Visual In-Context Learning (VICL) enables promptable segmentation by conditioning RGB predictions on visual demonstrations, making it attractive for auto-labeling, but its robustness under query-only access and unknown deployment prompts remains underexplored. We study query-only, prompt-agnostic targeted attacks where the attacker observes only returned RGB outputs, without access to model internals or deployment prompts. Treating prompt-conditioned RGB rendering as a distinct attack surface, we introduce two appearance-space attacks: Color Hijack (CH) drives foreground colors toward the background to suppress visibility, while Task Hijack (TH) collapses instance-distinguishing colors toward a shared color assignment. To transfer across unseen prompts, we propose a Self-Prompt Surrogate (SPS) constructed from the clean query and its output. In a paired, matched-budget SegGPT comparison, SPS outperforms the Random-Prompt Surrogate (RPS), reducing CH instance foreground retention from 0.918 to 0.155 and TH palette-aware tolerance distance from 0.345 to 0.225. Across four VICL models on COCO and HierText, CH suppresses foregrounds across semantic-, instance-, and text-style outputs, while TH collapses instance colors under held-out deployment prompts. After palette-based decoding, TH also reduces the decoded instance count and mask retention while increasing instance merges. Finally, at 30% corruption, CH-corrupted annotations from a publicly accessible VICL-based auto-labeling pipeline reduce downstream YOLO-seg mAP50 from 0.761 to 0.281.
Progress in 3D sign generation is constrained by the scarcity of high-fidelity 3D sign language datasets, a limitation that is particularly severe for under-resourced languages. We introduce MySign, the first motion-capture dataset for isolated signs in Bahasa Isyarat Malaysia (BIM). Its 5,000 samples cover 1,000 lexical items (glosses) and were produced by five Deaf native signers. Prior datasets that rely on SMPL-X parameters regressed from video provide pseudo-labels, which are often affected by depth ambiguity, inaccurate finger poses, and temporal jitter. MySign instead provides ground-truth SMPL-X sequences retargeted from recordings made with a motion-capture setup tailored for sign language; these sequences offer higher-fidelity motion, particularly in fine-grained hand and finger articulation. A controlled comparison on the same glosses and signers indicates that training on these sequences improves generation over training on such pseudo-labels. We establish MySign as a challenging benchmark for gloss-conditioned 3D sign generation: state-of-the-art models produce visually plausible outputs but fall short of the ground truth, most notably in fine-grained articulation, revealing persistent gaps in modeling articulated sign motion. In contrast, recognition baselines consistently achieve high lexical accuracy, indicating that the dataset reliably supports supervised learning. The dataset is available at https://huggingface.co/datasets/mysigner/MySign
Multi-stage large language model (LLM) cascades balance accuracy and inference cost, but reliable deployment also requires controlling error risk and identifying its sources across stages. We propose Compositional Conformal Risk Control (CompCRC), a risk-control framework for pre-specified multi-stage LLM cascades. We decompose the risk of accepting incorrect answers into additive stage-wise contributions over the full query population, linking local risk budgets to a system-level guarantee. CompCRC constructs and freezes candidate budget allocations and acceptance thresholds, selects one candidate on independent data, and certifies its stage-wise risk contributions on separate held-out data. Under conditional exchangeability, we establish a finite-sample guarantee: the marginal expected accepted-error risk of the complete deployment procedure does not exceed the global risk budget. At the reported comparison budget of $\alpha=0.45$, all three stages of CompCRC pass their local checks on MMLU, GSM8K, and HellaSwag. Compared with Global-CRC, which passes system-level risk certification at the same budget, CompCRC reduces observed accepted-error risk by 1.33--2.00 percentage points, improves final accuracy by 0.67--2.00 percentage points, and reduces weighted inference cost by 30.37\% on HellaSwag. The selected cost-based budget allocation also improves accuracy per unit cost by 10.6\%--46.7\% over uniform allocation. Further experiments show that stage-wise statistics identify stages that fail local certification checks, providing diagnostic information alongside system-level risk control.
Large language models (LLMs) are increasingly used for text annotation, but their outputs are often reduced to a single categorical label before downstream analysis. This hardening step discards the model's relative support across categories, which may contain useful information when category boundaries are ambiguous. We study whether retaining the full probability vector improves downstream regression when LLM annotations are combined with a limited amount of human annotation. Using matched soft and hard representations from three LLMs across three social-science text datasets, we compare direct regression and human-assisted correction. Soft inputs are not uniformly better when used directly as outcomes, but after human correction, using the probability vector consistently improves regression efficiency across models and datasets. Analyses of repeated human judgments and controlled simulations further show that these gains arise when probability variation beyond the hard category carries useful information about the human-coded outcome. Our results identify output representation as an important design choice in LLM-assisted measurement and support preserving probability outputs for downstream statistical analysis. Code and reproducibility materials are available at https://anonymous.4open.science/r/soft-label-LLM-D201/
Theoretical results in statistics and machine learning rest on long, technically delicate proofs, and fluent arguments generated by large language models are not necessarily correct. Automated formalization makes such arguments mechanically checkable, and automated proving goes further by discovering machine-checked proofs for statements whose proofs are incomplete or missing altogether. Both require reconstructing assumptions from source materials, accessing specialized mathematical infrastructure and managing long proof developments. We present Statopia, an ecosystem that jointly develops Statopia Prover, an independent agent specialized for formalizing and proving results in statistics and machine learning, and Statlib, a Lean~4 library built with the prover and organized for automated use. The prover controls its own execution without relying on external coding agents. It is built for formal proof discovery and checking existing proofs. A natural-language proof, when available, guides its search but is not required: from a statement alone, the prover develops the argument itself and constructs a Lean proof. Its workflow integrates source preparation, statement review, library supported planning and proof construction. Dynamic proof dependency graphs and task groups support parallel, resumable proof development. Statlib's shared definitions and theorem interfaces are organized for retrieval and composition by agents, and its coverage of statistics keeps expanding. Verified results and usage feedback from the prover inform reviewed library updates, so that proof development drives this expansion. A cross-platform application provides an interactive workspace for uploading source files and running tasks locally or in the cloud.
Generative samplers for lattice field theory are commonly evaluated by their effective sample size (ESS). We show that a small global ESS need not imply inaccurate long-distance observables: errors spread over many short-distance modes can dominate this diagnostic. We introduce a scale-resolved diagnostic based on the overlap between coarse-grained model and target distributions. At the lattice cutoff, this overlap determines the asymptotic ESS fraction; lower cutoffs isolate long-wavelength errors. In the exactly solvable free theory, two samplers with comparable ESS yield a $32\%$ bias and no bias in the susceptibility, respectively. In two-dimensional $\phi^4$ theory near criticality, trained diffusion samplers show the same separation of scales: their path-weight ESS fraction falls rapidly over the volumes studied, while low moments of the magnetization remain close to their target values. Varying the reference parameters substantially changes infrared observables, with little effect on the log-weight variance. Given the target magnetization distribution, a marginal Metropolis test maintains an acceptance rate of order one in our tests. Combined with exact local updates, it reduces the remaining mismatch, although residual bias remains at finite relaxation length. In the free field, uniformly bounded relative errors in the proposal mode variances give a volume-independent bound on the number of updates needed for a prescribed accuracy of the chain's stationary mode variances at fixed correlation length. The target marginal must be supplied or computed separately; estimating it from the samplers' own weights can inherit the poor global overlap.
In edge intelligence systems, computation offloading and dynamic voltage and frequency scaling (DVFS) are widely used to improve energy efficiency. However, as applications like deep neural network (DNN) inference evolve in complexity, their demands for memory resources have progressively increased, establishing memory as a critical bottleneck that impacts both latency and energy consumption. In this paper, we propose a joint optimization scheme integrating memory, computing and offloading for energy-efficient DNN inference. We formulate the energy minimization problem under deadline constraints as a mixed-integer nonlinear programming (MINLP) problem, where the main challenge arises from the coupling between offloading and resource optimization. To address this issue, we decompose the original problem into two subproblems under a fixed offloading policy. The local subproblem is transformed into an equivalent geometric programming (GP) problem to derive the optimal memory frequency and computing frequency, while the edge subproblem is solved via convex optimization. An efficient coordinate descent (CD) algorithm is further developed to jointly optimize all variables. Simulation results demonstrate that our proposed algorithm achieves near-optimal performance with the gap from the optimal policy being at most 9.3%, and reduces the energy consumption of devices by up to 16.1% compared with benchmarks without memory frequency optimization.
Mathematical performance depends on how a language model receives information, uses generated text and submits an answer. We study these dependencies through paired comparisons that separate submission requirements, draft reuse and candidate availability. On 65 development problems, an instruction permitting continued derivation produced no valid final answers in 390 short-output calls; a single-line submission instruction produced 390 format-valid outputs, including 389 parseable answers. Under the fixed submission protocol, 198 problems with failed initial submissions were answered correctly in 9.09p of cases with the question alone, 33.84p with its draft, 46.46p with candidates, and 59.60p with both. Drafts added 13.13 percentage points with candidates fixed, and candidates added 25.76 points with drafts fixed; both effects were significant after Holm correction and were subadditive on the accuracy scale. Retaining a trailing excerpt outperformed retaining a leading excerpt of similar average length by 10.10 points Holm-adjusted p 0.027. A second completion model showed the same directions of input effects on the fixed materials. In a cohort with parseable initial submissions, candidates added no correct answers when the previous draft was supplied, but added 27 points when it was absent. These results establish a conditional value of generated text: a failed submission can leave useful material, whose contribution depends on what else is supplied and what is retained.
Compressing continuous human motion into a finite vocabulary of discrete tokens is now the first stage of many motion understanding and generation pipelines. Vector quantization partitions a latent space into Voronoi cells around a finite codebook, but the geometry of that latent space is inherited from the input pose encoding, and it is unclear which encoding admits an efficient discretization. We study four encodings that place skeletal motion in different geometric spaces: Gaussian joint heatmaps on the pixel lattice, which we show to be a morphological dilation of the joint point set by a quadratic structuring function; root-relative joint positions in Euclidean space; local joint rotations on $\SO$ through a continuous 6D embedding; and a redundant full kinematic encoding. For each, we train VQ-VAE tokenizers whose codebooks lie on the unit hypersphere, varying codebook size, temporal stride, and code dimension. Beyond reconstruction, we measure codebook utilization and the effective vocabulary size, i.e., the number of equiprobable spherical Voronoi cells with the same entropy as the observed token distribution. Rotation and position encodings keep converting additional codewords into effective vocabulary, whereas the full kinematic encoding saturates at a few hundred effective cells and heatmaps at fewer than sixty. The smallest code dimension gives the highest utilization for three of the four encodings, and the effect of temporal stride depends on the encoding. Encodings of the same pose family can thus yield effective vocabularies differing by an order of magnitude, a difference that reconstruction loss alone does not reveal, which points to input geometry as a first-order design choice in motion tokenization.
In reinforcement learning from limited feedback (RLLF), only a small fraction of an offline dataset can be labeled with rewards, and the central question is which samples should be labeled to learn a high-performing policy from the resulting partially labeled dataset. Prior work formalized this as a reward-selection problem by focusing on the selection stage while treating downstream policy learning as a black box, in a regime where queried rewards are not retained for reward-model training. We instead study the retained-label setting, where queried rewards can be stored and used to fit a reward model before policy learning. We bound the suboptimality of the learned policy in terms of two sources of error: one arising from offline RL on a fixed, offline dataset, and the other from uncertainty in the reward model. Since reward selection cannot change the offline dataset, the limited labeling budget must be used to strategically reduce reward uncertainty. Motivated by RLLF's observation that useful rewards tend to keep the agent on high-return trajectories, we introduce Successor-Guided Uncertainty REduction (SURE), a novel method that uses successor features to select rewards that are both uncertain and associated with states from which high-valued states are reachable. We formally derive SURE from a bound-induced design objective and characterize its exact one-step marginal gain. Empirically, SURE reaches near full-feedback performance with few reward labels across a variety of domains and offline-RL backbones, yielding an effective method for feedback-efficient reinforcement learning.
We present SoftPAT3D, a physics-augmented framework for text-to-3D scene generation in which deformable objects, such as cloth, stuffed toys and ropes, are generated as physically valid, simulation-ready bodies and placed in intersection-free contact with the furniture that supports them, with one another and with themselves. Existing physics-augmented pipelines simulate rigid bodies only and cannot accommodate a generated deformable object, which text-to-3D generation returns as a geometrically invalid mesh, sometimes self-intersecting, that has no rest shape, no material model and no admissible initial state. These are the inputs a deformable simulation requires, and we construct each from information already in the pipeline. The rest shape comes from how the object is manufactured rather than from the generated physically inconsistent deformed configuration, and a material library assigns each deformable object its initial elastic parameters, which can then be optimized. The occluded support is regenerated and converted into a watertight collision mesh, each object is released from an optimizable placement above the supporting part that the prompt specifies, and the complete scene is driven to static equilibrium under one unified contact model, with completion verified by geometric checks and a vision-language model. When a scene fails the prompt, a grid search over placements corrects it, and the adjoint of a differentiable simulation of the same contact energy reaches the targets the grid does not, places several bodies jointly and recovers and optimizes material parameters. Multi-body scenes equilibrate in a single simulation, and the pipeline also runs from the text prompt alone. On all sixty-five prompts, SoftPAT3D produces no undesired interpenetration of any kind, whereas every other system does, and its scenes directly support scene editing and robotic manipulation.
We introduce HBAR (Hierarchical Brain Activity Representation), a hierarchical tokenization framework for learning compact, multi-scale representations of neuroimaging data. HBAR organizes voxel-level resting-state fMRI into a hierarchy spanning sub-parcel voxel groups, parcels, and large-scale functional networks, enabling up to 415x compression while preserving fine-grained spatial information beyond conventional parcellation. The framework supports fixed atlas-based, fully learned, and softly prior-guided hierarchies, allowing us to systematically study the role of structured priors in brain representation learning. On HCP-YA, at approximately 415x compression, HBAR achieves higher parcel and network correlations than rate-matched flat tokenizers, while at a lower 52x compression rate, shuffling the anatomical grouping reduces parcel correlation from 0.920 to 0.886. Demographic readouts with autoregressive training also reveal potential gains from the hierarchical structure. By combining sub-parcel inputs, compact discrete tokens, and multiscale brain structure, HBAR offers a practical framework for representation learning in high-dimensional neuroimaging data.
Low-bit post-training quantization reduces language-model memory but can sharply degrade downstream task accuracy. Calibration usually relies on generic text, but when task-informed data recover accuracy and which quantizers benefit remain poorly understood. We introduce FragCal-DC, which supplements generic text with training documents from task families with low retention of full-precision accuracy in a two-bit scan. We use a two- to four-bit sweep to locate functional regimes and four-seed calibration comparisons to evaluate six data-dependent methods across six dense models (1.7B-8B) and ten tasks. At their lowest tested functional bit widths, iterative SignRound V2 and OmniQuant improve at two bits, whereas single-pass GPTQ and AWQ show small mean responses at three bits. At two bits, SignRound V2 gains 1.83 percentage points in mean ten-task accuracy over generic-text calibration across six models. At 70 billion parameters, GPTQ and SignRound V2 both improve at two bits, with functional generic-calibration baselines. For two-bit SignRound V2, task-level controls show that recovery concentrates on families included in calibration, while incorrect answer designations retain about 90% of the mean accuracy gain. Changes in reconstruction error do not consistently track accuracy gains. These findings support task-informed calibration as a design principle for preserving downstream accuracy in low-bit quantization.
Dynamic neural networks are a class of adaptive models in which the computation assigned to each input sample is determined dynamically, potentially reducing computational cost while maintaining similar performance. For image classification, the Resolution Adaptive Network (RANet) is a strong approach that progressively allocates computational resources across input scales based on sample difficulty. In this work, we introduce RANet with Attention, Distillation and Calibration (ADC-RANet), an enhanced version of RANet that improves performance across all computational budgets. Our approach combines three tightly connected components: bottom-up channel attention, top-down cascaded knowledge distillation, and overconfidence correction using MixUp, in a synergistic design that coordinates both pathways. Experiments on CIFAR-100 and ImageNet datasets show consistent gains while preserving the efficiency of RANet. Our code is available: https://github.com/jefersonseverino/ADC-RANet
Large language model safety evaluations often focus on a compliance-refusal tradeoff: models should refuse unsafe requests without refusing safe ones. We argue that this misses a broader form of prompt-conditioned task selection, where the same request can elicit different behaviors depending on the system prompt. We investigate this using a controlled evaluation crossing eight safety-task prompt families with three clarity levels across HarmBench, XSTest, and IHEval. Holding benchmark items fixed, we evaluate 13 models and use signal detection theory to distinguish two possibilities: whether models become better at telling safe from unsafe requests (discriminability), or simply more willing to refuse (response criterion). We find substantial differences in discriminability across models, but system prompts primarily shift the willingness to refuse, with smaller and less consistent effects on discriminability. Increasing prompt clarity makes models more willing to refuse without reliably improving their ability to distinguish safe from unsafe requests. Beyond refusal, prompt families also shift responses among six distinct behavioral policies, including clarification and safe help. Together, these findings show that system prompts primarily change how models respond to safety-relevant requests rather than consistently improving their ability to distinguish safe from unsafe ones.
Automated brain tumor classification from magnetic resonance imaging (MRI) can support accessible medical image analysis, but conventional deep learning models often exceed the resource budgets of highly constrained edge-AI platforms. We present NanoNet, an ultra-lightweight convolutional neural network designed for resourceconstrained MRI classification and motivated by future edge-AI applications. Inspired by MobileNetV2, NanoNet substantially reduces network depth, channel capacity, expansion ratios, and final feature dimensionality, resulting in only five inverted residual blocks and 8,132 parameters. We further investigate dependency-aware structured pruning, knowledge distillation, and INT8 quantization-aware training for additional compression. Experiments on the BRISC 2025 benchmark show that NanoNet achieves 92.80% accuracy and 92.97% macro F1 with 8,132 parameters and 10.498 MFLOPs. Across three compression seeds, it exceeds the mean accuracy of 70%-pruned MobileNetV2 by 5.20 percentage points while using 60.2% fewer parameters and 9.8% less computation. Converted INT8 NanoNet achieves 89.70% accuracy and 90.22% macro F1 across three QAT seeds. Earlier TensorFlow Lite measurements indicate 27.7 KB of baseline model storage, motivating future edge-AI deployment.
Discrete diffusion is a promising framework for categorical generative modeling, but current evaluation protocols often fail to reveal whether models have actually learned the target distribution. We introduce the Discrete Diffusion Playground, a diagnostic benchmark that converts known 2D continuous distributions into binary token sequences through an exact discretized bijection. This creates a rare setting in which models are trained on discrete sequences, but generated samples can be decoded back into 2D and compared directly against a known ground-truth distribution. The benchmark spans smooth Gaussian mixtures and tilted checkerboards with sharp boundaries, disconnected support, and coordinate-dependent structure, allowing it to expose failures that are invisible in standard proxy metrics. It supports exact grid divergences, Wasserstein distances, density visualizations, and nonparametric two-sample tests, providing both quantitative evaluation and interpretable failure diagnosis. Using this testbed, we show that masked diffusion models can learn meaningful adaptive reveal orders but remain less stable than autoregressive baselines on smooth targets and largely fail on tilted checkerboards, where autoregressive models preserve the target geometry. Control experiments with reversed and interleaved token orderings show that autoregressive performance is not explained by favorable bit order alone. The Playground provides a simple but stringent benchmark for stress-testing discrete generative models and for developing evaluation metrics that detect real distributional mismatch rather than proxy performance. Our code is available at https://github.com/shaojus/DiscreteDiffusionPlayground .
Large language models (LLMs) are increasingly deployed in user-facing systems, where jailbreak attacks can bypass safety controls and induce harmful and policy-violating outputs. Recent model-internal detectors use hidden representations, gradients, or internal signals to identify such prompts, but their practical security value depends critically on behavior at low false-positive operating points rather than on area under the receiver operating characteristic curve (AUROC) alone. We study two data dependencies that shape this behavior: calibration data, which determine the detection threshold, and reference data, which can influence detector scores. Across four model-internal detectors, four LLMs, and twelve jailbreak families, thresholds calibrated on curated benign prompts frequently fail to preserve their false-positive rate (FPR) budgets on real-user traffic. Recalibration improves false-positive control but reveals substantial family-level coverage gaps, including cases with high AUROC but low true-positive rate (TPR). Causal reference-set interventions show that changing reference composition can substantially alter family-level detection and that improving coverage of one attack family can reduce coverage of others. These findings show that calibration and reference data are part of a detector's security behavior. We therefore recommend evaluating model-internal jailbreak defenses using realized FPR on deployment-like traffic and per-family TPR at explicit low-FPR operating points, with AUROC used as a complementary ranking metric rather than the primary security measure.
Earth system models underpin climate projections, but their cost limits how many scenarios, ensemble members and impact pathways can be explored. Machine-learning weather models forecast as skillfully as physics-based models at a fraction of the cost, yet coupled to an ocean they have produced weak El Nino variability or have learned it from climate-model output, and none carries the climate through rivers to the people exposed to it. Here we present Rescene-ESM, an Earth system model whose coupler, YAHO, yokes a frozen 20.8-million-parameter AI weather model to physically formulated ocean, sea-ice, land, river, carbon and energy components constrained by observations. Adding the couplings one at a time shows what each contributes: over a climatological ocean the AI atmosphere varies from year to year by a third of the observed amount over the oceans, and an interactive mixed layer, an equatorial wave ocean, the residual of the observed sea surface temperature (SST) equation as stochastic forcing and the land's moisture flux raise this step by step and bring El Nino and its teleconnections, leaving the mean climate almost unchanged. The full model simulates about 1,000 years per day on one graphics processing unit, holds the observed climate for a thousand years without drift, reproduces the observed 1979-2022 global warming, carries forced change with a physically based structure around the frozen network (transient climate response 1.60 K), routes runoff to 212 river gauges with a median seasonal-cycle correlation of 0.79 and, driven by emissions, separates the heatwave exposure of an SSP1 and an SSP3 world by a factor of 4.7 in 2081-2100, three-quarters of it through the warmer climate.
Generative agents are game characters whose plans, conversations, and memories are written on the fly by a large language model (LLM). The original Generative Agents demonstration used cloud inference in a custom simulation. We present GARP, a platform for real-time generative agents inside Godot, and describe the Character Behavior System used in GARP 2. The complete loop runs on local hardware using a 3B-parameter open-weight model and no cloud service. Alongside the original planning, memory, and reflection, GARP gives each agent an emotion system driven by LLM appraisal, four graded personality axes drawn from the Big Five that set simulation parameters as well as prompt text, six needs whose urgency grows over time, and pairwise relationships updated after every conversation. The system narrows 50+ candidate actions to typically 3-8, constrains model responses with JSON schemas, and surfaces important memories in decision prompts. A layered optimization strategy (an adjustable game clock, decisions taken ahead of need, multi-tier caching, prompt compression, and fully asynchronous execution) keeps the game responsive while the model thinks. In a nine-scenario evaluation, 25 of 27 runs passed log-based checks. Both failures occurred in the birthday party scenario, where one guest omitted the event from their daily plan.