Benchmark Tests AI Health Assistants on Time-Critical Self-Diagnosis
A new benchmark probes whether AI health assistants can recognise when a plausible explanation masks a time-critical emergency.
Research Digest
Preprints and papers, rewritten in plain English.
SPEM: Structure-Aware Memory Cuts Skill-Loading Overhead in LLM Agents
A new arXiv paper proposes a structure-aware procedural memory to help LLM agents select and use skills without the token and runtime cost of loading full skill libraries.
PROVISO Tracks Content Provenance to Secure LLM Agent Actions
A new defense framework, PROVISO, selects LLM agent actions based on whether their content originates from the user, addressing a blind spot in existing safeguards.
Preference Optimization Teaches Poker Agents Strategic Play
A new preprint examines whether preference optimization can push compact poker agents past legal but strategically poor moves under hidden information.
Entropy-Driven EDMA Framework Targets Multimodal QA Distractors
A new arXiv preprint proposes using entropy to help multimodal LLMs focus on relevant modalities and ignore distractors.
Magnolia Framework Brings Persistent Memory to Chemistry Agents
A new arXiv paper proposes Magnolia, an agent framework that uses persistent memory to help computational chemistry cope with growing data.
AI K-Space Method Targets Motion Artifacts in Brain MRI
A new AI-based technique called MoSAIK could cut repeat brain scans by correcting motion artifacts in k-space.
New Benchmark Suite EGI-BENCH Evaluates LLMs for Enterprise Agentic Workflows
The EGI-BENCH benchmark suite targets the evaluation gap that currently gates enterprise adoption of LLMs in agentic workflows.
TEX: Execution-based cross-validation for LLM test-case scaling
A new method called TEX uses execution-based cross-validation to improve test-case generation for long-horizon language-model tasks.
Graph-Based Automatic Differentiation for Constrained PDE Optimization
A new arXiv paper proposes representing approximate PDE solutions on graphs to enable automatic differentiation with constraints for forward and inverse problems.
Latent Control Model Aims for Realistic Rodent Motion
A new arXiv preprint proposes a latent control model for generating realistic rodent motion to support embodied comparisons in neuro-AI.
LLM Agent Targets Slow, Fragile Legal Compliance Checks
A new AI sub-agent aims to replace manual legal review and brittle keyword rules for assessing text against Canadian and Quebec law.
Trust-Based Wrapper Secures Robot Swarms Against Compromised Links
A new algorithm gives robot swarms a statistical guarantee that compromised communication links are excluded while legitimate links are quickly re-admitted.
KPHarness Uses Diagnostic Probing to Adapt Long-Horizon Gaming Agents
A new arXiv paper argues that failed gaming trajectories should trigger diagnostic probing rather than direct harness changes.
When Does an Added AI Planning Capability Actually Pay Off? A New Standard
A proposed evidence standard seeks to isolate whether a single addition to an industrial AI planning system independently improves operational value.
Graph Model Captures the Winding Path from Initial Idea to Published Paper
A new framework represents research ideation as a closed-loop trajectory, acknowledging that published ideas emerge from iterative experimentation and revision.
Correcting Attribution Patching with Few Exact Measurements
A new arXiv paper proposes a way to correct attribution patching in mechanistic interpretability using only a few exact measurements, potentially reducing the cost of identifying important components.
When PTQ Calibration Helps or Hurts Quantized Text Embedders
A new arXiv study examines how post-training quantization calibration affects text embedders and proposes a validation-gated abstention to avoid collapse.