This work presents the first in-depth security analysis of widely used agentic systems for offensive security operations, revealing common design flaws that allow adversaries to exfiltrate API keys and compromise operator machines even within sandboxes.
CrossPool is a serving engine designed for cold Mixture-of-Experts (MoE) models that disaggregates FFN weights and KV-cache into separate GPU memory pools to address memory inefficiencies in sparse request scenarios. By consolidating static weights and dynamically provisioning active KV-cache demand, the system aims to improve GPU memory utilization and support bursty long-context requests.
A custom quantization recipe applied to the HuiHui abliterated model demonstrates superior performance compared to the vanilla 3.6-35B-a3b variant in mathematics and coding tasks. The results suggest that removing refusal mechanisms allows the model to achieve greater accuracy and wisdom in these domains.
This Reddit post shares an image featuring the quote "Open Source Models Will Eat Your Children" attributed to Amodei. The content consists of a link to the image and a link to the associated comment thread on r/LocalLLaMA.
Dario Amodei, CEO of Anthropic, has expressed concerns that open source AI models could lead to dangerous outcomes. The statement highlights the potential risks associated with unrestricted access to advanced artificial intelligence technologies.
The article discusses reasons why the scaling exponents of current Large Language Model applications indicate an unsustainable regime regarding energy resources.
This study conducts a rigorous reevaluation of nine recent Graph Foundation Models (GFMs) for node property prediction, comparing them against strong Graph Neural Network (GNN) baselines to address the lack of unified evaluation standards in the field.
Researchers present RaDaR, an open-source 32B parameter reasoning large language model designed to accelerate the diagnosis of rare diseases by addressing challenges in clinical deployability and data scarcity. The model was trained on nearly 50,000 public cases and over 100,000 synthetic cases, demonstrating superior performance across benchmarks and external validation centers.
The authors propose a reinforcement learning fine-tuning framework that utilizes autonomous vision-language evaluation as a scalable supervision signal for GUI agents, eliminating the need for manual labels or task-specific heuristics. By treating evaluator feedback as a noisy binary reward channel and deriving a noise-corrected estimator for Proximal Policy Optimization, the method addresses the difficulty of obtaining machine-readable rewards in open-ended desktop environments.
The authors present AdversaBench, an end-to-end red-teaming pipeline that generates hard inputs for large language models using five structured mutation operators and confirms failures through a three-judge panel with a meta-judge tiebreaker.
A lawsuit has been filed in the United States against major memory chip manufacturers Samsung, SK hynix, and Micron regarding allegations of price fixing.
DeepReinforce has released Ornith-1.0, an open-weight model licensed under MIT that achieves state-of-the-art performance among open-source models of comparable size on coding benchmarks. The model is built upon pretrained Gemma 4 and Qwen 3.5 foundations and includes variants with 9B Dense, 31B Dense, 35B MoE, and 397B MoE parameter counts.
A researcher submitting their first paper to arXiv reports that the manuscript has been under moderator review for two months despite passing automatic qualification checks. The author inquires whether this delay is normal and asks for advice on whether to resubmit or continue waiting.
The llama.cpp b9842 release introduces a change to deduplicate preset and cached model entries in the /v1/models endpoint. This update is signed off by Adrien Gallouët from Hugging Face.
This research investigates the use of large language models to detect scam phone calls in Turkish, a low-resource language where annotated data is scarce. The study introduces the first public multi-modal dataset containing 100 aligned audio-transcript pairs of scam and benign conversations.
This paper formalizes the fleet-memory problem in multi-agent LLM environments, identifying four foundational failure modes: unauthorized leakage, stale propagation, contradiction persistence, and provenance collapse. To address these issues, the authors define explicit systems-level primitives including scoped retrieval, temporal supersession, provenance tracking, and policy-governed memory propagation.
This research tests whether Benjamin Graham's classic value investing rules can act as a mathematical filter to prevent complex machine learning models from memorizing market noise. The study compares pure Graham rules, modern factors, and a combination of both against XGBoost and AutoGluon models using 20 years of S&P 500 data.
A study investigates the impact of overrefusal on small, on-device large language models when processing legal prompts, finding that authority-style prefixes systematically increase refusal rates by 2 to 20 times compared to a no-prefix baseline. While role-play jailbreak prefixes showed mixed effects across different models, the results indicate that these small LLMs are unstable under contextual framings typical of real institutional users.
This paper introduces ASALT, a method for lateral transfer learning in multi-agent reinforcement learning that accommodates mismatched state-space dimensionalities between source and target domains. The approach uses observation-level and state-level adapters to map inputs into a shared embedding space, enabling effective knowledge transfer across heterogeneous environments.
The author argues that upgrading from a single to dual GPU offers greater benefits through parallel processing rather than enabling the use of larger, higher-quality model quantizations. For coding tasks, the quality difference between Q4 and Q6/Q8 quantizations is minimal, making increased context window and throughput more valuable.