prompt injection as role confusion
research shows llms prioritize text style over content, enabling jailbreaks through role confusion.
topic
research
measuring cs curriculum alignment with new guidelines
a pipeline uses semantic retrieval and human review to measure how well a computer science program covers evolving curricular standards.
distilling misaligned foundation models for lightweight scientific forecasting
a new method extracts useful knowledge from large time-series models that don't fit scientific data, creating small, fast forecasters for sensor networks.
hidden anchors in multi-agent llm deliberation
a new model shows llm agents have hidden internal beliefs that shape group decisions, explaining how confidence can exceed initial levels.
deontic policies for agentic ai governance
a new policy framework extends beyond permit/prohibit to handle obligations, waivers, and conflict resolution for llm-driven agents.
space data centers face harsh realities
orbital ai data centers promise abundant solar power and cooling but must overcome radiation, heat rejection, and high maintenance costs.
adaptive uncertainty-aware refinement for llm judge auditing
aura iteratively learns human-consistency signals to audit llm-as-a-judge decisions with minimal human verification.
information lattice learning as graphical model structure learning
information lattice learning can be interpreted as a method for learning the structure of probabilistic graphical models by projecting probability distributions onto partition lattices and lifting rules back.
u-calibration algorithm adapts to smooth losses
a new forecasting algorithm achieves optimal regret for both general and smooth proper losses, overcoming previous limitations in u-calibration.
research agents leak private data through web queries
mosaicleaks shows that deep research agents often expose private information when they search the web, and training them to be better at tasks makes the leakage worse.
first in-orbit vision-language model demo on spacecraft
navi-orbital runs a vision-language model on a low earth orbit satellite to classify scenes, describe content, and answer operator questions in plain english.
shared workspace human-ai teams need coordination scaffolding
adding human collaborators to ai teams can hurt performance without structured coordination, but shared memory and approval gates help.
ceo-bench tests ai agents as startup founders
a new benchmark evaluates language model agents on long-horizon business management by simulating 500 days of running a startup.
securing ai agents with defense-in-depth
google deepmind outlines a control roadmap to manage internal ai agents by treating them as potential insider threats and using layered monitoring.
point observations barely help radar nowcasting
a graph neural network study finds adding sparse station data to radar forecasts gives minimal gains, while numerical weather prediction and satellite inputs matter more.
shock-wave theory linked to symmetry-reduced sgd dynamics
a new mathematical framework connects shock-wave theory to the learning dynamics of stochastic gradient descent after removing parameter symmetries.
training task generators without the solver bottleneck
a new method trains task generators using a lightweight probe instead of costly solver rollouts, making it practical to create frontier tasks for reinforcement learning.
trajectory-level supervision in offline rl
a statistical theory for offline policy optimization using only trajectory-level outcome labels instead of per-step rewards.