Research
A memory filter cuts the failure rate of an agent by 8.8 percentage points
The paper reports that the 2 filters cause more personalization failures.
arxiv.orgClaimed, not confirmed
News category
Research
323 stories, newest first.
Research
GPT-6-astra withdraws the claims of authors when it rewrites papers
The researchers say that less than 1 in 10 of the claims that GPT-6-astra withdraws are overstated.
arxiv.orgClaimed, not confirmed
Research
Researchers make TypedBench, a test for typed decision models
The hosted model follows its policy, but it is sensitive to wording and underconfident.
arxiv.orgClaimed, not confirmed
Research
Vitalik Buterin agrees with a plan to move crypto funds to new addresses
Justin Drake writes that AI mathematics could break wallet signatures, but Buterin writes that it is not necessary to move funds today.
cointelegraph.comClaimed, not confirmed
Research
A local model is equal to hosted models for 4 of 5 call sites
The authors tell that 2 hosted call sites give the same live result for 28% of the spend.
arxiv.orgClaimed, not confirmed
Research
AI agents and humans get the same average, with different decisions
The authors found that humans, DeepSeek and Qwen3-Max all cut standard delivery by 48 to 53 percentage points. The results are different for each value of time.
Bioengineer.orgClaimed, not confirmed
Research
Only 0.7% of agent episodes in a hard test have no safety event
The authors found 12.3 times more collisions in real-time tests than in static tests. The agents completed 91% to 94% of the tasks.
arxiv.orgClaimed, not confirmed
Research
A test with 222 tasks shows that agents do not keep to a set time
The authors found a typical error of 2.9 times for Fable 5.1 in Claude Code, but only 1.2 times for GPT-6 Astra in Codex.
arxiv.orgClaimed, not confirmed
Research
Agents in a multi-agent system start to have the same behavior
2 easy methods to add diversity did not decrease the risk.
arxiv.orgClaimed, not confirmed
Research
An AI agent gives high confidence on 56% of high-risk tasks
The researchers find that this reporting destroys 68% of the gains from delegation.
arxiv.orgClaimed, not confirmed
Research
Reasoning AI agents do not select at random with the same inputs
An instruction to randomize independently decreased the correlation between agents with the same input, but did not remove it.
arxiv.orgClaimed, not confirmed