A daily curated roundup of Claude Code news, discussions, and links for developers.

Krea AI launches Krea 2, its first in-house foundation model focused on aesthetics and creative control, with features like style references and moodboards. User reviews are mixed, praising speed and interface but noting inconsistent quality and weak precision for small edits.

Eric Schmidt's commencement speech praising AI was met with boos, reflecting growing public skepticism.

Despite predictions that AI will render higher education obsolete, economic principles suggest AI may actually increase demand for certain forms of higher education by increasing the value of frontier knowledge, sophisticated tool use, and credible signaling.

Montage M1 lets AI agents generate interactive, persistent UIs with a single API call, reducing token usage and cost while enabling saved, brandable artifacts.

AnyFrame is the official Python SDK for the AnyFrame control plane, enabling developers to point an agent at a repo and get a sandbox running Claude Code. It provides comprehensive APIs for managing agents, sessions, skills, MCPs, connectors, builds, previews, and more, with both sync and async support.

This paper presents a method for training reference-guided, perceptive reinforcement learning locomotion policies for humanoid robots, where reference trajectories are modulated during training to be consistent with terrain geometry. By synthesizing SE(2)-controllable reference trajectories inside the RL training loop, projecting footsteps onto valid footholds and adjusting swing-foot and center-of-mass trajectories, the resulting policy provides a clean SE(2) velocity interface. Simulation shows improved tracking performance. On hardware, integrated with an MPC and control barrier function planner, the method enables long-horizon (>70m) autonomous navigation on the Unitree G1 through rough terrain and stairs, with onboard sensing and computation.

This paper presents a QUBO-based framework for robot design optimization using kinematic structure-level metrics. It combines classical computation with quantum annealing, using a robotic hand as a case study. The framework incorporates individual rewards, workspace interactions, one-hot constraints, and dependency penalties into a unified quadratic model. Simulated and quantum annealing verify feasibility, yielding valid design combinations.

This paper presents LAPS, a replay management framework that combines reliability-based active pooling and uncertainty-guided active sampling to address catastrophic forgetting in incremental neural distance field mapping, improving reconstruction completeness while maintaining geometric accuracy.

Researchers introduce FLASH Policy, a new visuomotor policy that uses Legendre polynomial trajectory representation and sparse temporal sampling to enable single-step inference, achieving state-of-the-art success rates and dramatically reducing inference time compared to diffusion and flow matching policies.

This study introduces intelligent frameworks that use Large Language Models (LLMs) to improve task scheduling for construction robots. The system utilizes a Natural Language Processing interface for communication and adapts in real-time to unexpected site conditions. It concurrently uses two LLM agents: a generator (GPT-4) and a supervisor (Gemma 3/Llama 4/Mistral 7b) to provide precise task schedules. Evaluation results highlight the crucial role of LLMs in construction robotic tasks.

Researchers propose a hybrid control framework, delay-resilient RL, combining an LSTM state estimator with a residual RL policy to address stochastic communication delays in teleoperation. Experiments on Franka Panda robots show superior stability and tracking accuracy over state-of-the-art baselines.

This study presents a vision-guided method for autonomous drones to select ideal perching spots on trees. Using image processing algorithms including machine learning, segmentation, and binary morphology, the method evaluates branch width, slope, and curvature rather than simply choosing the nearest branch. Tested on over 10,000 urban tree images, it succeeds for 76% of feasible targets. Future work will incorporate depth perception and attitude sensors.

This paper presents a diffusion-based visuomotor control policy for opening heavy self-closing doors and passing through. It learns an end-to-end policy that coordinates a nonholonomic base and dual arms, achieving high success rate and robustness. Traditional state-machine approaches lack generalization.

This report presents HoloMotion-1, a humanoid motion foundation model for zero-shot whole-body motion tracking. It scales control-policy training with a large-scale hybrid motion corpus that combines video-reconstructed motions for diversity with high-fidelity motion capture and in-house data for supervision. The model addresses challenges like reconstruction noise and domain mismatch through a sparse Mixture-of-Experts Transformer with KV-cache for real-time control and sequence-level training. Experiments show robust generalization, improved tracking accuracy, and direct transfer to a real humanoid robot without fine-tuning.

PhysBrain 1.0 is a new approach that transforms large-scale human egocentric video into structured physical commonsense knowledge to improve robot understanding and control. By extracting scene elements, spatial dynamics, and depth-aware relations from video, it trains vision-language models that are then adapted to robot policies. The method achieves state-of-the-art results on multiple embodied AI benchmarks, particularly excelling in out-of-domain tasks.

PanoWorld is a panoramic video world model that generates geometry-consistent 360° video from a single image and a caption. It enhances geometric constraints via depth consistency loss and trajectory consistency loss, and introduces the PanoGeo dataset, significantly improving geometric consistency while maintaining visual realism, catering to embodied AI spatial understanding.

ChangeFlow is a generative framework for remote sensing change detection, reformulating change mask synthesis as rectified flow in latent space. It uses a lightweight conditioning signal and stochastic design to support sampling-based ensembling. On four benchmarks, it achieves an average F1 of 80.4%, outperforming previous best by 1.3 points with comparable inference speed.

This paper proposes sampling in feature space for group-convolutional neural networks (GCNNs), replacing geometrically dense samples with representative samples selected by feature similarity to decouple geometric resolution from memory and processing costs. Empirical results show that coarse feature-space sampling preserves classification accuracy well, enabling substantial acceleration of equivariant 3D classifier training.

Minerva-Ego is a new benchmark for evaluating complex egocentric visual reasoning, featuring multi-step multimodal questions and densely annotated reasoning traces. Experiments show a large gap between state-of-the-art models and human performance, with significant improvements when providing 'where' and 'when' hints.

Reliable detection of humans beneath forest canopy remains difficult due to occlusion. This paper presents a multimodal pipeline integrating LiDAR evaluation, visible-thermal fusion, and synthetic-aperture imaging via AOS. A YOLOv5 detector fine-tuned on FLIR dataset achieves mAP ~0.83, establishing a baseline for UAV-based search-and-rescue and surveillance in forested environments.

COPRA is a novel framework for video anomaly detection that dynamically adapts a frozen vision-language model by generating input-specific parameter updates, addressing the mismatch between training and inference in data distribution and model configuration. It outperforms static baselines on standard benchmarks and generalizes to unseen tasks like video question answering and dense captioning.

This paper proposes Deep Pre-Alignment (DPA), a novel architecture that replaces the standard ViT encoder with a small VLM as perceiver, ensuring visual features are deeply aligned with the text space of the target large language model. On the 4B parameter scale, DPA outperforms baselines by 1.9 points across 8 multimodal benchmarks, with gains widening to 3.0 points at the 32B scale. DPA also achieves a 32.9% reduction in language capability forgetting and demonstrates consistent gains across different LLM families like Qwen3 and LLaMA 3.2.

ReactiveGWM is a reactive game world model that synthesizes dynamic interactions between player and NPC by decoupling player controls from NPC behaviors. It injects player actions via a lightweight additive bias into the diffusion backbone and grounds high-level NPC responses through cross-attention modules. These modules learn a game-agnostic representation of interactive logic, enabling zero-shot strategy transfer to off-the-shelf world models of different games without retraining. Evaluated on two Street Fighter games, it maintains fine-grain player controllability while achieving robust NPC strategy adherence.

Surprisal theory links word predictability to processing difficulty, but language model (LM) surprisals systematically underpredict human difficulty in syntactic ambiguities like garden path sentences. This study tests the Parse Multiplicity Mismatch Hypothesis, which suggests LMs can consider more simultaneous parses than humans. Using Recurrent Neural Network Grammars (RNNGs) with beam search, varying the number of simultaneous parses, they found that reducing parses increased predicted garden path effects but not enough to match human data, indicating this hypothesis alone cannot reconcile LM and human processing.

This paper analyzes neural activation patterns across six large language model architectures on twelve cognitive tasks, revealing that mathematical reasoning yields the highest attention entropy across all architectures, while decoder models exhibit significantly higher sparsity than encoders. The findings offer insights for model selection and optimization in big data applications.

Large language model personalization typically adapts outputs to user preferences and style but does not account for differences in user evaluation capacity across domains of expertise. This limitation can encourage Professional Domain Drift, where users rely on AI generated reasoning in domains they cannot reliably evaluate. We introduce Capability Conditioned Scaffolding, a typed framework that partitions expertise into strong, mixed, and weak domains and conditions intervention behavior on structured capability profiles. A pilot evaluation across multiple MMLU subsets and four LLM substrates shows consistent profile conditioned intervention behavior, including categorical inversion under profile swapping and selective activation in mixed domain risk zones. These findings suggest that c…

This study introduces Eskwai for Students, a retrieval-augmented generation (RAG) system designed to assist Ghanaian law students with legal education. Leveraging a curated database of over 12K case laws and 1.4K legislation, the system was deployed in a 30-month longitudinal study with 3.1K students making 32K queries. The evaluation assessed its helpfulness and analyzed query types, raising ethical concerns.

Sub-Saharan Africa faces high student-teacher ratios and teacher shortages, limiting personalized learning. Adesua, a WhatsApp-based AI Teaching Assistant for science, integrates textbooks and 33 years of exam questions with generative AI to provide conversational Q&A and automated assessment. A 6-month feasibility study in Ghana with 56 active users showed a 93.75% helpfulness score for AI answers, indicating potential for scalable, low-cost learning support.

This study investigates how humans produce language under limited vocabulary (down to 250 high-frequency words) and compares their behavior to greedy and globally optimal sampling algorithms using Sequential Monte Carlo inference with large language models. Humans generally resemble greedy sampling, but more skilled individuals backtrack and revise. In high-constraint settings, humans rely on semantically light words, a pattern also seen in both algorithms. Results have implications for resource-rational cognition, psycholinguistics, L2 communication, and language impairments.