vertex-softmax tightens transformer verification
new method finds exact softmax bounds from score intervals, reducing slack in transformer robustness checks.
aitag research
summaries filed under research
balora adds bayesian uncertainty to lora fine-tuning
balora extends lora with bayesian adaptation, improving accuracy and providing uncertainty estimates for large model fine-tuning.
aidecentralized novelty detection with quantized model exchange
a method for decentralized novelty detection that controls false discovery rate without sharing raw data, using low-precision model exchange.
aicapability elicitation vs creation in post-training
a free-energy perspective distinguishes whether post-training reweights existing model behaviors or creates new ones.
aiactive multi-predictor inference cuts label costs
a new method routes each case to a cost-appropriate set of predictors and samples labels based on uncertainty, reducing variance under a fixed budget.
aiwhere reliability lives in vision-language models
a mechanistic study finds attention sharpness is a near-zero predictor of correctness in vision-language models, while hidden state geometry and self-consistency offer stronger reliability signals.
aipreference embeddings for collective decisions
standard text embeddings capture semantics, not agreement, so new methods are needed to map free-text opinions for fair clustering and facility location.
aithinking machines lab unveils ai that talks and listens at once
thinking machines lab announced interaction models, a full-duplex ai that processes input and generates responses simultaneously, aiming for natural conversation speed.
airl-kirigami speeds inverse design of shape-shifting structures
a new framework combines flow matching and reinforcement learning to quickly generate valid kirigami cut patterns for target shapes.
aipathboost: gradient boosting learns graph features from paths
pathboost is a gradient tree boosting method for graph-level prediction that learns path-based features directly from graph structure, outperforming graph neural networks on half of benchmark datasets.
aihigh-dimensional multi-component ica phase structure
mean-field theory reveals initialization-driven competition and decoupling regimes in online independent component analysis.
aihow attention helps signal recovery in sequence models
a random matrix analysis reveals how attention weights affect signal detection in high-dimensional sequence representations.
aikernel selection as model selection for two-sample tests
a new method treats kernel choice in mmd tests as a model selection problem, using a complexity penalty to avoid overfitting and scale to deep kernels.
aiai writing is breaking my brain
jason koebler describes the zombie internet where ai-generated content blends with human writing, making it exhausting to filter and distorting communication.
aiai guides chemists to design molecules with plain language
a new ai system called synthegy lets chemists describe synthesis goals in everyday language, then scores and explains the best reaction pathways.
ai