JetBrains ran 80 paired A/B tasks on Claude Code's Ponytail skill, finding actual code reduction of 15%, cost savings of 10.3%, and time savings of 11%—far below the advertised 54% code reduction and 22% token savings. The skill works by prompting the model to use existing functionality or write minimal code, but it only shows significant savings where the baseline code is already verbose. The skill does not self-activate in Claude Code; it requires a plugin to inject its ruleset.

OpenAI CEO Sam Altman says he doesn't believe people would want an AI to replace his job, emphasizing accountability and human connection. He also acknowledges his earlier views on AI replacing jobs may have been overstated.

This is a chat share with Claude.

MonkeysCode introduces Capuchin AI, an agentic IDE that users can fully own, offering complete control over the development environment.

‘Racist bias’ overestimating ages in Home Office’s facial-recognition software will lead to solo children being housed with adults, says Human Rights Network

This paper proposes a robot localization method that leverages rich semantic data from OpenStreetMap using VLMs to extract landmarks from panoramas and match them with prior maps, distilling a lightweight matcher for efficiency, and demonstrating generalization across diverse environments including snowstorms.

Efficient autonomous exploration with palm-sized nano-UAVs is challenging due to severe limitations in sensing, computation, and flight endurance. This paper presents a lightweight sensor-driven Lévy walk (SDLW) controller for aerial robots under 50 grams with sparse local sensing. The method combines discrete Lévy step-length sampling with a sensor-reactive heading policy using directional range measurements. Each robot independently samples its Lévy exponent from a uniform prior, eliminating the need for inter-robot communication. Simulations show coverage improvements of 79.6% in open arenas, 43.1% in rooms-and-corridors, and 13.6% in cluttered environments, with collision reductions of 13.0%, 7.1%, and 1.4% respectively, compared to a uniform-heading Lévy walk baseline.

This paper investigates reactive 3D control of a Franka Emika Panda manipulator using Star-World workspace reshaping. Intersecting inflated obstacles are clustered and replaced by star-shaped proxies before a dynamical-system-based end-effector controller is evaluated. In six PyBullet scenarios, reshaping reaches the goal in five of six scenarios, compared with four of six for the unreshaped baseline. It resolves the overlapping-wall case and requires 0.68–8.70 ms per workspace update. However, it increases path length, produces near-equilibria in two cases, and closes a navigable corridor due to over-aggressive merging. The results show both promise and limitations of transferring Star-World guarantees to a redundant manipulator controlled via inverse kinematics.

A cable-driven continuum robot arm is prone to residual vibration at the end of rest-to-rest maneuvers. This work applies time-delay filters as input shapers—both non-robust and robust—to suppress vibratory modes. Experiments show significant reductions in overshoot and settling time, with the robust shaper further improving performance, marking progress toward precise control of continuum robots.

This paper compares 13 path planning algorithms and proposes a hybrid approach combining Artificial Potential Fields (APF) with a Spatio-Temporal (ST) Transformer, demonstrating superior balanced performance for AUV navigation.

New research proposes Environmental Constraint Exploitation (ECE), an alternative approach that deliberately uses contact with the environment to simplify robot motion planning, reducing dimensionality and computational complexity, with benefits demonstrated on RRT-based planners and a real-world task.

Classical grasp quality metrics assume a single deterministic friction coefficient, failing to predict force closure loss under varying friction. This paper presents FIRMGrasp, a family of friction-volatility-aware metrics using Conditional Value-at-Risk (CVaR) to evaluate a risk-adjusted force-closure margin. Experiments on 1,599 grasps show that 53% of grasps certified by the traditional Ferrari-Canny metric lose force closure under adverse friction, while FIRMGrasp provides significantly better ordering and success rates.

This paper introduces an AD-assisted implicit derivative for regularised smooth contacts based on the Implicit Function Theorem, applied to Mujoco MJX. The method differentiates the stationarity residual at the tolerance-converged solution, avoiding solver unrolling and hand-assembled KKT systems, with memory growing significantly slower than unrolled AD. Additionally, optimiser distillation amortises batched full-horizon iLQR into a policy guiding short-horizon residual iLQR, achieving 28-98 percentage point improvements on Finger, Franka, and Unitree robots.

This paper proposes a differential-torque drift control method for a steering-free four-wheel independently driven (4WID) vehicle. Using a double-track vehicle model and a closed-loop controller, experiments on a 1:10-scale vehicle demonstrate steady circular drifting with a sideslip angle of approximately 20° and figure-eight drift tracking, proving the feasibility of drift control using only differential wheel torques.

Researchers propose the SWiFT framework enabling robotic fish to hold station in unknown turbulent flows using egocentric feedback and reinforcement learning, achieving significant improvements over existing methods.

MorphUNet is a novel diffusion-based morphing framework that formulates two-parent generation as alpha-controlled biometric transport, using CLIP appearance and ArcFace identity evidence with trainable dual cross-attention. It achieves state-of-the-art Morphing Attack Potential (MAP) on FEI and FRLL datasets, and remains highly difficult to detect under cross-dataset transfer. The framework includes stress testing with unseen identities from CFD.

Researchers introduce Diff-ID, a diffusion-based framework for generating identity-consistent facial images. It integrates ArcFace and CLIP embeddings to preserve identity while maintaining photorealism, and includes a morphing pipeline. The method achieves a strong trade-off between identity similarity and image quality.

The paper introduces RegAL, a unified active semi-supervised framework using topology-aware Pareto optimization to couple sample acquisition with unlabeled data utilization, outperforming baselines under extreme annotation scarcity.

This paper introduces I-LW-DETR, the first fully integer-only lightweight DETR, addressing deployment challenges of Vision Transformer detectors on NPUs and microcontrollers. It achieves 3.6x model size reduction and over an order of magnitude computational cost reduction with moderate accuracy loss.

PerceptionBench is a benchmark designed to evaluate atomic visual perception in multimodal large language models (MLLMs). It uses a bottom-up approach, building an error taxonomy from 42 existing benchmarks to define ten atomic perceptual capabilities, and creates 3,000 verified questions. Tests on 16 frontier MLLMs show no model exceeds 60% accuracy, with perception-related hallucination being the weakest capability.

Mage-VL is a novel streaming multimodal foundation model that reduces visual token consumption by over 75% using a custom Mage-ViT tokenizer that selectively encodes dynamic regions via motion vectors and residual energy, and employs a dual-system architecture for real-time perception. It matches or outperforms larger models on various tasks with up to 3.5x inference speedup.

State-of-the-art toxicity detectors for text-to-image generation are found to fail marginalized communities: about 35% of images labeled safe are considered harmful by disability communities. This paper advocates for community-specific toxicity detection (CTD) and demonstrates its feasibility with safety guidelines for dwarfism and blind/low vision communities. Experiments reveal that general-purpose detectors achieve F1 scores lower than random (0.32, 0.37) in zero-shot settings, while prompt-based methods improve up to 0.78 (GPT-4o). However, CTD performance still lags far behind general-purpose detection (F1≈0.9), highlighting the need for sustained research.

Social media imagery often contains vague or ambiguous geographic references, making accurate geolocalization challenging during disasters. The proposed DisasterTD framework integrates multimodal large language models (MLLMs) for semantic reasoning with cross-view geolocalization. It first extracts toponyms and generates candidate locations from noisy text, then refines them via cross-view matching among social media, remote sensing, and street-view imagery. Evaluated on the Hurricane Harvey dataset, it achieves 47.01% accuracy within 50 meters, significantly outperforming baselines, especially for ambiguous toponyms.

This paper proposes a track-leakage-free hold-out protocol for self-validation of photogrammetric reconstructions without ground truth. Experiments show it measures internal geometric consistency but not absolute accuracy, and fails to detect globally distorted but self-consistent models.

Human language exhibits a new scaling law: the influence of word context decays as approximately 1/d, with exponent near 1, across ten corpora from six language families. The effect vanishes in scrambled controls and does not appear in genomic sequences.

TabRank presents a framework for training reasoning rerankers for tabular retrieval by distilling chain-of-thought reasoning traces from Large Reasoning Models (LRMs). It achieves significant improvements across multiple datasets, with Acc@10 gains up to 52.9%, and demonstrates strong generalization to out-of-distribution and multi-table scenarios.

A new deep learning model combining temporal convolutional networks with label-wise attention significantly improves medical coding accuracy, achieving a 9% increase in F1 score and a 28% increase in recall over previous methods.

Human language relies on unspoken beliefs and belief updates. This study creates the first expert-annotated implicature cancellation dataset and evaluates LLMs' ability to recognize unspoken beliefs through implicatures and understand their updates. Results show LLMs lag behind humans, especially in natural scenarios, and their performance depends on prior beliefs and update type.

This paper describes a system for CLEF 2026 CheckThat! Task 2, focusing on automated verification of numerical claims in English and Arabic. Two approaches are explored: an LLM-based verifier fine-tuned with LoRA, and a lightweight TF-IDF reward model. Results show the LLM approach outperforms on most metrics, especially Recall@5, while the reward model excels on conflicting claims. Sub-claim decomposition did not improve performance. For Arabic, AraBERT outperforms multilingual baselines.

CogArena is a procedurally generated 13-paradigm benchmark for evaluating cognitive ability structure in LLMs. Testing on 55 open-weight models shows all paradigm correlations are positive, a common factor explains about half the variance, but within-grouping advantages are small and fail to meet validation criteria. The study concludes that current evidence does not support stable five-dimensional cognitive profiles.