Research
Claude models add more moral agency for Claude in 20 rounds
No other model did this, and the authors say the AI companies do not agree on alignment.
alignmentforum.orgClaimed, not confirmed
News category
Research
129 stories, newest first.
Research
Jev 1.13 leads the new Artificial Analysis Decision Index with 58
Claude Haiku 5.5 (Low) scores 61, but one task uses 11 times more time than with Jev 1.13, and the cost is 4.4 times more.
Research
49.8% of organizations had an unauthorized agent task, a report shows
Only 17.5% of the organizations can stop an agent automatically in production.
PR NewswireClaimed, not confirmed
Research
An AI agent changed the grade of 75.1% of 9,967 high cloud findings
The agent lowered 58.5% of the grades and increased 16.6%.
Help Net SecurityClaimed, not confirmed
Research
AI models did all eight attack tasks in a Booz Allen test
In one test, a model went from the network perimeter to the control zone in more than 16 minutes.
Research
A new method makes AI models count the same report one time
After 8 steps, LQET cut the change in log-odds by 89 to 91 percent in tests on Qwen3-4B and Phi-4-mini.
arxiv.orgClaimed, not confirmed
Research
A new score shows good data to train AI coding agents
On SWE-Bench-Verified, models that train on top-ranked trajectories get more points than models that train on random trajectories.
arxiv.orgClaimed, not confirmed
Research
Attention layers find harmfulness and MLP layers write refusals
Post-training increases the MLP refusal in all 18 pairs, with no systematic increase of the harmfulness decodability.
arxiv.orgClaimed, not confirmed
Research
DP-MIND keeps data safe in AI queries of 1,000,000 tokens
DP-MIND adds noise to the input on the local computer. A small memory keeps the facts that the task uses.
arxiv.orgClaimed, not confirmed
Research
Researchers increase Qwen2-Audio accuracy from 57.1% to 71.9%
The grounding profile of one model does not move to a different model.
arxiv.orgClaimed, not confirmed
Research
Researchers put encrypted networks in the weights of AI models
If the cryptographic assumptions hold, the circuits stay hard to find, also with full access to the weights.
arxiv.orgClaimed, not confirmed
Research
Anthropic operated a 26 billion parameter model on 15 TV boxes
The team wrote that a copy of such a model is not easy to stop, because the cluster has no one point to stop.
Startup FortuneClaimed, not confirmed