A new benchmark called FlipSet reveals that most vision-language models fail at Level-2 visual perspective taking, with 75% of errors stemming from egocentric bias. In an evaluation of 103 models, the…
Researchers have developed GIAVA (Gaze Integrated Active-Vision ALOHA), a robotic vision system that mimics human foveated vision by integrating eye-tracking and perspective control data from human op…
AI search engines like ChatGPT are reshaping digital intermediation by resolving information needs internally, reducing outbound clicks to just 5.2% of sessions compared to Google's much higher referr…
The Neuro-Symbolic Knowledge- and Gradient-Guided Reinforcement Learning (KGRL) algorithm, developed to exploit domain knowledge stored in a Datalog knowledge base, significantly outperforms state-of-…
A new framework called IQA-T1 challenges traditional image quality assessments by integrating explicit perceptual observations with multimodal reasoning, achieving the best overall performance across …
Public sentiment toward AI in education, particularly ChatGPT, remains cautiously positive despite ethical concerns over academic integrity, according to a study using BERT-based topic modeling and Se…
AutoTrace, a new tool for vulnerability trigger localization, achieves a VulnHit rate of 75.0% and FuncHit of 80.8% on the InterPVD benchmark by using layer-by-layer exploration of code property graph…
Researchers have introduced Causal Rules and Maximally Specific Causal Relationships (MSCRs) to resolve statistical ambiguity in AI, building on Nancy Cartwright's definition of causes. The approach r…
A new study reveals that intermediate representations from models like CLIP and DINOv2, combined with semantically aligned references, significantly improve synthetic image attribution accuracy. The r…
Researchers enhanced Ensemble Determinization Monte Carlo Tree Search (MCTS) with dynamic resource allocation strategies, including Dynamic Number of Determinizations and Dynamic Simulation Allocation…
A new theoretical framework called the Compression Calculus proposes that intelligence is fundamentally about compressing information and recombining those compressed units, challenging current AI app…
A study on subjective expected utility (SEU) sensitivity finds that while the sensitivity parameter α is sharply recoverable in decision-making models, the belief and utility parameters β and δ remain…
Google researchers introduced DiffusionGemma, a 26-billion parameter mixture-of-experts speech recognition model that uses discrete diffusion to refine whole transcripts in parallel, achieving a 6.6% …
Researchers have developed a training-free counterfactual report-coordinate (CRC) clamp method that achieves a resist and update score of 1.00 on a Bayesian-witness benchmark, enabling AI language mod…
The European AI Act imposes new transparency, traceability, and human oversight requirements on high-risk AI systems, particularly in algorithmic hiring, but lacks comprehensive standards for recruitm…
Australia is introducing strict regulations on data center development and bolstering copyright protections for creatives against AI exploitation, Prime Minister Anthony Albanese announced. New rules …
A study evaluating 26 AI-generated routes found that all 57 deletions of steps improved scores, with an optimizer uncovering better structures in 21 routes, but the deletions did not necessarily impro…
TRACE (Typed Reasoning And Commitment Evidence), a new framework for recording AI decision-making processes in a structured and auditable way, aims to introduce transparency and accountability into ho…
A novel self-evolving critic framework improves LLM agent performance by learning from past actions, reducing Expected Calibration Error (ECE) by up to 54% across three agent benchmarks and three crit…
A study using Terminal-Bench 2.1 with GPT-5.4 and Claude Opus 4.6 found that automatic harness evolution for LLM agents fails to consistently outperform simpler test-time scaling methods, especially o…