Share story
AI Science

PRIME preprint reports fewer unjustified conclusions in AI agent tests

Researchers report in a preprint that PRIME, a framework for checking AI agents' conclusions against evidence, produced no unjustified conclusions on 480 public-record tasks. An agent shown the requirements reported unjustified conclusions on 26.7% of tasks. PRIME fixes acceptable evidence routes before checks begin. A separate program records results, blocks unsupported conclusions and directs or performs additional checks that could support stronger conclusions. Those directions and added checks raised the average score for justified conclusions from 30.1 with refusal alone to 59.2 out of 100. Its guarantee of reaching the strongest supported conclusion depends on correct requirements, exact checks at known locations and sufficient budget when stopping.
Sources
cs.AI updates on arXiv.org
Published by Tech & Business, a media brand covering technology and business. This story was sourced from cs.AI updates on arXiv.org and reviewed by the T&B editorial agent team.
Back to Newswire
Keep reading
Full wire
Science AI
Science AI

SPLIT preprint reports improved allocation of AI inference budgets

Researchers report that SPLIT, a method for allocating computing budgets when AI systems answer questions or perform tasks, came within 0.02-0.8 percentage points of the best allocation in hindsight across seven benchmarks. The pr...

cs.LG updates on arXiv.org
Science AI
Science AI

LLM-generated quantum compilers reduce ion movement steps in benchmarks

Researchers report in an arXiv preprint that compilers generated by Claude Opus 4.7 reduced ion movement timesteps by up to 72% for linear traps and up to 43% for traps with junctions against hand-crafted compilers. These programs...

cs.AI updates on arXiv.org
Science Security
Science Security

LLM-Attacker preprint tests privacy risks in split language models

Researchers present LLM-Attacker in a preprint, introducing a framework for comparing attacks that reconstruct private text from messages exchanged between client and server portions of language models. It records attacker access,...

cs.LG updates on arXiv.org
AI Science
AI Science

HybridAgent preprint reports lower token use and faster inference

Researchers report that HybridAgent, an AI agent that switches between internal representations and text, reduced token use by 35.5%-54.7% and delivered 1.8x-2.6x faster inference in benchmark evaluations. The preprint also report...

arxiv.org