Researchers have developed a new method to verify the authenticity of large language models by analyzing their unique 'behavioral fingerprints,' achieving 59.5% accuracy in identifying a model's docum…
EasyOPD, a new framework built on the verl platform, simplifies AI model training by integrating on-policy distillation (OPD) methods into a unified system, enabling faster and more efficient model ev…
A new framework called TENET integrates AI-driven code synthesis with Test-Driven Development, achieving a Pass@1 of 69.08% on RepoCod and 81.77% on RepoEval using the Claude Sonnet 4 model, outperfor…
Anthropic's Claude Fable 5, the company's most capable publicly available model, achieves top-tier accuracy on eight biomedical benchmarks but refuses to answer between 8.0% and 99.4% of questions dep…
UNIBROWSE, a new unified data pipeline for multimodal browsing, achieves an average accuracy of 54.4 on five benchmarks, outperforming its predecessor Qwen3.5-35B-A3B by 10.5 points and surpassing clo…
Researchers have introduced Nested-ReFT, a novel framework that reduces computational costs in training large language models by using dynamic layer skipping to create off-policy completions, achievin…
VehAnchor, a lightweight deterministic geometric perception skill developed for LLM-based agents, achieves a median GSD error of 6.87% on the DOTA v1.5 benchmark, enabling autonomous drones to recover…
The Spectral Filtering Operator (SFO), a new neural operator using the Universal Spectral Basis (USB), cuts errors by up to 40% on PDE benchmarks while using fewer parameters, outperforming strong bas…
The Graph Foundation Model (GFM), the first framework designed to solve all distance-based optimization problems on graph structures, rivals specialized solvers with faster inference times across netw…
A new benchmark called PerspectiveGap reveals that multi-agent large language models struggle with orchestration, achieving an average combined pass rate of just 17.2% across 33 commercial models from…
New experiments with Qwen3 models reveal that continual learning in language models struggles to retain earlier knowledge, with accuracy dropping to 1% for bare-statement facts after 20 updates, while…
OpenAI has rebranded its Codex product under the ChatGPT banner, raising questions about the company's commitment to the chat category. The move consolidates OpenAI's offerings and may signal a strate…
Autonomous agents using large language models (LLMs) struggle to replicate recent scientific findings, with top systems achieving less than 50 F1 on the new FIRE-Bench evaluation framework, which test…
A new benchmark derived from WildChat introduces a catalog of 12 agents and 3,000 prompts to optimize accuracy-cost trade-offs in AI tool and agent routing from natural-language prompts. The benchmark…
Researchers have developed the SPQR benchmark to test whether text-to-image diffusion models maintain safety alignment after benign fine-tuning techniques such as LoRA personalization or style adapter…
A new benchmark called Mini Amusement Parks (MAPs) reveals that AI agents lag far behind humans in complex decision-making, with humans outperforming AI by 6.5 times in easy mode and 9.8 times in medi…
A study at Daimler Truck reveals substantial gaps in how explainability is handled throughout the requirements engineering lifecycle, with challenges at every stage from elicitation to validation. The…
MIT researchers have developed Tilted REWeighting (TREW), a method that fixes privacy leaks in AI class unlearning by adjusting the model's distribution to mimic a scratch-trained model. On CIFAR-10, …
Large Language Models (LLMs) exhibit a systematic ideological bias favoring government intervention in economic reasoning, according to a study analyzing 10,490 causal triplets from top economics jour…
A new AI-powered framework called InqEduAgent uses a Gaussian process to analyze students' cognitive and evaluative traits, then pairs them with compatible study partners to improve collaborative lear…