Entropic view of adversarial attacks enables data-minimal AI training
A single-author study proposes an entropy-optimising framework that turns adversarial attacks into a tool for training simpler, cheaper models with far less data.
Research Digest
Preprints and papers, rewritten in plain English.
Adaptive Pricing for Memory Allocation in Long-Horizon LLM Agents
A new arXiv paper proposes using adaptive pricing to help long-horizon agents decide what information to retain.
SQuIREL: A Context-Relative Way to Measure Entity Richness in Wikidata
A new framework proposes measuring Wikidata's 'richness' metric at the entity level, relative to context, rather than only at the ontology level.
Physics reasoning failures traced to world modeling, not knowledge
A new arXiv paper argues that LLMs' physics errors stem from flawed world modeling and proposes an axiomatic framework to address it.
EEG Super-Resolution Reframed as Dynamic State Decomposition
A new arXiv paper proposes treating EEG spatial super-resolution as a dynamic state decomposition and filtering problem, rather than a purely spatial reconstruction task.
M^2Net Unifies Scale-Specific Structures for Urban Forecasting
A new model tackles the coupled spatial and temporal challenges in urban forecasting by combining scale-specific structures with prototype decomposition.
Can AI Draft Laws for Unforeseen Situations?
A new study asks whether large language models can move beyond legal text to write rules for scenarios they were never trained on.
Population Health World Model Aims to Turn Data into Action
A new framework uses diverse digital health data to model both individual trajectories and population dynamics, moving beyond prediction toward actionable insights.
Paper Proposes Audit of Validation Signals for Clinical Causal KGs
A new arXiv paper highlights the need to audit extracted assertions in clinical causal knowledge graphs used for retrieval-augmented generation.
Hierarchical inference recovers accurate agent benchmark scores from small task subsets
A new method, HSPR, uses hierarchical posterior inference to correct score inaccuracies when evaluating software agents on small benchmark subsets.
Frequency-Domain Membership Inference Does Not Beat Confidence Thresholding
A band-resolved analysis finds that frequency-domain membership inference is no more effective than simple confidence thresholding, challenging a common assumption about memorized details.
MetaSkill: Formalizing Only Skill Boundaries for Agent Composition
A new approach proposes formalizing only the coordination points between agent skills, not the skills themselves, to enable reliable composition.
MM-LeJEPA: Gaussian Targets for Multimodal Self-Supervised Learning
A new paper extends LeJEPA's Fisher-information-based Gaussian embedding target to multimodal architectures.
Inhibitory Message Passing Cuts Token Costs in Multi-Agent Reasoning
A new arXiv paper proposes a neuro-inspired inhibitory mechanism to make multi-agent communication more accurate and token-efficient.
ServiceBench: Benchmarking LLM Agents for Complaint Handling
A new benchmark evaluates whether LLM agents can resolve service failures in multi-turn complaint scenarios marked by information asymmetry.
ReTRACE: LLM agent adds feedback-guided judgement to retrosynthesis planning
A new arXiv paper describes an LLM-based agent that uses feedback within a chemistry harness to handle whole-route decisions in retrosynthesis, beyond single-step prediction and tree search.
Verifying Human-Written Reasoning References for AI Evaluation
A new arXiv paper highlights that human-written reasoning chains used to evaluate AI systems are rarely checked for logical validity.
VLM Token Pruning Needs Better Metrics Than Accuracy and Token Count
A new arXiv paper argues that standard evaluations of visual token pruning in vision-language models rely on misleading proxies and should look beyond accuracy and token count.