Matching Retrieval Scope to Certification Prevents Agent Memory Interference
A new arXiv paper argues that aligning retrieval scope with certification scope makes persistent memory edits in language-model agents more reliable.
Research Digest
Preprints and papers, rewritten in plain English.
New benchmark tests Indian banking AI assistants for process errors
IndicBankBench evaluates language model assistants in Indian retail banking by examining the full action sequence, not just the final answer.
Repeated Benchmark Evaluation Needs Replication for Honest Uncertainty
New results characterize when repeated evaluation can certify narrow uncertainty, showing replication is required even when scores are accurate.
New 'Entropy Triangle' Framework Aims to Predict Heart Rhythms
A preprint proposes a machine-learning framework, tested on data from over 10,000 patients, to improve prediction of cardiac arrhythmias.
Automated Assessment for AI Courses Beyond Input-Output Tests
A new arXiv preprint argues that automated programming assessment should move beyond simple input-output tasks for non-trivial AI courses.
Deviation-Guided Skill Evolution for LLM Agents
A new method treats deviations from reference paths as learning signals rather than failures, helping LLM agents improve skills for tool-use tasks.
Reinforcement Learning Picks Classifiers for Fatty Liver Prediction
A new arXiv paper proposes using reinforcement learning to automate the selection of classification algorithms for predicting non-alcoholic fatty liver disease.
Task-Substitution Framework for Automated Governance
A new framework proposes replacing human review tasks with automated systems while preserving accountable authority.
Auditing Generative Search for Omitted Source Relationships
A new arXiv paper proposes a claim-gated audit method to catch when generative search answers cite supported passages but omit source relationships that change meaning.
Deep learning model assesses quality of ambulatory ECG signals
A new deep-learning signal quality assessment model aims to distinguish clean from noisy ambulatory electrocardiograms.
RD-JEPA: Few-trajectory transfer for reaction-diffusion surrogates
A new self-supervised architecture aims to make PDE surrogates transferable across different governing operators with minimal new simulation data.
New Benchmark Measures How Coding Agents Improve Engineering Governance
SWE-Prometheus evaluates LLM coding agents on real-world repository tasks rather than starting from pre-identified issues.
Knowledge graphs proposed for chest X-ray device reasoning
A new arXiv paper outlines a knowledge-graph approach to reason about catheters and tubes in chest radiographs, bridging image labels and text reports.
Auditable framework for open-ended lung cancer decisions
A new arXiv preprint proposes a conditional-strategy framework that makes complex lung cancer care pathways explicit and auditable.
Differentiable Parameterization Lets Networks Learn Their Own Size
A new approach treats network size as a learnable, differentiable quantity, merging architecture search with weight optimization.
Live Assistant: Learning When and Whom to Help in Livestreams
A new framework models the evolving needs of livestream audiences to decide whether, when, and whom to assist in real time.
Discrete Diffusion via Evolving Variational Autoregressive Networks
The paper proposes a framework that gives discrete diffusion models both sampling and direct likelihood estimation.
KITE: Architecture-aware scaling for efficient agentic LLMs
A new arXiv paper argues that model architecture determines how much compute is spent across training, prompt processing, and autoregressive decoding.