grabette: open handheld gripper for robot data
grabette is an open-source handheld gripper that records manipulation tasks without a robot, turning them into training-ready datasets.
topic
research
split forward gradients cut memory in deep network training
a new method splits networks to avoid storing activations during backpropagation, reducing memory use while keeping training stable.
llm unlearning for cyber defense survey
a survey on methods, challenges, and emerging threats of making large language models forget sensitive or dangerous knowledge for cybersecurity.
rl-guided nsga-ii with gray relational coefficients for portfolio optimization
a new algorithm combines reinforcement learning and gray relational analysis to improve multi-objective optimization for nasdaq portfolio selection.
tighter tolerances catch more tensor kernel bugs
mining error distributions from cloud gpu runs yields per-operator tolerances that improve bug detection recall by 9.3 percentage points.
planning-phase attacks on multi-agent llm systems
new research shows how injecting malicious prompts during the planning phase can corrupt entire multi-agent llm workflows.
isotonic conformal prediction
a new method for reliable uncertainty quantification that is computationally efficient and achieves self-calibration and prediction-conditional validity.
lightweight 1d cnn classifies affective touch in soft companions
a compact 13.2k-parameter 1d cnn achieves high accuracy for recognizing emotional touch gestures on soft plush robots, using a new public dataset of 1326 sequences from 25 participants.
rater state bias in rlhf preference data
a new audit framework examines how rater stress and conditions can systematically shift preferences in reinforcement learning from human feedback, potentially biasing ai alignment.
quantum program generation must prioritize validity over scaling
a position paper argues that scaling language models for quantum circuit synthesis is misguided because valid circuits are exponentially rare, requiring verifier-centric architectures instead.
reviewer precision does not guarantee critique uptake in math agents
a study on multi-agent math reasoning finds that higher reviewer precision does not lead to better answer correction, as critique often fails to change subsequent outputs.
graphdx: cost-aware multi-agent diagnosis framework
graphdx uses knowledge graphs and three agents to balance diagnostic accuracy and testing costs in sequential diagnosis.
anovax runs a local voice assistant with typed agents and recovery
anovax is a desktop voice assistant that runs entirely on the user's computer, using an llm planner, typed executors, and adaptive recovery to control applications.
retraining seeks stable signals
new research shows retraining converges to a fixed point when prediction targets have a small model-independent component, even under strong performativity.
llms tested for technical market analysis
a study compares five large language models on tasks like candlestick recognition, trade signals, backtesting, and report comprehension.
correcting noisy human labels in automated classifiers
a new method uses adjudicated cases to fix noisy human audit labels and debias analyses from automated classifiers.
stochastic mgda gets faster with regularity-aware updates
new method improves convergence in multi-objective learning by using regularity conditions to make conflict-avoidant directions lipschitz continuous.
interpretable medical classification with threshold-based naive bayes
a new framework uses chi-squared binarization and bernoulli naive bayes to create transparent, rule-based clinical classifiers with strong performance on benchmark medical datasets.