LLM prompt strategies for extracting ophthalmic report data
A new arXiv study describes a method for using large language models to turn ophthalmic PDF reports into structured data, though results are not yet reported.
Research Digest
Preprints and papers, rewritten in plain English.
Scoring Sentence Specificity in Technical Documentation
A new arXiv study proposes sentence-level specificity scores to help technical documentation support collaboration between people and AI teammates.
Madeleine: Recall for conversational AI when memories don't match the prompt
A new arXiv paper proposes a system that learns to surface relevant long-term memories even when they are not similar to the current user message.
Task Operators: A New Lens on In-Context Learning
A new preprint proposes task operators to model in-context learning dynamics and address the efficiency cost of processing every demonstration at inference.
AGO Quality Gate: Evidence-First Release Decisions for RAG
A new framework treats RAG version releases as evidence-based decisions, accounting for incomplete metrics and unreliable LLM judges.
Polish-Silesian MT benchmark and system beat larger models
A new benchmark and translation system for the Polish-Silesian dialect outperforms existing open-source and commercial models, highlighting a gap in dialectal machine translation.
New Reasoning Method Targets Thai Clinical SOAP Notes
A proposed approach, ASCRIBE, aims to reduce omissions and unsupported content in automatic Thai SOAP note generation.
Design Framework for Human-AI Collaboration in Industrial Work
A new arXiv preprint outlines a design framework for human-AI collaboration, focusing on GenAI's role in decision support, automation, and assistance.
Automating AI R&D Could Trigger an Intelligence Explosion
A new analysis warns that as AI systems take over more of the research pipeline, progress could accelerate radically and become hard to control.
High-Bandwidth Flash Studied as Remedy for LLM Memory Bottlenecks
A new arXiv paper characterizes high-bandwidth flash memory as a possible way to ease the memory capacity and bandwidth pressures that limit large language model serving.
Decision Titan: Test-Time Training for Long-Term Memory in Offline RL
A new method called Decision Titan uses test-time training to improve long-term memory in offline reinforcement learning.
Testing the Stability Assumption Behind Action Chunking
A new arXiv paper examines why action chunking improves behavioral cloning policies by probing the stability assumption underlying the technique.
Local-First Harness Aims to Make Small Open Models Useful for Real Tasks
A new agent harness, Mingbird, targets the failure modes that prevent small open-weight models from completing real-world tasks on laptops.
Causal Memory Policy: Identifying Memory Utility in LLMs
A new arXiv paper proposes intervening on retrieval to make memory utility estimates in memory-augmented LLMs identifiable.
Cognitive Continuity Test Checks Whether AI Agents' State Changes Are Governed
A proposed test aims to verify that persistent AI agents' belief revisions and memory consolidations follow authorized transitions rather than merely resembling them.
SW-KAN Brings q-Orthogonal Polynomials to Kolmogorov-Arnold Networks
A new arXiv paper proposes replacing standard basis functions in Kolmogorov-Arnold Networks with Stieltjes-Wigert q-orthogonal polynomials, potentially improving approximation and efficiency.
Constant-Memory Recall Depends on Task and Training
A study of a DeltaNet variant shows that fixed-size recurrent memory can recall learned associations, but success hinges on the task and training.
RPTune: A Learned Approach to Context Curation for LLM Catalog Search
A new method, RPTune, learns to curate catalog context for long-context LLMs, offering small merchants a simpler alternative to multi-stage retrieval.