Boundary-Aware Obfuscation Shields On-Device Models from Cross-Node Attacks
A new arXiv paper proposes obfuscating model boundaries to protect on-device machine learning from attackers who gain white-box access via compromised edge nodes.
Research Digest
Preprints and papers, rewritten in plain English.
AI Safety Tradeoffs Are Fundamental, New Impossibility Proofs Argue
A new arXiv paper formalizes safe-AI goals as a three-player game and shows that desiderata like harmlessness and helpfulness cannot all be satisfied at once.
Simpler Logarithmic Approximations for Optimal Decision Trees
A new arXiv paper offers simpler logarithmic approximation algorithms for optimal decision tree and adaptive set cover problems.
Scope-Transfer Poisoning Can Make RAG Systems Misapply Rules
A new arXiv paper describes a retrieval-augmented generation attack in which rules valid in one documented context are applied to questions outside that scope.
ClawMMU: Protection-Coherent Memory for Stateful LLM Agents
A proposed virtual-memory design for LLM agents tracks who may act on long-term memory, addressing a gap where hidden instructions in fetched pages can go unchecked.
Conversational AI Needs Human Heuristics, Study Says
A new arXiv paper argues that conversational AI should be designed around the heuristics and interactional strategies people use in everyday decisions.
Bagging-Based Detector Targets Malicious LLM Prompts
A new arXiv preprint proposes bootstrap aggregation to make detection of malicious prompts for large language models more efficient and adaptable.
Unified framework benchmarks privacy attacks on split LLMs
A new red-team suite aims to make privacy evaluations comparable across split and distributed large language models.
Black-Hole Attack Poisons Vector Databases with Malicious Embeddings
A new arXiv paper describes a poisoning attack that exploits embedding-space defects to redirect retrieval in vector databases.
Spatio-Temporal Attack Targets Cognitive Radar Deep Networks
A new arXiv paper describes an adversarial attack that combines digital deception with physical fidelity to threaten cognitive radar systems.
EGGHead: Single-Image 3D Heads with Efficient Gaussian Splatting
A new method uses Gaussian splatting to generate photorealistic 3D head avatars from a single image, balancing quality, speed, and generalization.
ECHO method links anonymous reviews to submissions
A new technique exploits review-process artifacts to link anonymous reviews to their submissions, threatening reviewer anonymity.
Causal attribution for alarms with scarce incident data
A new Shapley-based framework traces alarm-score changes to specific causal mechanisms, even when incident data are limited.
Kernel Framework Handles Non-Hilbertian Distances via Sparse Landmarks
A new Sparse Landmark Embedding kernel guarantees positive semi-definite kernel matrices for arbitrary distance measures, extending Gaussian processes to manifolds and probability spaces.
Beckmann World Models Learn One-Step Predictions Directly
A new arXiv preprint proposes a world-model variant that learns a terminal map for one-step action-conditioned prediction, potentially reducing training compute.
Quadratic Sample Complexity for Strictly Replicable Learning
New bounds pin down the sample cost of forcing two learning runs to agree when the marginal is unknown.
Sage-A1.5: 35B MoE Model for Long-Horizon Scientific Tasks
A new 35B mixture-of-experts agentic model aims to improve long-horizon problem solving through adaptive capability composition.
Causal World Model Generates Counterfactual Tumor Evolution
A new arXiv paper proposes CaMeWM, a causal medical world model that simulates how tumors and survival would evolve under competing treatments, addressing the inherently counterfactual nature of treatment planning.