Masked latent prediction methods such as I-JEPA aim to learn predictable features, but what determines which information they retain is less clear. We built a simulator in which every factor behind an image is known, so a probe can measure how readable each factor remains after training. We trained small encoders (0.7M parameters) with I-JEPA, SimCLR and BYOL on the same pairs of views and probed a factor that both views share. I-JEPA had a predictor with half the encoder's width and depth, the proportions it uses for ViT-B. Its median score was about 40% of SimCLR's. Widening the predictor to the encoder's width reversed this ordering at every training seed, against both other methods. We release the simulator and the trained models.
LLM agents increasingly rely on external tools exposed via the Model Context Protocol (MCP), yet current benchmarks assume tools respond correctly and on time, leaving agent resilience to realistic tool failures untested. We introduce MCP-Injector, a drop-in, protocol-aware fault-injection middleware that interposes on the MCP tool-call path with controlled persistence (permanent, transient, intermittent) and seedable schedules. Its fault model is calibrated to the observed incident distribution (failure mode x persistence) of the public MCP ecosystem, mined from 57,240 issue and pull-request threads across 1,774 open-source MCP repositories. In a 1,350-cell experiment crossing two backbone sizes (20B, 120B), three agent stacks (custom multi-round plan-and-execute, OpenAI Agents SDK, LangChain), and three fault injection conditions, we find that faults drawn from this permanent-heavy stress profile cause severe quality degradation (task-completion score falls by 60%, from 8.08 to 3.23 out of 10), and that scale does not purchase fault tolerance for the two gpt-oss backbones we test: the 120B retains no larger share of its fault-free score under faults (resilience-ratio model, p = 0.96), and its completion rate under faults is, if anything, lower (91.1% vs. 95.6%; difference 4.4 points, 95% CI [-0.2, 9.3]). Agent stacks differ in how well they absorb faults, which suggests that framework-level error handling matters more than model scale. These results demonstrate that capability benchmarks alone can mask operational brittleness, motivating protocol-aware fault injection as a complementary evaluation dimension for tool-calling agents.
We study squared-error estimation when a nuisance parameter is unidentified on a submodel but affects a target that vanishes there. For models with higher-order profile expansions, we give sufficient conditions under which the largest critical-path minimax risk is the exact asymptotic minimax risk over the full parameter space. An efficient observable feature, negligible calibration error and stable recovery yield one estimator attaining the leading constant uniformly, with the profile coordinates unknown. The estimator uses calibration estimates, a local feature mean and an empirical recovery feature. The constant is expressed through a bounded-normal decision problem. We verify the conditions in two models. In centered Gaussian logistic errors-in-variables regression with one surrogate and two unknown variances, we determine the fifth-order global coefficient risk and its exact constant. A sparse extension separates variance-decomposition uncertainty from direction estimation and specifies when the scalar constant is preserved. In a degree-balanced replicated-network model, we determine the sixth-order global edge-contrast risk when block proportions are unknown. Triangles identify a parameter combination; a bowtie supplies the additional information needed for the contrast.
This paper proposes incorporating channel configuration (CFG) information into Radio Foundation Model (RFM) pre-training and demonstrates that the resulting CFG-enriched embeddings serve as effective inductive priors for downstream tasks under label scarcity. We compare the proposed embeddings with raw channel tokens and multiple baseline models, under three different data regimes, for multiple downstream tasks --- Line-of-Sight/Non-Line-of-Sight (LoS/NLoS) classification, timing advance (TA) estimation, and beam selection. Evaluation is done in both in-distribution and out-of-distribution configuration settings. Experimental outcomes show that the proposed embeddings consistently outperform and/or are competitive to the baselines, across experiments in both seen and unseen configuration scenarios.
Low Earth Orbit (LEO) satellite mega-constellations such as Starlink require frequent handovers due to high orbital velocities, resulting in substantial Round-Trip Time (RTT) fluctuations. Accurate RTT prediction can mitigate these fluctuations, yet it fundamentally relies on continuously updated serving- satellite information. Existing serving-satellite identification methods typically complete identification only at the end of each scheduling cycle, creating an identification blind zone that limits real-time prediction. To overcome this limitation, we develop StarRTT, an end-to-end real-time framework that eliminates the blind zone by integrating per-second serving-satellite identification with RTT prediction. Within this framework, we propose Real-Time Long Short-Term Memory (RT-LSTM), an LSTM-based RTT prediction model designed to capture both non-stationary RTT evolution and time-varying geometric relationships among the serving satellite, user terminal, and gateways. Specifically, a historical LSTM encodes historical RTT observations and real-time serving-satellite and gateway-relative features, while a future LSTM processes Two-Line Element (TLE)-propagated future features of the identified serving satellite up to the next scheduling instant. Under strict real-time back-testing, RT-LSTM achieves a Mean Absolute Error (MAE) of 1.405 ms averaged over all 15 prediction horizons and 15 reference time-slot indices, reducing the error by 4.3 %, 11.6 %, and 22.1 % compared with the existing baseline models, LSTM, StarNet, and Attention-based Spatiotemporal Model (ASTM), respectively. Notably, the reduction over LSTM reaches 12.5 % at the scheduling instant, highlighting the benefit of real-time serving-satellite information when recent RTT observations are limited.
GUI grounding benchmarks ask whether an agent can localize a target that is already visible in a static screenshot. Real interfaces often violate this assumption: a target may be off-screen, appear only after a state change or hover, remain occluded, or become actionable after a delay. This leaves the ability to recover target visibility unevaluated. To measure this reveal step, we introduce Interactive GUI Grounding (IGG), a benchmark in which an agent must first make the target visible through interaction and then click it. IGG contains 601 tasks across 16 controllable GUI environments and seven visibility-constraint sub-types, with a compact action space of scrolling, moving, dragging, waiting, and clicking. In our human study, participants achieve 94.6% accuracy, while GPT-5.5 reaches 79.4% and the strongest evaluated open-source VLM reaches only 31.6%. To identify the source of this gap, we conduct two controlled interventions. Telling Qwen3-VL which interaction to perform raises its accuracy only from 30.8% to 33.9%, whereas performing the interaction for the model raises it to 81.9%. The bottleneck is executing the reveal interaction, not recognizing the target once it is exposed. We further show that this capability is learnable, demonstrating that visibility recovery can be strengthened consistently across model families.
Retro video games are everywhere lately - especially in places where they aren't supposed to be. Halo in a web browser, Wind Waker on the…
Learned representations often encode sensitive attributes, such as gender or race, that may be undesirable or unsafe to retain. Concept erasure aims to remove this information while preserving task-relevant content. Existing methods remain incomplete: linear methods provide guarantees only against linear probes, while nonlinear methods achieve stronger empirical erasure but lack formal guarantees and often sacrifice utility. In this work, we introduce a principled conditional autoencoder framework for nonlinear concept erasure. We follow the principle that erasure should remove only the sensitive attribute and nothing else, and we show that, under the stated assumptions, perfect reconstruction together with population-level independence is sufficient for perfect nonlinear guardedness, with preservation guarantees in the imperfect-reconstruction regime. Empirically, we attain a superior utility-erasure trade-off across multiple datasets, sensitive attributes, and language models, surpassing even the supervised variants of competing methods despite using no target-task labels, while being simpler to optimize and converging substantially faster. Beyond erasure, our framework yields concept conversion (representation steering) as a by-product: by changing only the target concept at inference time, it produces representations corresponding to different concept values - at no additional training cost - a capability no prior erasure method provides, and on which we likewise demonstrate strong empirical performance.
This paper examines speech characteristics of the Austrian National Council members (CMs). We introduce analysis methodology based on a fine-tuned BERT-Model which assigns to every speech a probability distribution over Austrian parties. These are then aggregated at the speaker and party level. We perform an extensive evaluation to ensure the quality of the generated distributions prior to further analysis. Instead of considering mis-assignments of predicted parties as a performance issue, we leverage them to derive insights about the CMs and political developments. Our findings show that parties that govern together become more similar in their speech characteristics, while, analogously, parties that are in opposition move further away from governing parties. Moreover, CMs who switch party usually adapt their speech characteristics only after the switch occurs. Finally, we publicly release a data set containing party affiliation predictions for the speeches of the Austrian National Council from 1996 to 2018.
Thompson sampling is a widely used strategy for sequential decision-making under uncertainty, offering a principled way to balance exploration and exploitation. This strategy involves maximizing a target objective under a single sample drawn from the current belief over unknown quantities. Conventionally, this optimization is performed once the belief has been inferred from data, and it becomes computationally burdensome when many Thompson samples are required, especially in continuous action spaces. In this work, we introduce an approach that amortizes this process ahead of time, before any data is collected, so that an approximation of the full Thompson sampling distribution is available immediately once observations arrive. We employ this approach to solve the multi-round setting, involving sequential data acquisition and Thompson-sample recommendation across rounds. Beyond accelerating Thompson sampling, our approach enables efficient estimation of expected information gain about the optimal action. This can be used to select more informative Thompson samples or to obtain maximally informative designs with optimization. Combining this estimate of expected information gain with immediate regret supports information-directed sampling (IDS), which can outperform Thompson sampling by explicitly balancing these two quantities. We demonstrate our approach on two simulator-based design problems in synthetic biology, where the simulators have intractable likelihoods.
Resources for toxicity detection in Punjabi remain limited, particularly for musical content that combines linguistic and multimodal information. We present the Punjabi Geet Toxicity Dataset (PGT-FG), a segment-level resource developed to support research on toxicity detection in Punjabi musical content. The dataset was constructed from publicly available Punjabi songs collected from YouTube. An initial collection of 620 songs resulted in 465 successfully downloaded songs and 3,402 preliminary segments from 450 songs. After quality control and annotation, the final dataset contains 2,396 unique segments from 392 songs. Four native Punjabi annotators produced 2,870 annotation records, including a shared subset of 158 segments annotated by all four. Each annotation includes binary toxicity, eight fine-grained toxicity categories, severity, sentiment, mode, and an uncertainty flag. Inter-annotator agreement was measured with Fleiss' kappa on the shared subset, and majority-based resolution produced the final segment-level labels. Of the 2,396 segments, 705 (29.42%) were labelled toxic and 1,691 (70.58%) non-toxic. Violence was the most frequent category, occurring in 539 toxic segments (76.45%), followed by gang references (227, 32.20%) and drug references (133, 18.87%). The overall toxicity label achieved substantial agreement (kappa = 0.6521). PGT-FG provides a fine-grained resource for Punjabi toxicity detection and multimodal content analysis. The resolved dataset, with one final record per segment, is publicly available on Zenodo (https://doi.org/10.5281/zenodo.23040156). Original lyrics, lyric excerpts, and copyrighted audio or video content are not redistributed.
We analyze HTML content and hyperlinks from an offline Tor crawl containing 383 onion hosts and 2,067 directed cross-host links. The study combines semantic analysis with degree-controlled quantum--classical link prediction. Two directory groups emit 94.8\% of the observed host links. Directed preferential attachment reaches ROC-AUC 0.976 with random non-edges, but falls to 0.5500.014 under source-conditioned target-degree matching. With disjoint training and test pairs, multimodal logistic regression reaches 0.6720.021, GAE 0.8990.016, and VGAE 0.9060.016. A four-qubit variational quantum classifier (VQC) comparison tests two encodings across 20 disjoint-pair graph splits under exact and 1,024-shot simulation, with classical controls using the same inputs and training examples. Its best exact mean ROC-AUC is 0.5490.023, compared with 0.5510.016 for RBF SVC and 0.5730.029 for a 16-parameter MLP on the same inputs and training examples. These results support predictive signal in combined graph/content features and show why degree control, feature compression, and optimization conditions matter in quantum--classical comparisons.
Spectroscopic instruments with limited wavenumber coverage require accurate scan positioning or multiple measurements to complete the material analysis. This often requires prerequisite knowledge about the material under diagnosis to measure the spectral regions with distinguishing peaks. In our study, we simulate a Bayesian Adaptive Acquisition with Reconstruction-driven Classification (BAARC) framework for Raman spectral measurements, using masked training to learn spectral features that dynamically drive laser channel selection until the material is identified. We present a jointly trained two-level framework that takes progressive partial measurements of small wavenumber patches across the complete spectral range and, at each step, first attempts to reconstruct the full spectra until the model uncertainty is low. Secondly, material classification is performed at each reconstruction step until stoppage. Uncertainty quantification during reconstruction steps helps determine the measurement position of our next patch, while combining it with classification uncertainty determines measurement termination. By closing the loop between hardware acquisition and deep probabilistic modeling, the proposed system minimizes physical measurement steps while maintaining both spectral fidelity and diagnostic accuracy. Experimental evaluations demonstrate that our framework achieves high-accuracy sample identification using only a fraction of full-spectrum channels, offering a scalable solution for real-time, high-throughput spectroscopic instruments.
Single-stage fine-tuning is pervasive, yet catastrophic forgetting persists when the source pre-training data is inaccessible. Data weighting affords a natural lever to modulate sample influence. Yet existing weighting methods lack a principled mechanism to explicitly arbitrate knowledge retention against new-task adaptation, and provide no unified differentiable objective to accommodate auxiliary signals such as sample diversity and sample characteristics. This is because current schemes privilege one goal over the other, and no explicit gradient-compatible encoding exists for such signals. To address this, we propose DRIVE, a principled and efficient weight assignment framework. It derives per-sample weights via a unified differentiable objective where three coupled signals co-evolve: Shapley additivity splits contribution into anti-forgetting and adaptation components, which are reshaped by a covariance-based diversity term and stabilized by a sample-characteristic prior. Far from a mere combination, this yields an end-to-end gradient-based scheme balancing stability against plasticity. Theoretically, we prove that DRIVE enjoys superior stability-plasticity guarantees over existing weighting schemes and enhances parameter estimation robustness. Empirically, it attains superior overall performance, faster convergence, and stronger robustness. Code at https://github.com/luciabcd/DRIVE.
