Abstract

Vision-Language-Action (VLA) models have demonstrated significant potential in robotic manipulation but suffer from a performance gap when comparing generalist models to task-specific specialists. Furthermore, deploying these models in unstructured indoor environments requires robust mapping and navigation capabilities. In order to close this gap, this paper presents a unified framework that combines a novel neuro-symbolic task configuration engine with a modular semantic mapping system. In order to make navigation and grounding easier, our method first builds a time-constrained semantic voxel map of the surroundings. By using relaxed waypoint algorithms, we optimize exploration efficiency by 50%. When a task configuration framework arrives at a workspace, it analyzes the spatial arrangement using zero-shot object detection and dynamic scene graph generation. 78-100% of the performance difference between generalist and specialist approaches is recovered by this analysis, which directs the manipulation request to the best specialized model. Empirical findings show that our system maintains real-time computational feasibility appropriate for practical implementation while increasing success rates from 74% to 98% in challenging long-horizon tasks.

Abstract

Robotic manipulation in unstructured environments requires accurate perception, pose estimation, and semantic understanding of object affordances. While vision-language models enable zero-shot detection, they lack structured reasoning for functional grasping. Geometry-based grasp synthesis ignores mass distribution and performs only static collision checking, causing failures in cluttered scenarios. We present the first end-to-end neurosymbolic framework integrating zero-shot perception, dynamic scene graph generation, 6-DoF pose estimation with temporal tracking, and grasp synthesis with symbolic reasoning. Key novelties are: (1) scene graphs encoding collision risk scores for real-time execution monitoring, (2) center-of-mass aware grasp optimization prioritizing mechanically stable grasps, and (3) automatic anchor generation for few-shot pose estimation eliminating manual intervention. In simulation, on a Franka Panda robot with 2-finger parallel-jaw gripper, we achieve 100% success on most reasoning-intensive tasks versus 0-50% for pure neural methods, 17 objects detected versus 5-13 for baselines, 99.97% mean ADD-S versus 23-98.7% for state-of-the-art on Any6D benchmarks, 95% success with 2% collision rate versus 45% success with 38% collision for geometry-only methods, and 98% grasp stability with 42% grip force reduction through center-of-mass optimization, demonstrating the superiority of neurosymbolic integration for robust robotic manipulation.

Abstract

A DeepSet-based regression model is employed to predict the superconducting critical temperature (Tc) of potential inorganic superconducting compounds. The model is trained on the SuperCon database using a set of 23 descriptors, primarily derived from chemical composition, while explicitly preserving permutation invariance to avoid artefacts arising from arbitrary ordering of elements. Model performance is assessed on the independent Konno COD candidate dataset, as well as on a collection of superconducting compounds reported in recent years. The model exhibits strong predictive accuracy for a substantial fraction of these newly reported materials, demonstrating its ability to generalize beyond the training set. Furthermore, SHAP value analysis is used to interpret the model predictions and to identify the elemental and compositional features that most strongly influence . These results underscore the effectiveness of DeepSet architectures in superconductivity-focused materials informatics and highlight the value of explainable machine learning for guiding the discovery of new superconductors.

Abstract

Conspiracy believers use significantly more
psycholinguistic markers per post than non-
believers (Cohen's d = 0.56, p < 10−80),
a pattern we term narrative density, sug-
gesting that belief manifests as structurally
denser conspiratorial frames distributed across
the full text rather than concentrated in spe-
cific lexical cues. We present Truth Gradi-
ent's system for SemEval-2026 Task 10 Sub-
task 2 (Samory et al., 2026): a DeBERTa-
v3-large model with mean pooling and a 5-
seed probability-averaging ensemble achiev-
ing macro F1 = 0.829 on the 77-sample de-
velopment set and 0.75 on the official test
set. The 5-fold CV estimate (0.734 ± 0.007)
proves the more reliable predictor of test perfor-
mance, and we recommend it as standard prac-
tice for low-resource shared tasks. Two con-
vergent tests support the narrative density ac-
count: masking annotated marker spans drops
F1 by 5.3 pp, and direct marker-count fusion
recovers +0.9 pp, though we note these are
not conclusive given the small dev set. Cross-
validated ablation identifies encoder fine-tuning
as the dominant design factor (−7.2 pts), and
layer-wise probing confirms belief informa-
tion peaks at mid-stack layers (layer 16/24).

Abstract

We present a system for the MEDIQA-SYNUR 2026 shared task on extracting structured clinical observations from nurse dictation transcripts. The transcripts contain spoken-style clinical language with disfluencies, filler words, and hesitations. Our approach is a four-stage LLM inference pipeline preceded by an offline knowledge enhancement step: (1) knowledge-enhanced concept detection using medical domain clustering, (2) evidence-grounded value extraction, (3) schema-constrained value normalization, and (4) deterministic post-processing with fuzzy matching and unit pairing. In the offline step, we use the task ontology and training examples to generate per-concept clinical definitions and extraction rules, and group the 193 concepts into 19 non-exclusive medical domain clusters. These are injected into all downstream prompts as domain priors. All LLM stages use gpt-oss-120b with structured JSON output and chain-of-thought reasoning. The task requires exact matching on concept ID and value pairs across a 193-concept ontology, making precision particularly challenging. We iteratively refine concept definitions and prompt guidelines based on error analysis of the training data. Our system achieves an F1 score of 0.806 on the test set.

Abstract

Video Situation Recognition (VidSitu) addresses the challenging problem of "who did what to whom, with what, how, and where" in a video. It tests thorough video understanding by requiring identification of salient actions and associated short descriptions for event roles across multiple events. Grounding with VidSitu requires spatio-temporal localization of key entities across shots and varied appearances.
We posit that coherent video understanding requires consistent identification of entities that play different roles. We propose Multimodal Entity Coreference (MEC) to unite entity descriptions in text with grounding across the video. Towards this, we introduce CineMEC, a multi-stage approach that unites event role mention groups with visual clusters of entities, without explicit grounding supervision during training. Our approach is designed to exploit the synergy between visual grounding and captioning, where improving one influences the other and vice versa. For evaluation, we extend the VidSitu dataset with grounding annotations. While previous work focuses primarily on descriptions, CineMEC improves consistency across both: captioning (+2.5% CIDEr, +7% LEA) and visual grounding (+18% HOTA).

Abstract

We describe our system for SemEval-2026
Task 10, Subtask 1: Conspiracy Marker Extraction (Samory et al., 2026), which involves
identifying spans of five marker types (Actor,
Action, Effect, Evidence, and Victim) in English social media text. Standard token-level
classifiers predict each position independently,
which can cause span fragmentation and ignores inter-marker dependencies. We present
a Span-Consistency Network (SCN) built on
DeBERTa-v3-large designed to address these
issues. First, a Span Consistency Layer (SCL)
propagates span-level signals to constituent tokens via a min-over-span, max-over-position
mechanism. Second, Cross-Marker Attention
(CMA) models co-occurrence patterns between
marker types through a learned correlation matrix. Third, we adopt a recall-biased Tversky loss augmented with Span Count Regularization (SCR), which penalizes mismatches
between predicted and gold span counts, penalizing diffuse probability distributions in favor of discrete spans. We ensemble five models trained via stratified cross-validation with
probability averaging. Our system achieved
a Macro F1 of 0.24 on the test set, placing
second among participating teams. Ablation
experiments on the development set show that
SCR has a large effect: its removal leads to prediction collapse for three of five marker types,
and that the Tversky loss outperforms binary
cross-entropy. We additionally provide a latency analysis of CMA, an analysis of how
SCL handles cross-type overlapping spans, and
an ablation on SCR warm-up scheduling.

Abstract

Chest X-ray (CXR) segmentation is an important step in computer-aided diagnosis, yet deploying large foundation models in clinical settings remains challenging due to computational constraints. We propose AdaLoRA-QAT, a two-stage fine-tuning framework that combines adaptive low-rank encoder adaptation with full quantization-aware training. Adaptive rank allocation improves parameter efficiency, while selective mixed-precision INT8 quantization preserves structural fidelity crucial for clinical reliability. Evaluated across large-scale CXR datasets, AdaLoRA-QAT achieves 95.6% Dice, matching full-precision SAM decoder fine-tuning while reducing trainable parameters by 16.6× and yielding 2.24× model compression. A Wilcoxon signed-rank test confirms that quantization does not significantly degrade segmentation accuracy. These results demonstrate that AdaLoRA-QAT effectively balances accuracy, efficiency, and structural trustworthiness, enabling compact and deployable foundation models for medical image segmentation.

Abstract

We propose MA-SafeDiffuser, a multi-agent extension of SafeDiffuser that equips diffusion-based trajectory planners with finite-time diffusion invariance guarantees for joint-agent safety specifications. Building on the single-agent construction that embeds control barrier function (CBF) constraints into reverse-diffusion updates, we: (i) formalize joint safe sets as intersections of per-agent and pairwise barrier sets; (ii) derive centralized and decentralized (communication-aware) constrained denoising procedures with provable invariance under mild assumptions; (iii) address local-trap and deadlock phenomena via time-varying specifications and liveness CBFs; and (iv) develop a lightweight benchmarking suite including a multi-agent Maze2D domain. Empirically, MA-SafeDiffuser reduces violation counts relative to unconstrained diffusion baselines, while retaining planning quality. We provide algorithms, proofs, and reference implementations.

Abstract

Radio frequency (RF) sensing is a fast and efficient technique for biomolecule detection using small sample volumes. Interdigitated capacitor (IDC) sensors are widely used in biosensing due to their compactness, ease of fabrication, and high quality factor. However, the design of the IDC sensor and the frequency tuning range (FTR) of the Vector Network Analyzer (VNA) are interdependent and must be considered jointly. In this work, we present a methodology to determine optimal IDC dimensions for detecting biomolecules based on dielectric constants for a given VNA, or conversely, to identify the required VNA frequency range for a given IDC-based biosensor. For breast cancer cell detection, the lowest calculated frequency range was 644.2-859.5 MHz (FTR: 28.64%), while the minimum FTR of 22.09% occurred at 4.23-5.28 GHz. Theoretically obtained results are compared with Ansys HFSS simulations and the impact of various IDC parameters on resonant frequency is analyzed.