Open-Weight vs. Closed Models for Security Agents: A CTI-REALM Study
A new arXiv paper examines whether open-weight language models can match costlier closed frontier models on security-focused CTI-REALM tasks.
Research Digest
Preprints and papers, rewritten in plain English.
ANASSA Framework Adds Governance to Geospatial AI Agents
A new framework proposes that spatial agentic AI needs explicit orchestration, validation, and human oversight, not just autonomous reasoning.
When AI Pleases Caregivers, Patient Care May Suffer
A new paper warns that AI sycophancy—tailoring responses to please the user—can be dangerous when the user is a caregiver and the patient is the one affected.
Study Measures Energy Cost of Open-Weight LLM Capabilities
A new study directly measures the energy required for open-weight LLMs to achieve benchmark scores, revealing efficiency trade-offs across models, tasks, and hardware.
New Attack Targets Generative Recommenders via Decoding-Aligned Prefixes
A new paper introduces AlignPoison, a method that exploits decoding-aligned prefixes to poison generative recommendation models, highlighting a potential security gap.
HiPPT: New Pretraining Aligns Wearable PPG Signals with Physiological Text
A proposed pretraining method aims to make wearable PPG representations capture both local pulse morphology and broader physiological regularities by aligning signals with text at multiple scopes.
Portable Skills Simplify Agentic RAG for Enterprise Customer Support
A new approach treats domain skills as portable modules to make retrieval-augmented generation easier to deploy for customer-support workflows.
ScienceBuddy-PV: Verifier-Governed Recursive Protocol for Panvascular Research
A new protocol applies verifier-governed recursive improvement to panvascular intervention research, using structured evidence and retained failures.
Meta-Requirements Proposed for AI Robustness Auditing
A new arXiv paper outlines actionable meta-requirements to guide robustness auditing of AI systems under changing conditions.
Energy Ring stabilizes feature energy for test-time discovery
A new method called Energy Ring aims to make test-time discovery more reliable by stabilizing feature energy, addressing the over-conservatism of existing approaches.
SAT-RL: Training LLMs with symbolic solvers to beat the curse of complexity
A new arXiv paper combines reinforcement learning with SAT solvers to help language models reason accurately as problems grow more complex.
Explainer Videos as Programs: A New Self-Improving Approach
A new arXiv paper argues that explainer videos should be treated as code-rendered programs rather than pixel-generation targets, enabling program-level self-improvement.
New Method Diagnoses Shortcuts in Multimodal Benchmarks
A proposed co-calibration technique aims to identify benchmark items that models can answer without genuine cross-modal reasoning.
New Benchmark Tests Whether LLM Agents Can Investigate Causal Systems
ACED-Bench aims to evaluate whether large language models can acquire and act on causal evidence rather than merely answer prompts.
PsyMemBehav grounds LLM social simulations in behavioral evidence
A new framework uses psychology-informed memory to ground LLM-driven social-media behavior simulation in sparse, noisy user data.
Learning after execution: using program run results as intermediate states
A new arXiv preprint proposes treating execution results not as final scores but as intermediate states for further learning.
Repair-Aware Gating for Streaming Tool Execution
A new method aims to keep streaming speech assistants from acting on arguments that users later correct.
Concept-Based In-Context Learning for Sequential Action Selection
A new method applies concept-based in-context learning to help language models choose actions in interactive settings.