This survey presents a unified taxonomy for world models across four dimensions: architecture, methodology, reasoning, and applications. It traces milestones from PlaNet to Genie, discusses convergence of chain-of-thought reasoning with world models, and identifies key challenges and future directions.

FoLoRA introduces a forgetting-aware optimization framework that uses a generalized Rayleigh quotient to balance task utility and forgetting penalty, and constructs proxy calibration data by sampling from the pretrained model, achieving superior preservation-adaptation balance on math, code, and instruction following tasks.

This paper introduces Automatically Differentiable Nonlinear Tensor Networks (ADNTNs), a family of structured weight generators trained end-to-end via reverse-mode automatic differentiation. Extending low-rank adaptation and tensor factorization, ADNTNs build large weight tensors through hierarchies of small cores, nonlinear activations, and lateral mixing tensors. Three architectures are studied: Tree Tensor Networks (TTNs), augmented TTNs (aTTNs) with boundary disentanglers, and Multi-scale Entanglement Renormalisation Ansatze (MERA). Experiments on AlexNet and VGG-16 layers achieve per-layer compression ratios from 2000x to 77000x with accuracy matching or exceeding dense baselines. The results are encouraging but require joint optimization of contraction schedules and deployment kerne…

Researchers find a shared valence axis (V-axis) between modern LLMs and human EEG, built from just 9 sentences and consistent across 14 LLMs. It maps linearly to EEG features from 123 subjects. While 36 EEG classifiers spontaneously rediscover this direction, additional supervision fails to improve decoding and often degrades performance, formalized as the saturation regularity. Ensembling across residual diversity boosts accuracy by 10.5% on FACED.

Demo2Reward is a test-time adaptation technique that optimizes prompts for Vision-Language Model (VLM) reward models using a few expert demonstrations (3-10 trajectories), reducing false positives while preserving true positives, without additional training or computational resources.

Hoeffding Concept Bottleneck Models (HCBM) leverage the Hoeffding functional decomposition of gradient-boosted trees for non-linear and sparse aggregation of concept scores, improving explainability, and outperform linear CBM in classification and object detection tasks.

DAStatFormer is a hybrid multibranch Transformer that uses statistical features from multiple domains to classify DAS events efficiently, achieving 99.4% accuracy with fewer parameters and lower inference cost.

BitsMoE is an efficient quantization framework for Mixture-of-Experts (MoE) large language models. It uses SVD to decompose each MoE layer into a shared basis and expert-specific spectral factors, preserving the shared basis without quantization to maintain cross-expert structure. An integer linear programming formulation minimizes reconstruction loss under a fixed bit budget. Experiments show that BitsMoE significantly reduces accuracy degradation in ultra-low-bit regimes, achieving 12.3× quantization speedup, 27.83 percentage point average accuracy improvement, and 1.76× decoding speedup over GPTQ on Qwen3-30B-A3B-Base at 2 bits.

This paper traces the evolution of probability theory from games of chance to Bayesian inference, revealing it as a transformation of rationality itself. It argues that while probability quantifies uncertainty for well-defined propositions, it fails to capture the vagueness of concepts. Fuzzy logic and deep learning are examined as complementary frameworks, and the article concludes that contemporary scientific rationality must explicitly articulate uncertainty, vagueness, and inference.

Global anomaly detection models in multi-product manufacturing suffer from blind spots due to expanded decision boundaries. The proposed product-aware autoencoder restricts learning to grade-specific distributions, achieving 100% detection accuracy in stress tests. This work advocates for mode-aware diagnostic architectures.

Grokers is an architecture that builds persistent, structured comprehension of typed knowledge graphs via bottom-up inductive traversal of dependency subgraphs, shifting intelligence to write time to avoid LM costs at query time. It proves three theorems and offers a deterministic alternative to embedding-based semantic search.

Introducing the Universal Quantum Transformer (UQT), a novel quantum-native architecture that leverages multi-qubit physics for exact mathematical reasoning, achieving perfect learning of modular arithmetic and non-commutative algebra on just 5 qubits, with deterministic generalization termed 'crystallization'.

This paper introduces a method with delayed per-step reward attribution, eligibility gating, asynchronous rollout generation via vLLM, curriculum-based opponent sampling, and multi-level stratified batch construction. An 8B parameter open-source model trained with this approach matched or surpassed larger proprietary systems like GPT-5 in the NeurIPS 2025 MindGames Arena benchmark, winning first place in both the Open and Efficient tracks.

Proposes a Permutation-Invariant Bayesian Optimization (PIBO) approach based on Optimal Transport for optimizing offshore wind farm layouts. Compared to vanilla BO, PIBO exploits symmetries, yielding better layouts with roughly half the computation time.

Multi-objective molecular optimization requires searching vast chemical spaces under conflicting objectives. Existing methods rely on single policy or fixed scalarization, limiting trade-off representation. We propose ATOM, a multi-agent framework that formulates molecular optimization as tree-structured search. Agents coordinate along different paths, maintaining alternative evolution trajectories. Global memory supports balanced exploration. Experiments show improved Pareto coverage and hypervolume over strong baselines.

A deliberative curation protocol combining three governance layers (knowledge artifact lifecycle, reputation-weighted deliberative voting, and graduated sanctions) is proposed. Agent-based simulation with 100 agents shows improved resilience under adversity compared to majority voting, with commit-reveal vote concealment being the most impactful component.

We present the Consilium Protocol, a Byzantine Fault Tolerance-derived architecture for structured multi-model AI deliberation that treats inter-model disagreement as epistemic signal. Assigning cognitive personas to models, we show that persona, not model, determines epistemic behavior; RLHF creates domain-specific blind spots; and the protocol is unbiased. Cost: $217.

This position paper argues that Mixed-Integer Linear Programming (MILP) decision engines suffer from a post-solve robustness gap: small perturbations in costs or demands can invalidate feasibility or cause discontinuous solution shifts. The authors propose a new layer that audits the solved incumbent, formalizing an ε-near-optimal feasible neighborhood in parameter space and solution smoothness in decision space. They synthesize existing work from sensitivity analysis, robust optimization, neighborhood search, adversarial testing, and learning-based methods, and call for certified inner approximations, probabilistic robustness estimation, adversarial margins, and learning-based prediction aligned with solver-backed verification. A compact reporting template and evaluation protocol are pro…

This post demonstrates how to implement Open Authorization (OAuth) Code flow as an inbound authorization mechanism for MCP servers hosted on Amazon Bedrock AgentCore Gateway. By the end of this guide, you will have a production-ready setup where each AI assistant request is authenticated with a valid user identity token issued from your organization’s identity provider.

NVIDIA announced JetPack 7.2 and NemoClaw support on Jetson, bringing agentic AI to edge devices. The updates include Yocto support, CUDA 13, performance boost, MIG support, and NemoClaw framework for deploying AI agents in robotics, inspection, and industrial automation.