Technology & AI
AI agent evaluation ignores time: this preprint fixes it
An arXiv preprint proposes replaying temporally-evolving enterprise worlds to test AI agents at any moment, fixing a blind spot in current evals.
AI Models
AI model releases and benchmarks — what each new model actually changes for the people who use it.
191 stories
Technology & AI
SkillBoost stops AI agents forgetting old skills
New arXiv paper proposes SkillBoost, a three-stage framework that stops LLM agents overfitting to limited experience and forgetting solved tasks.
Technology & AI
EvoPINN: AI agent invents new neural network for physics
EvoPINN uses an LLM agent to automatically discover and validate algorithms for physics-informed neural networks, replacing manual tuning.
Technology & AI
TAPR auto-rewrites LLM prompts to lift benchmark accuracy
TAPR uses reinforcement learning to turn vague user prompts into task-optimized instructions, lifting accuracy on Natural Questions and GSM8K benchmarks.
Technology & AI
AI science papers score 2.47 out of 5 in first AI peer-review test
AI-generated research papers scored below the midpoint in the first automated multi-model peer review, and one AI reviewer disagreed with the others
Technology & AI
AI agents do 100 office tasks cheaper than humans, but not as well
OmegaUse-OfficeVal pairs 100 real office tasks with labour costs, showing AI agents are cheaper and faster but still fall short on quality.
Technology & AI
OpenAI 'abundant intelligence' post lays out AI cost flywheel
OpenAI's July 31 post describes a full-stack push for cheaper, more capable AI, arguing falling intelligence costs change what work is worth doing for
Technology & AI
MTGuard paper targets malicious MCP tool use in AI agents
An arXiv preprint proposes MTGuard, a hybrid analysis framework that catches malicious MCP tool use in AI agents while preserving normal task performance.
Technology & AI
RSMeM: satellite AI agents learn from mistakes, 6% accuracy gain
RSMeM adds a knowledge-enhanced memory system to remote sensing AI agents, improving DeepSeek-V3.2 accuracy 6% with under 1% extra tokens, per arXiv.
Technology & AI
TraceCoder adds audit trail to AI-generated code
A new arXiv preprint tracks every repair attempt in LLM code generation, giving developers a replayable audit trail for AI-written code.
Technology & AI
AI search lifts e-commerce discovery to 80% at 30% cost
A preprint proposes a hybrid AI architecture for e-commerce search, extending discovery from 60% to 80% of queries at 30% of teacher model cost.
Technology & AI
23 AI agents tested on breach response: zero passed
SecRespond, a new arXiv benchmark, tested 23 frontier LLMs on real-world post-compromise incident response across 10 cyber ranges. Zero passed.