WAQS: A New Method for Efficient Activation Steering in LLMs
A proposed technique aims to balance efficiency and expressiveness in controlling large language model behavior without fine-tuning.
Research Digest
Preprints and papers, rewritten in plain English.
MELD: Hierarchical Coordination for Self-Improving Multi-Agent Science
A new framework tackles the coupled problem of how AI research agents choose experiments and learn from them, enabling recursive self-improvement.
DKnown-Jev: A Fine-Tuned Model for Normative Knowledge Tasks
A new arXiv paper describes DKnown-Jev, a JEV-style model fine-tuned to handle tasks that depend on published rules and facts.
Training-free self-improvement for LLM agents via hippocampal replay
A new arXiv preprint proposes letting LLM agents learn from experience stored in external memory, without any parameter updates.
Study: Pluralistic AI alignment conflates three distinct questions
A new paper argues that pluralistic alignment research must separate questions about objectives, human references, and preference representations to avoid muddled results.
Hybrid Framework Infers Air Trips from Sparse LBS Data
A new two-stage approach uses mobile location data to reconstruct air travel when airport sightings are incomplete.
Recursive skill stacks improve zero-shot robot task composition
A new arXiv preprint proposes recursive self-improvement of shared executable stacks to make long-horizon robot manipulation work without task-specific training.
MSCR: A New Method to Balance Modality Use in Multimodal Learning
An arXiv preprint proposes MSCR, a technique that jointly balances how much each modality contributes to training while uncovering synergistic cross-modal information.
TorchTPU Brings Native PyTorch Execution to Google TPUs
A new open-source backend from Google and Meta aims to let PyTorch models run on TPUs without code changes.
Remix, Don't Retrain: A New Way to Adapt Web Agents
A proposed method uses mixture hypernetworks to let web agents adapt to changing sites without full retraining.
Self-Distillation Can Lead to Progressive Misalignment, Study Finds
New research shows that training AI models on their own outputs can gradually produce misaligned behavior even from benign starting points.
SPOT method speeds up diffusion language models via commit-order prediction
A new technique called SPOT enhances speculative decoding to close the speed-quality gap in parallel text generation.
Structure-Aware Kernels Improve Gaussian Process Port-Hamiltonian Models
A new arXiv paper proposes structure-aware spatial kernels to address inefficiencies in Gaussian process distributed port-Hamiltonian systems.
Quantile Certification for Policies from Logged Bandit Data
A new arXiv paper constructs a PAC-Bayes Bernstein lower-confidence envelope to certify upper tail quantiles for policies chosen from the same logged contextual-bandit sample.
Forced Aligners Fail on Spontaneous Noisy Speech
A new arXiv study shows that forced alignment tools degrade on real-world spontaneous speech, prompting a return to classical GMM-HMM methods.
Layer-Wise Probe Entropy Diagnoses GNN Failure Modes
A new diagnostic tracks how task-relevant information evolves through graph neural networks, revealing where message passing goes wrong.
Deterministic bounds for metric-space regression and classification
A new arXiv preprint derives worst-case accuracy and stability guarantees for regression and classification in metric spaces without relying on probabilistic assumptions.
Singletons Matter for Estimating Unseen Species
A new arXiv paper underscores the outsized role of singletons in estimating missing mass and unseen species, while flagging their variability in small samples.