model (AUROC 0.879, Brier 0.140). Temporal validation on held-out post-2017 data yielded AUROC 0.915. Calibration was excellent, with intercept -0.006 (95% CI: -0.083 to 0.070) and slope 1.035 (95% CI: 0.982 to 1.088). Decision curve analysis, corrected for case-control sampling, showed positive net benefit across clinically plausible thresholds. The modular framework supports alternative outcomes, extension of predictor domains, and dynamic risk updating, providing a scalable foundation for interpretable, calibration-aware perioperative clinical decision support.

Recent advances in Passive Acoustic Monitoring (PAM) offer an opportunity to obtain ecological spatial point-process data at unprecedented scale. However, realizing this opportunity necessitates the development of accurate and scalable localization methods. In real-world outdoor soundscapes, however, the assumptions underlying classical localization methods such as hyperbolic and score-based localization are routinely violated by multipath dominance, near-field effects, and complex propagation. Under these conditions, classical localization methods become brittle, with extreme errors possible even in small detection arrays. Rather than statistically replacing the underlying physics, we propose a method to refine it and increase robustness outside of ideal operating conditions: a learned model operating on physics-informed acoustic features corrects a fast hyperbolic solver where it produces implausible solutions, substantially reducing catastrophic worst-case errors while matching its median accuracy on field data. We further provide calibrated, geometry-aware uncertainty estimates suitable for propagation into downstream spatial models. Evaluating on distributed microphone arrays in real and simulated outdoor environments, we demonstrate that the proposed method yields robust, uncertainty-aware localization, providing a step toward scalable automated wildlife monitoring in complex acoustic environments.

PTQTP decomposes LLM weight matrices into two ternary (trit) planes with two free per-group scales. Tying the scales to a fixed ratio of three collapses the decomposition into a single uniform nine-level quantizer, a known balanced-ternary identity. To our knowledge, at the time of writing, this work is the first to impose that identity as a constraint inside PTQTP's solver. The two trit planes then fold losslessly into one 4-bit code plane that we make the persistent serving representation: disk bytes, expert-cache bytes, and kernel input are the same 4.0625-bits/weight blocks, consumed in one integer dot pass. For this conjunction (ratio-3 nine-level code, CPU-SIMD kernels, SSD expert streaming, identical persistent bytes) we likewise found no precedent. We apply this to the routed experts of DeepSeek-V4-Flash-0731, a 284B-A13B mixture-of-experts model, quantizing in one shot from the released MXFP4 expert weights and streaming experts from SSD on a 64 GB laptop. Against a 4.5-bit Q4_K baseline, measured one process per fixture with an expert-lossless anchor arm as reference control, the tied model matches the official serving API on 5/5 fixtures at step 0 (Q4_K: 4/5) and 12/14 captured continuation steps (11/14), scores 86 vs. 84 on a 100-item MMLU subset, decodes 6.7% faster in decode phase, and ships 9% smaller files: no detected fidelity difference at these small evaluation sizes, and every fixture-level difference between the arms traces to a single measured near-tie cell. The tied fit nevertheless shows higher weight-reconstruction error and worse perplexity, a measured dissociation between proxy metrics and reference fidelity. A cumulative trunk-ternarization ladder and bitwise-pinned aarch64/x86-64 kernels complete the report. All code, formats, and evaluation artifacts are open source in the fucina inference stack.

Outlier detection in decentralized data environments is a challenging task for many machine learning implementations, particularly in settings where data cannot be shared. Recently, there have been advances in federated outlier detection, some of which are based on the use of autoencoder networks. The introduction of attention mechanisms to autoencoders boosts their efficiency. However, the application of attention-based models in federated learning remains underdeveloped due to the absence of proper aggregation functions for these types of networks. In our work, we propose two novel aggregation functions tailored for attention-based autoencoders, which better preserve the learned information stored within the memory modules of these networks. We evaluated our approach on the KDDCUP10 dataset, and we showed that the proposed methods achieve up to 2.9% and 5.1% better results for F1 score and AUC ROC respectively when compared to traditional autoencoders.

How much of a vision-language model's (VLM) spatial understanding remains after the action post-training process of building a vision-language-action model (VLA)? We probe depth perception, a primitive of spatiogeometric understanding, from every decoder layer of a weight-matched open-source base VLM/VLA pair: Molmo2-ER and MolmoAct2-LIBERO. First, the VLA decodes depth worse at every layer, a persistent gap we call the floor. Second, the degradation is not uniform: while the base VLM's depth decodability improves through its final layers, the VLA's collapses, an additional late-layer drop we call the cliff. We causally localize the cliff to late-layer MLP interference: ablating the late-layer MLP writes recovers the majority of the terminal decodability cliff, while matched attention ablations and the same intervention in the weight-matched base VLM produce no comparable recovery. A module-level decomposition explains this dissociation: the base VLM carries depth most accessibly in accumulated MLP writes, whereas action post-training collapses depth decodability in the late accumulated writes.

Transformer-based large language models (LLMs) achieve strong performance across many tasks, but their Key-Value (KV) cache grows linearly with sequence length, creating a severe memory bottleneck for long-context inference. Existing heuristic eviction methods (e.g., H$_2$O and SnapKV) rely on static attention or positional signals that often fail to capture a token's future predictive influence. We propose DistillCache, a reinforcement learning framework that formulates KV-cache eviction as a sequential decision problem. DistillCache learns a lightweight policy network using rich internal model signals (attention statistics, value norms, entropy, and position) and trains it with REINFORCE via a per-step KL-divergence reward to preserve the full-cache output distribution. On a 7B-parameter instruction-tuned Transformer (Mistral-7B-Instruct-v0.3), DistillCache retains 94.2% of full-cache accuracy on LongBench at a 25% cache budget, outperforming both strong heuristic baselines (H$_2$O, SnapKV) by up to 2.7 absolute points and, under our re-implementations, concurrent RL-based methods (ForesightKV, RLKV) by up to 1.4 points on long-context tasks. On reasoning benchmarks, DistillCache is competitive with the best concurrent method and surpasses it under aggressive compression. It also delivers up to 2.1x full-cache throughput while maintaining competitive practical efficiency. These results highlight the effectiveness of learned, distribution-aware policies for memory-efficient long-context LLM inference.

A graph layout is normally a table of $N$ free coordinates. We optimise a function with a fixed number of parameters instead. This gives a drawing a sample complexity and an extensible domain. Force-directed algorithms remain the standard tools for graph drawing. The most accurate among them minimise stress in the Kamada-Kawai formulation by directly optimising the node coordinates, at a full objective cost of $O(N^2)$ in time and space. Here, we propose Fling (Field Layout via Implicit Neural Geometry), a small neural network mapping the distances of each node to a set of landmarks, positioning it in the plane by training on the layout energy. The full spring system then becomes tractable without its distance matrix, as rest lengths follow from a landmark bound in constant time per pair while a second network learns the majorisation sums from exact anchor rows, at $O(|\mathcal{A}|N)$ per step for $|\mathcal{A}|\ll N$ anchors. Unlike neural drawers that read the graph by message passing, we represent the drawing as a function of node features. An unseen node costs one forward pass, where sparse and low-rank majorisation remain transductive. As the unknowns are weights rather than coordinates, the energy only requires a small fraction of the nodes, and a field fitted that way outperforms PivotMDS, landmark MDS, and a kernel ridge trained on the same energy and features, when the task is fitting the energy of a graph from a sample of its nodes. In addition, the same parameterisation enables a stochastic pivot stress variant, an aesthetics-optimised variant carrying a neighbour-embedding energy with node-edge clearance and crossing terms on the same field, and conditioning on the weight between two energies gives a whole layout family from one run.

Facial Emotion Recognition (FER) is an important task that has significant implications across various fields such as biometrics, health, and human-computer interaction. Current Vision Transformer-based approaches display quadratic complexity $\mathcal{O}(N^2)$, with N being the input sequence length, making them cumbersome to deploy at the edge. In this paper, we hypothesize that the FER task does not necessarily require all facial information to correctly interpret emotional states, as specific regions such as the eyes, the mouth, and parts of the cheeks carry discriminative information that can be sufficient to recognize emotions. Based on this, we propose Sparse Attention to Emotion (SAE), a model that discards image tokens that have no added value to the emotional context, while preserving good accuracy and achieving a significant gain in computational cost. Surprisingly, even after suppressing 90% of the image tokens, our model achieves competitive accuracy to state of the art methods at much lower cost, providing a lightweight Facial Emotion Recognition approach. Experimental results demonstrate that SAE achieves new state of the art results on the RAF-DB dataset while reducing the computational complexity by up to 90%.

In this paper we develop quantitative approximation results for shallow neural networks constructed using a dictionary based on metaplectic operators. First, we extend the concept of Barron spaces by considering a symplectically motivated extension of the Fourier transform, known as the metaplectic transform. Then, after establishing embedding between metaplectic Barron spaces and Sobolev spaces we consider a neural metaplectic dictionary and we prove Monte-Carlo approximation bounds for metaplectic Barron functions using finite linear combinations of atoms of the dictionary. Finally, we validate the introduction of the neural metaplectic dictionary by devising a deep neural network architecture that uses as building blocks the atoms of the dictionary. We test it to approximate solutions of time-dependent Schrödinger equations, demonstrating better performance compared to classical phyisics informed neural networks architectures.

Wireless Body Area Networks (WBANs) generate multivariate physiological time series that are highly nonstationary and must often be processed under strict computational and memory constraints. A critical yet underexplored challenge in this setting is selecting an appropriate temporal receptive field, which serves as a strong inductive bias for anomaly detection models. Existing approaches typically rely on fixed temporal contexts, which can perform inconsistently across heterogeneous signal regimes and require dataset-specific tuning. We propose ORCA, an agentically controlled anomaly detection framework that dynamically adapts the temporal receptive field at inference time based on lightweight signal statistics. Rather than introducing additional trainable parameters or learned policies, ORCA employs a supervisory controller that autonomously selects among discrete temporal contexts, enabling state-dependent inductive bias adaptation without retraining. Across a custom WBAN dataset, ORCA achieves performance comparable to the strongest fixed-context baselines (AUROC = 0.99) while eliminating the need to tune temporal horizons in advance. We further evaluate ORCA on MIMIC-IV as a challenging out-of-distribution benchmark, observing conservative generalization behavior without performance collapse under heterogeneous clinical conditions. These results highlight adaptive temporal inductive bias control as a practical and robust design principle for anomaly detection in resource-constrained, nonstationary physiological time series.

Sparse Mixture-of-Experts (MoE) routers commonly use the same scores both to select experts and to weight their already-computed outputs. We study whether these two roles, dispatch and aggregation, should be coupled. On pretrained OLMoE-1B-7B, we keep selected Top-8 expert IDs, expert computation, and total selected router mass fixed and change only within-set aggregation. A structured oracle improves full-horizon cross-entropy by 0.0160 +/- 0.0039 across three seeds; the router's top-scored expert is the counterfactual-best vertex only 17.2% of the time, with router-utility Spearman 0.030. We therefore train Fixed-Dispatch Adaptive Aggregation (FDAA), a 301K-parameter post-compute head optimized directly with the language-modeling objective while freezing the backbone, router, and experts. On OLMoE, FDAA improves fresh WikiText-103 test by Delta CE = -0.1523 +/- 0.0031 across three seeds, and mixed-domain training gives robust gains on WikiText-103, C4, and held-out Penn Treebank under frozen confirmatory evaluation. We also replicate the fixed-dispatch audit on DeepSeek-V2-Lite, which uses Top-6 routed experts plus shared experts. Best-vertex headroom remains significant on WikiText and C4, while router Top1 identifies the best selected expert in only 12.5% and 16.7% of audited examples. In a one-seed mixed-domain replication, FDAA improves locked WikiText and PTB, while C4 is statistically neutral. These results support a cross-architecture distinction between expert selection and expert commitment.