Sam's News — tech-research — 2026-09-24¶
Healthcare AI¶
6.5 EviStreams: Human-in-the-Loop AI for Systematic Review Data Extraction¶
A system automates data extraction from medical studies for systematic reviews through human-in-the-loop AI, reducing the expert-labor bottleneck in the protocol-driven extraction workflow.
Sources: arXiv — Computation and Language RSS
Efficiency¶
6 Distilling Sequential Computation: Speeding Transformer Inference by Compressing Predictable Token Spans¶
A method reduces transformer inference cost by recognizing that many adjacent token spans are highly predictable and can be processed as stable units rather than token-by-token.
Sources: arXiv — Computation and Language RSS
6 Parallel Speculative Decoding: Concurrent Drafting and Verification in Transformer Inference¶
A method overlaps the draft and verification phases of speculative decoding to remove serialized drafting from the critical path of token generation.
Sources: arXiv — Computation and Language RSS
Reasoning¶
6 Planned Test-Time Scaling with Coordinated Reasoning Paths¶
A method improves test-time scaling by coordinating parallel reasoning branches rather than drawing them independently, reducing redundancy in reasoning attempts.
Sources: arXiv — Computation and Language RSS
Evaluation¶
6 Uncheatable Eval: Dynamic Compression-Based Language Model Evaluation Resistant to Data Leakage¶
A dynamic compression-based evaluation method makes it difficult for language models to memorize benchmark data, addressing benchmark contamination risks.
Sources: arXiv — Computation and Language RSS
Neurotechnology¶
6 Brain-to-Language Decoding: Methods and Applications for Communication Restoration¶
A comprehensive survey of brain-to-language decoding systems that translate neural activity into linguistic outputs for speech restoration and cognitive research.
Sources: arXiv — Computation and Language RSS
Legal AI¶
6 LabourCrew: Multi-Agent RAG Framework for Statutory Reasoning Over Labour Law¶
A multi-agent system with adversarial deliberation ensures that every claim in labour law question answering is traceable to verified evidence, not merely relevant.
Sources: arXiv — Computation and Language RSS
Tech Research¶
5.5 Goal-oriented agentic AI enables non-expert users to repair ubiquitous robots via natural language¶
A goal-oriented agentic AI architecture enables non-expert users to perform maintenance and repairs on ubiquitous robots through natural language interaction without visual interfaces.
Sources: arXiv — Artificial Intelligence RSS, arXiv — Robotics RSS
5.5 SR-Fraud: outcome-supervised LLM agent for real-time payment fraud detection¶
Researchers introduce an outcome-supervised reflective LLM agent framework designed to detect non-stationary payment fraud patterns in real-time streaming systems.
Sources: arXiv — Machine Learning RSS, arXiv — Cryptography and Security RSS
AI¶
5.5 COMED: Bridging Routing and Collaboration in Multi-LLM Inference Systems¶
A new approach addresses the gap between routing and output combination in multi-model LLM inference systems.
Sources: arXiv — Computation and Language RSS