Sam's News — tech-research — 2026-09-16¶
Medical AI¶
7 AI technique improves minimally invasive surgical navigation using X-rays¶
MIT researchers developed XVR, a patient-specific AI technique that matches X-rays captured during surgery with preoperative 3D medical scans in seconds with sub-millimeter precision. The system generates thousands of synthetic X-rays from a single preoperative scan to improve real-time surgical navigation in minimally invasive procedures.
- Achieves sub-millimeter precision matching X-rays to 3D scans
- Generates approximately 1,000 synthetic X-rays per second
- Designed for orthopedic and neurosurgery procedures
- Enables real-time navigation using flat X-ray images
- Improves positioning of surgical tools through tiny incisions
Sources: MIT News AI Web Searched, MIT — Computers RSS, MIT — Artificial Intelligence RSS
AI Safety & Fairness¶
6.5 Bias Audits Detect Bias but Disagree on Ranking: Evidence from Ten Instruments and Ten Frontier Models¶
Different bias audit tools produce inconsistent rankings of frontier AI models, questioning the validity of using audit scores to compare systems.
Sources: arXiv — Computation and Language RSS
6 Are We Grading Properly? Understanding Failure Modes in Medical Benchmarks¶
An analysis of how rubric-based evaluation used to grade open-ended medical LLM outputs at scale introduces systematic biases and failure modes.
Sources: arXiv — Computation and Language RSS
Security¶
6 Latent Undertow: How Ordinary Typos Break Probes¶
Probes designed to detect malicious prompts in LLMs fail reliably on ordinary typing variations despite models handling them fluently.
Sources: arXiv — Computation and Language RSS
NLP¶
5.5 The Functionalizer: Lossless Functional Decomposition for Subword Tokenization¶
A new method for subword tokenization preserves orthographic variations without fragmenting embedding spaces or requiring lossy normalization.
Sources: arXiv — Computation and Language RSS
Applications¶
5.5 Crash Narrative-Guided Countermeasure Recommendation Using Large Language Models: A Retrieval-Augmented Generation Framework for Intersection Safety¶
An LLM-based retrieval-augmented system automates the identification of crash mechanisms and recommendation of safety countermeasures at intersections.
Sources: arXiv — Computation and Language RSS
LLM Architectures¶
5.5 Retrieval-Driven Memory Reconsolidation for Long-Term LLM Agents¶
A memory system for long-horizon LLM agents that reconsolidates memory during retrieval rather than only during information arrival.
Sources: arXiv — Computation and Language RSS
5.5 The Immutable Past: Formalizing State Mutability and Conflict Resolution in Mutable RAG¶
Formalizes the semantic shadowing failure mode in retrieval-augmented generation when shared memory conflicts with updated information.
Sources: arXiv — Computation and Language RSS
LLM Capabilities¶
5.5 State of Thought Enables Endogenous Reasoning¶
A test-time reasoning paradigm for LLMs that relies on internally driven control rather than external reasoning programs.
Sources: arXiv — Computation and Language RSS
Multimodal AI¶
5.5 Towards Scalable RLVR: Multimodal Instruction Following Data Synthesis and Distillation¶
A data synthesis and distillation approach enabling multimodal instruction following through reinforcement learning from video rewards.
Sources: arXiv — Computation and Language RSS