Sam's News — tech-research — 2026-10-10¶
Privacy & Security¶
6.5 Context-Sensitive Memorization in LLMs via Prefix Extraction¶
Study shows memorized training sequences can be extracted in realistic deployment contexts beyond isolated prefix conditions.
Sources: arXiv — Artificial Intelligence RSS
Self-Improvement¶
6.5 Recursive Self-Improvement via Multi-Agent Self-Supervision¶
Framework overcomes supervision bottleneck in non-verifiable tasks by using the model itself as evaluator for recursive improvement.
Sources: arXiv — Artificial Intelligence RSS
Security¶
6.5 Option-Channel Attack: Exploiting Typed Decision Models as Agent Guardrails¶
Single-word injection attack circumvents typed decision model guardrails deployed to protect LLM agent systems.
Sources: arXiv — Artificial Intelligence RSS
AI Compliance¶
6.5 Regulatory Rule Sensitivity Audit of LLM Compliance Systems¶
Audit reveals LLM compliance verdicts sometimes ignore regulatory rules when deleted, swapped, or negated.
Sources: arXiv — Artificial Intelligence RSS
AI Transparency¶
6.5 Counterfactual Audit of Legal Chain-of-Thought Faithfulness¶
Audit reveals legal LLM reasoning may cite statutes without actually following them when facts are held constant.
Sources: arXiv — Artificial Intelligence RSS
AI Safety¶
6 Guard Models Insufficient for LLM Agent Safety Without Obligation Tracking¶
Guard models that only block forbidden actions fail to ensure agent safety; unfulfilled obligations also require monitoring.
Sources: arXiv — Artificial Intelligence RSS
Research Methodology¶
6 Survey: Benchmarks and Evaluation of Automated Research Systems¶
Comprehensive review identifies fragmented evaluation practices across literature synthesis, experimentation, and peer review automation.
Sources: arXiv — Artificial Intelligence RSS
Agent Learning¶
6 Hippocam: Intent-Structured Memory for Long-Horizon LLM Agents¶
Framework consolidates continuous agent experience into structured, reusable knowledge for sustained autonomous operation.
Sources: arXiv — Artificial Intelligence RSS
Model Training¶
6 Hindsight Hierarchies for Self-Improving Reasoning Models¶
Self-improvement loop trains reasoning models on solutions to problems beyond current capability, extracting generalizable insights.
Sources: arXiv — Artificial Intelligence RSS
Agent Safety¶
6 OnTrack: Real-Time LLM Agent Monitoring via Optimal Transport¶
Framework monitors and intervenes in agent trajectories in real-time to prevent costly or unsafe irreversible actions.
Sources: arXiv — Artificial Intelligence RSS