[research]
73 stories · page 6
A federal judge approved the landmark settlement over pirated training data, but the core legal question remains unsettled.
A new paper reveals how prompt injection can be weaponized across multiple cooperating LLM agents, creating systemic risks that single-agent defenses can't handle.
Researchers propose Oracle, a memory system that lets AI agents maintain persistent, structured user profiles across sessions without retraining.
Researchers propose a method for AI agents to autonomously refine their own capabilities through structured feedback loops.
Leaked source code suggests the AI music generator built its training dataset by ripping audio from protected platforms, including YouTube Music and Deezer, amid ongoing copyright lawsuits.
New research proposes sticky routing for mixture-of-experts models, keeping tokens on the same expert across layers to cut communication overhead and improve throughput.
A new paper reveals that large language models struggle when sub-tasks depend on the truth of intermediate premises, pointing to a fundamental reasoning gap.
New research on compressing the key-value cache in transformer models could reduce memory usage and latency, enabling longer context windows and cheaper deployment.
Researchers propose Oyster II, a novel approach to align large AI models with human values, emphasizing robustness against adversarial attacks.
A new method enables LLMs to adapt to new tasks or domains while retaining previously learned capabilities, addressing a key challenge in model deployment.
GPT-5.6 tiers, Claude Fable, and an open-source surprise go head-to-head on raycaster, Rubik's Cube, calculator, and SVG tasks.
Researchers introduce a formal taxonomy and defense framework for prompt injection attacks, bridging theory and practice.
