Diffusion Model Designed for Block Missing Air-Quality Data
A new diffusion-based imputation method accounts for the geometry of sensor networks to handle long contiguous gaps in air-quality records.
Research Digest
Preprints and papers, rewritten in plain English.
Can Softmax Attention Act as a Policy Mirror Descent Controller?
A new preprint asks whether causal softmax attention can be viewed as a repeated policy mirror descent controller, not just a one-step algebraic operation.
Neurosymbolic AI Gets a Method for Acquiring Constraints
A proposed architecture addresses the overlooked upstream problem of formalizing requirements from multiple stakeholders for neurosymbolic systems.
ENDOPROMPT Shows Prompt Injection Can Silently Reduce Task Utility
A new white-box method uses victim-side pseudo-references to degrade benign task performance without triggering harmful content.
When Long-Horizon Agents Can Forget Reasoning History
A new arXiv paper asks when language-model agents can discard old reasoning to keep context length and inference costs from growing without bound.
Neuro-Symbolic AI Proposed for Industrial Configuration Limits of LLMs
A new paper argues that the probabilistic nature of large language models makes them fundamentally unsuited for industrial product configuration, where exact outputs are required.
Augur Rehearses Reactions to Product and Policy Changes Offline
A new arXiv paper describes Augur, a synthetic decision lab that uses a typed knowledge graph and grounded personas to simulate how people might react to changes before they ship.
Let Specifications, Not Agents, Sign Off on LLM Output
A new arXiv paper argues that LLM agents should not be the final judges of their own work; external specifications should hold the pen.
Synthetic hospital benchmark opens EHR data for clinical AI
A new physician-validated synthetic EHR dataset aims to give frontier models the realistic longitudinal benchmarks they lack.
EEG Decoding of Imagined Speech Moves Toward Subject-Independent Models
A new arXiv study proposes a strictly subject-independent EEG approach to decode imagined speech, addressing a key limitation in prior evaluations.
PrivDrift Audits Secret Leakage in Active LLM Conversations
A new arXiv preprint proposes an auditing method for user-secret leakage when topics drift in persistent LLM sessions.
EnigmaForge benchmark hides logic puzzles inside document stacks
A new arXiv benchmark tests AI by burying a unique logic puzzle in documents with no explicit question.
LLM Ideation: Optimised for Novelty and Clarity, Not Impact
A new arXiv preprint notes that LLM-based scientific ideation is optimised for easily judged proxies such as novelty and clarity, leaving genuine scientific impact as the open question.
Search-Aware RL Improves Query Understanding in Roblox Game Search
A new arXiv paper describes a reinforcement learning approach that aligns query understanding with downstream search performance in Roblox game search.
NNV3 Expands Formal Verification to New Neural Network Architectures
A new release of the MATLAB-based NNV tool broadens formal verification to more deep learning models and learning-enabled cyber-physical systems.
Style, Not Self-Recognition, Explains LLM Code Attribution
A new study suggests language models identify their own code by surface style, raising concerns about collusion in AI-as-judge systems.
HEXIS Compiles Agent Skills into State Machines
A new approach turns agent skills into extended finite state machines to separate task reasoning from control flow.
AgentX-Model Framework Links Experiments for Industrial Recommender Research
A new model research framework aims to sustain industrial recommender systems by using prior experiment results to guide future investigations.