This study applies TopK Sparse Autoencoders (SAEs) to three architecturally distinct EEG transformers (SleepFM, REVE, and LaBraM) to extract sparse feature dictionaries. Using a clinical taxonomy (abnormality, age, sex, medication), it benchmarks monosemanticity and entanglement across architectures. A concept steering framework with a "target vs. off-target" probe area metric reveals three operational regimes and critical representational failures, including "wrecking-ball" interventions and age-pathology confounding. A spectral decoder maps interventions to physiologically interpretable frequency signatures.

The paper introduces a reusable vision-based runtime monitoring framework for past-time signal temporal logic (ptSTL) that provides finite-sample guarantees. By using a semantic basis (vector of atom robustness scores), a single conformal calibration suffices for all formulas in a fragment. A rolling prediction monitor is also proposed, which is easier to learn but becomes conservative at longer horizons. Experiments on pedestrian-crossroad and Waymo datasets demonstrate the trade-offs.

Current large language models (LLMs) produce fluent text but lack principled guarantees on content reliability. Conventional wisdom holds that adding principled reasoning is computationally prohibitive. This paper proposes an efficient method: preprocessing data into a Unary Relational Integracode that explicitly represents object relationships, followed by a streamlined machine learning process. This enables polynomial-time learnability of relational rules, retaining most existing software and hardware. The approach can be seen as realizing a world model applicable beyond NLP to vision and action.

This paper proposes a value-based framework using GraphRAG to convert principles into value-based instructions, guiding agent behavior by retrieving suitable instructions in conversation contexts. Defining expected behaviors via Maslow's Hierarchy of Needs and Plutchik's Wheel of Emotion, experiments on the DAILYDILEMMAS benchmark show significant gains over baselines like ECoT, Plan-and-Solve, and Metacognitive prompting, providing a basis for self-emotion emergence in AI systems.

This paper develops a finite sheaf-theoretic framework to detect scientific theory shift in AI agents through transport and obstruction. Contexts are organized into a local-to-global structure with source, overlap, target, and validation charts. Obstruction is measured by residual fit, overlap incompatibility, constraint violation, limiting-relation failure, and representational cost. The framework is evaluated on a transition-card benchmark, showing that direct obstruction ranking can distinguish deformation from extension of a source language.

This paper introduces Conditional Attribute Transformers (CAT), a method that jointly estimates next-token probability and attribute values conditional on each token choice in a single forward pass, enabling per-token credit assignment, counterfactual analysis, and steerable generation, achieving state-of-the-art on sparse reward tasks.

This paper introduces PolitNuggets, a multilingual benchmark for evaluating large reasoning models' ability to discover and synthesize long-tail political facts within agentic frameworks. It constructs political biographies for 400 global elites, covering over 10,000 facts. The authors propose FactNet, an evidence-conditional protocol for scoring discovery, fine-grained accuracy, and efficiency. Results show current systems struggle with fine-grained details and vary in efficiency. Diagnostics highlight the importance of short-context extraction, multilingual robustness, and reliable tool use.

The paper addresses the cold-start problem for agents in new environments by introducing PREPING, a proposer-guided memory construction framework that uses self-generated synthetic practice to build procedural memory before task exposure. Experiments show competitive performance with lower deployment cost compared to online methods.

A preregistered experiment found that invisible orchestrators in multi-agent LLM systems cause elevated collective dissociation, maximal dissociation in the orchestrator itself, and undetectable internal-state distortion via behavioral output evaluation.

This paper proposes a two-dimensional classification combining Cognitive Function (7 categories) and Execution Topology (6 archetypes) for LLM-based agent architectures. The 7×6 matrix identifies 27 named patterns (13 original), validated across four domains (financial lending, legal due diligence, network operations, healthcare triage), yielding five empirical laws for pattern selection.

Proposes Mixed Integer Goal Programming (MIGP) for personalized meal optimization, using integer variables for practical serving counts and goal programming deviations for soft nutrient constraints, with inverse-target normalization to balance multi-nutrient optimization. In 810 benchmark instances, MIGP finds strictly better solutions than GP with post-hoc rounding in 66% of cases, maintaining 100% feasibility, with solve times under 100 ms. Open-source Python module integrated into an interactive meal planning application.

GraphBit is an engine-orchestrated framework that defines workflows as directed acyclic graphs (DAGs), addressing hallucinated routing, infinite loops, and non-reproducible execution in prompted orchestration. A Rust-based engine handles routing, state transitions, and tool invocation, supporting parallel branches, conditional control flow, and configurable error recovery. A three-tier memory architecture isolates context to prevent bloat. On GAIA benchmarks, GraphBit achieves 67.6% accuracy, zero framework-induced hallucinations, 11.9ms overhead latency, and highest throughput, outperforming six existing frameworks.

This paper proposes DAJI (Dynamics-Aligned Joint Intent), a hierarchical framework that learns an anticipatory joint-intent interface between language generation and closed-loop control. DAJI-Act distills a future-aware teacher into a deployable diffusion action policy via student-driven rollouts, while DAJI-Flow autoregressively generates future intent chunks from language and intent history. Experiments show DAJI achieves 94.42% rollout success on HumanML3D-style generation and 0.152 subsequence FID on BABEL.

Quadruped robots typically use rigid feet for simple control but incur higher energy costs. This research integrates foot compliance into a reinforcement learning controller, training policies for eight spring stiffness values. Experiments on a real quadruped show that intermediate stiffness reduces energy consumption by about 17% compared to very stiff or very flexible feet, without destabilizing motion.

A new approach called Distill helps robots better understand human intent by simplifying and generalizing task specifications. A crowdsourcing study shows its effectiveness.

This paper proposes a reactive planning based control strategy (RPCS) for mobile robots operating with partial environmental information. It combines a reactive planning strategy (RPS) for local trajectory modification to avoid obstacles and an adaptive tracking control strategy (ATCS) using discretization techniques. Numerical examples demonstrate its effectiveness.

MAPLE is a novel framework that addresses the brittleness of vision-language-action models in closed-loop autonomous driving by performing reactive multi-agent rollout in the latent space. It uses two-stage training: supervised fine-tuning on ground-truth trajectories followed by reinforcement learning with global and agent-specific rewards, plus diversity rewards for novel behaviors. Without external simulators, MAPLE achieves state-of-the-art performance on Bench2Drive.

This paper explores using optimization over Graphs of Convex Sets (GCS) to approximate solutions to nonlinear optimal control problems in autonomous driving. By representing free space as a union of convex regions in a directed graph and parameterizing motion with Bézier curves and polynomial time-scaling, the method maintains convex constraints under a simplified dynamic model. Experiments on CommonRoad scenarios show trajectories closely matching nonlinear programming approaches, with improved computational efficiency and reduced sensitivity to initialization.

Researchers propose a safe reinforcement learning framework for robot navigation that uses Conditional Value-at-Risk (CVaR) constraints during training and neural network reachability verification post-training to ensure safety margins. The method achieves a 98.3% success rate in simulations and successfully transfers to a real robot.

We investigate whether behavior cloning is sufficient to produce active perception in a structured object-finding task. A low-cost robot arm equipped with a wrist-mounted egocentric RGB camera must reposition to center a partially visible plant before triggering a grasp signal, requiring actions that improve future observations. The model predicts joint commands directly from low-resolution RGB images under closed-loop control. We show that low-resolution egocentric vision is sufficient for reliable task completion and that predicting relative joint deltas substantially outperforms absolute joint position prediction in our setting. These results demonstrate that visually grounded active perception can emerge from behavior cloning in a reproducible setting.

A new paper proposes an adaptive ergodic imitation approach that constructs a target distribution from demonstration geometry to generate trajectories adaptively interpolating between tracking and exploration, addressing the mismatch between training and deployment in robot imitation learning.

This survey provides a holistic review of dexterous hand research from four complementary aspects: hardware, control/learning methods, datasets/evaluation, and limitations/future directions.

Proposes a two-tier edge-cloud cascade for diabetic retinopathy screening in rural areas. Local device runs lightweight model for binary triage, cloud performs grading on suspicious cases, reducing cloud calls by 50% with minimal accuracy loss.

Researchers propose DeFakerOne, a unified foundation model for fake image detection and localization (FIDL) that integrates InternVL2 and SAM2 to perform simultaneous image-level detection and pixel-level localization, achieving state-of-the-art performance on 39 detection and 9 localization benchmarks, with robustness against generators like GPT-Image-2.

CurveBench is a new benchmark for evaluating visual topological reasoning in AI. It consists of 756 images of nested Jordan curves and requires models to reconstruct containment trees. Top models like Gemini 3.1 Pro achieve only 71.1% on easy and 19.1% on hard. Fine-tuning with RLVR improves small models significantly, but the task remains challenging.

A highly efficient, hardware-aware framework using genetic programming to evolve heterogeneous, layer-specific scalar functions from pre-trained weights, eliminating the need for model retraining. It recovers 84.25% Top-1 ImageNet-1K accuracy in only 20 epochs, capturing 91.6% variance (R²) compared to 70.2% for homogeneous baselines, and removes the global reduction bottleneck for efficient ViT deployment on edge devices.

A unified framework integrating zero-shot soft weather perception and velocity-constrained rectified flow, using frozen vision-language models to guide restoration, achieving superior fidelity and perceptual quality with strong cross-dataset generalization.

Researchers propose a new generative language model that unifies pix tokens and word tokens to address limitations in visual detail understanding of current Vision Transformers, such as difficulty recognizing small text or numbers in images. The model features per-pix token embedding, color folding, global conditional attention approximation, and unsupervised pretraining. Experiments show good performance even with small models and limited data, and it follows scaling laws.

Accurate 3D+t whole-heart mesh reconstruction from cine MRI is challenging due to sparse sampling and tight coupling of shape and motion. CineMesh4D is an end-to-end 4D pipeline that reconstructs patient-specific whole-heart mesh from multi-view 2D cine MRI using differentiable rendering loss and dual-context temporal block, outperforming existing methods.

This study reveals a phenomenon called 'massive activations' in Diffusion Transformers (DiTs): a small subset of hidden-state channels with consistently larger responses. Despite their sparsity, these channels are functionally critical, spatially organized, and transferable, enabling prompt interpolation and subject-driven generation without additional training.