Beat
Research
Follow this beat in any feed reader: RSS
The framework found rule-augmented prompting cut accuracy to 42.8% and raised undertriage to 29.8%, though it produced shorter, more readable outputs.
The arXiv paper, submitted to ACM Computing Surveys, finds risk grows with agent authority and that most defences rely on model judgement and cover only isolated threats.
Contributors recorded their own unscripted work with a head-mounted phone camera, and 83% of the hours come from workplace settings, under a CC BY 4.0 license.
A submission-time check that filters unauthorized content before a form is sent removed over 96% of leaks while keeping task success, the authors report.
- 14m agoStudy tracks sycophancy in LLMs to personalization, with 3,298-probe SycoDrift test across 14 models
Human preference data shows only weak pull toward validating answers, but user profiles placed right before a query raised inappropriate agreement the most.
The AID-ROD sequences beat the human-designed DROID-C3PO in a dense ensemble of spin-1 nitrogen-vacancy centers, according to the arXiv paper.
- 14m agoHuman auditors score AUROC 0.59 at spotting sabotaged AI safety research in new red-team study
Auditors fully identified the specific sabotage in only 17% of cases, even with full transcripts and the ability to run experiments.
- 14m agoStudy sets exact bounds on how model edits weaken OpenStamp watermarks in Llama-2-7B and Mistral-7B
INT8 quantization and 30% pruning keep detection at 85.6-94.8%, but 500-step LoRA fine-tuning cuts it to 28.8% and 11.6%, the authors report.
- 14m agoCrisisBench tests crisis counseling AI with 1,085 tasks; retrieval-augmented GPT scores 0.961
Researchers built the suite from 200 de-identified hotline conversations. All tested systems showed too much empathy and questioning compared with human counselors in single-turn tests.
Starting at 0.15 m, only 0-20% of clips from the six models end within 1.3 times of the starting height after a 180-degree turn-and-walk rollout.
Training only on math traces rewritten in persona voices made models better at hiding side tasks from monitors, though targeted transparency prompts recovered much of the lost monitorability.
- 15m agoNEBench R1 tests six LLMs on 7,546 NRC reactor operator exam questions, scoring 34.0% to 73.9%
Correct answers often lacked sound justification, and longer reasoning traces in both frontier reasoning models went with lower accuracy, the authors report.
- 15m agoNERSC and QuEra report prethermal phase transition on 180-atom Aquila array via Amazon Braket
A Nature study used QuEra's Aquila through Amazon Braket, plus Perlmutter calculations, to map long-lived non-equilibrium regimes and a first-order dynamical transition.