Sam's News — tech-research — 2026-09-07¶

LLM Reasoning¶

6.5 Study probes how large language models integrate external evidence in reasoning¶

Researchers investigate mechanisms by which LLMs integrate externally-supplied evidence into their decision-making.

Sources: arXiv — Computation and Language RSS

Security¶

6 Research reframes indirect prompt injection as test-time search problem¶

Researchers formulate indirect prompt injection attacks as a test-time search problem and develop an agentic attacker methodology to explore the attack surface.

Sources: arXiv — Artificial Intelligence RSS, arXiv — Cryptography and Security RSS

6 Study proposes federated learning for cross-organizational cyberattack campaign detection¶

Researchers develop a federated learning approach to detect orchestrated cyberattack campaigns across organizations without requiring sensitive telemetry sharing.

Sources: arXiv — Machine Learning RSS, arXiv — Cryptography and Security RSS

AI¶

6 Study shows diffusion model utility for both trajectory planning and safety-critical scenario generation in autonomous driving¶

Researchers demonstrate that a single pretrained diffusion traffic model can simultaneously perform guided trajectory planning and safety-critical scenario generation for autonomous vehicles.

Sources: arXiv — Computer Vision RSS, arXiv — Robotics RSS

6 Reducing False Refusals While Maintaining Safety in Language Model Alignment¶

Research proposes methods to balance safety-tuning in large language models to refuse harmful requests while avoiding false refusals that make models unhelpfully evasive on benign queries.

Sources: arXiv — Computation and Language RSS See also: Boundary-Aware Self-Distillation for contextual LLM safety refusal policies

6 Mechanistic Analysis of Reasoning Operations in Language Model Chain-of-Thought¶

Study investigates how large language models geometrically represent distinct reasoning operations like problem formulation, goal decomposition, and deduction during chain-of-thought reasoning.

Sources: arXiv — Computation and Language RSS

Healthcare AI¶

6 VERGE: NLP method for extracting early-onset colorectal cancer symptoms from clinical notes¶

VERGE is an LLM-based framework designed to identify red-flag symptoms of early-onset colorectal cancer in clinical notes.

Sources: arXiv — Computation and Language RSS

AI Safety¶

6 Boundary-Aware Self-Distillation for contextual LLM safety refusal policies¶

A method enables LLMs to apply topic-specific safety boundaries that vary by application context.

Sources: arXiv — Computation and Language RSS

6 Study detects cultural misalignment in open-weight LLMs across USA, Poland, and China¶

Researchers benchmark three open-weight LLMs for cultural alignment using World Values Survey data across multiple countries.

Sources: arXiv — Computation and Language RSS

LLM Interpretability¶

6 Study reveals miscalibrated readouts mask knowledge in language models¶

Research demonstrates that internal probes can reveal correct reasoning in LMs even when behavior suggests complete ignorance.

Sources: arXiv — Computation and Language RSS