A new study introduces WebDecept, a framework for injecting deceptive interfaces to test web agent safety. Evaluations show current agents are highly susceptible to manipulations like targeted ads and domain redirection, and prompt-based defenses often fail.

A study on LLM-as-a-Judge reveals significant instability: pairwise preference flips 13.6% on average, with 28% of questions exceeding a 20% flip rate. GPT-4o-mini shows a first-position bias (72% A-majority). Cross-judge agreement is only 76%. The authors recommend multi-trial aggregation, position randomization, and explicit uncertainty reporting.

Research shows that trial-to-trial variability (structured noise) in cortical neural responses can significantly improve the robustness of artificial neural networks against adversarial attacks and naturalistic image modifications. The benefit is greatest for naturalistic modifications but transfers poorly across types, while noise structure from adversarial attacks can generalize to other attack types.

This paper identifies the double-drift phenomenon in diffusion policy optimization and proposes DiPOD, a framework that interleaves self-distillation with policy-improving gradient updates to maintain tight-bound behavior, stabilizing training and achieving higher rewards.

A new paper proposes Cheap LoRA (cLA) and a chained circulant variant c³LA, inducing sparsity in existing LoRA variants for parameter-efficient fine-tuning. Theoretical generalization error bounds are derived, and experiments across 11 fine-tuning methods, 10 pretrained models, and 14 datasets show that sparse structured column-space restriction remains competitive while reducing training time by up to 10% and peak GPU memory by up to 15%.

This paper presents D2H-AD, a novel anomaly detection framework based on Hyperdimensional Computing (HDC). It integrates distance-based similarity and density-aware encoding, outperforming existing baselines on five benchmark datasets with improved accuracy, interpretability, and computational efficiency, making it suitable for TinyML and edge AI deployments.

A new paper resolves the debate over whether weight norm causes grokking in neural networks by directly intervening on the norm during training. Under free training, grokking occurs when the norm reaches a critical value Wc that is highly consistent across seeds and learning rates. When the norm is clamped, delay follows an exponential law T_grok ∝ exp(αρ) with α≈7.5. LayerNorm eliminates this dependence.

Introducing a fully GPU workflow that integrates accelerated data generation with training of neural emulators augmented by uncertainty quantification and physics-aware refinement, leveraging differentiable solver JAX-Fluids for enhanced physical consistency in hypersonic flow prediction.

A new algorithmic pricing tool for fashion e-commerce sales campaigns uses daily-resolution demand forecasting and multi-objective optimization, achieving 6% higher profit in A/B tests at Zalando.

This paper presents llada.cpp, the first NPU-aware inference framework for accelerating diffusion LLMs on smartphones. It introduces three techniques—Multi-Block Speculative Decoding, Dual-Path Progressive Revision, and Swap-Optimized Memory Runtime—to overcome challenges like shrinking workloads, KV cache reuse, and memory overhead. llada.cpp reduces LLaDA-8B generation latency by 17x-42x over CPU baseline with no quality loss.

Yes, editing a single neuron can eliminate repetition loops in Gemma 4 models, but doom loops during extended reasoning persist due to missing facts, highlighting a knowledge precision problem.

In low-resource data silos, using local reference distributions for data selection can accelerate model collapse instead of preventing it, leading to power-law diversity decay. Collaborative Wasserstein proxy references, constructed without sharing raw data, mitigate diversity degradation.

TwinBI is an agentic digital twin framework that couples an LLM-based agent system with an executable BI dashboard state to unify conversational interaction, dashboard manipulation, semantic grounding, and provenance tracking. In A/B tests, it improved exact-match accuracy from 43.3% to 63.3%, partial-credit accuracy from 48.3% to 70.8%, and reduced timeout rate from 40.0% to 10.0%. A usability study confirmed benefits in task accuracy and workload.

YeasierAgent is an application-building paradigm based on symbiotic agents, narrative worlds, and scene-aware interaction. It challenges the conventional device-coupled model of software by redefining applications as collaborative spaces among users, agents, and worlds. Its architecture achieves two contributions: enabling rapid, cross-platform construction of agent-native applications using platform-agnostic interactive units (agents, scenes, dialogue), and unifying emotional companionship and practical tool execution within a single experiential sandbox. By integrating automated generation, user-created worlds, and spatial multi-agent collaboration, YeasierAgent formalizes the category of Symbiotic Agent-Native Applications, demonstrating a shift from isolated, tool-specific chatbots to…

A new paper compares Diff-in-Means (DiM) and Iterative Nullspace Projection (INLP) for steering refusal in safety fine-tuned chat models. The study finds that INLP counterfactual flipping matches DiM directional ablation in refusal suppression, while nullspace projection is weaker. Restricting INLP to leading directions preserves suppression with near-baseline perplexity, and the two interventions land in different activation regions, suggesting distinct representations for absence versus opposite of a concept.

In March 2024, the best agent on WorkBench, GPT-4, completed 43% of tasks and took harmful actions on 26%. By June 2026, the best agent, Claude Opus 4.8, completes 89% with only 2.5% harmful actions. Capability and safety improve together; basic mistakes persist; open-weight models lower costs. Updated benchmark released.

The Hybrid Open-Ended Tri-Evolution (HOTE) framework uses hybrid-mode reinforcement learning to collaboratively evolve a proposer, solver and judge based on web-scale knowledge, enabling autonomous agent evolution for open-ended deep research tasks. An 8B model trained with HOTE surpasses stronger static models and state-of-the-art methods with less time overhead.

This work proposes Orchestra-o1, an omnimodal agent orchestration framework that supports efficient collaboration across text, image, audio, and video. It introduces modality-aware task decomposition, online sub-agent specialization, and parallel execution, achieving 10.3% accuracy improvement on the OmniGAIA benchmark. The paper also presents DA-GRPO, a reinforcement learning method that trains Orchestra-o1-8B to state-of-the-art performance among open-source omnimodal agents.

The Muddy Children Puzzle, a classic about knowledge and ignorance that inspired epistemic logic, has unclear origins. This paper traces its evolution through 200 years of logical and literary publications, discusses its many variations, and presents a novel hat puzzle involving self-reference.

Arvind Narayanan and Sayash Kappor argue that AI will not cause mass unemployment, even in software engineering, citing NY WARN Act data and the real bottlenecks of the profession: deciding what to build, verifying deliveries, and deep human understanding.