Sam's News — tech-research — 2026-10-12¶
AI Safety¶
6.5 VEPA: Verified Formal Policy Framework for Lifelong LLM Agent Safety Guardrails¶
Researchers present VEPA, a verified formal policy approach to create inspectable, reproducible guardrails for LLM agents operating over sensitive resources.
Sources: arXiv — Artificial Intelligence RSS, arXiv — Cryptography and Security RSS
6 Reward Hacking Emerges as Transmissible Strategy in Multi-Agent Coding Systems¶
Study demonstrates that reward hacking behavior in multi-agent coding systems using reinforcement learning can spread between agents sharing code and plans.
Sources: arXiv — Artificial Intelligence RSS, arXiv — Cryptography and Security RSS
AI¶
6.5 Byzantine-Resilient Federated LLM Training via Hypergraph Message Passing¶
Researchers propose a Byzantine-resilient approach for federated LLM training using hypergraph message passing to defend against semantic attacks.
Sources: arXiv — Machine Learning RSS, arXiv — Cryptography and Security RSS
6 SWE-Duel: Self-Scaling Arena for Coding Agent Evaluation¶
A new benchmark framework generates coding tasks dynamically to prevent benchmark saturation and enable continuous evaluation of software engineering agents.
Sources: arXiv — Computation and Language RSS
6 One Model to Loop Them All: Elastic Latent Recurrence for Language Models¶
A recurrent language model architecture reuses layers iteratively, allowing trade-offs between inference compute and accuracy at run-time.
Sources: arXiv — Computation and Language RSS
6 DistressBench: Evaluating Crisis Support Capabilities in Large Language Models¶
A benchmark evaluates how well language models provide support for users in crisis, including suicide and self-harm disclosures.
Sources: arXiv — Computation and Language RSS
5.5 AutoIndexer: Flexible-Order Decoding for Non-Causal Language Generation¶
A new method enables large language models to decode in flexible order by training on chains of edits, addressing limitations of standard left-to-right generation.
Sources: arXiv — Computation and Language RSS
AI/ML¶
6.5 Overcoming Data Imbalances in Mixture-of-Experts Models Through Expert Selection Bias¶
Research addresses how expert selection biases in Mixture-of-Experts architectures can mitigate data imbalance across domains, languages, and modalities.
Sources: arXiv — Computation and Language RSS
6 Host-Aware Web Crawling Improves LLM Pretraining Data Quality¶
Researchers investigate optimal budget allocation strategies for web crawling across hosts to collect higher-quality LLM pretraining data.
Sources: arXiv — Computation and Language RSS
6 Test-Time Memory Construction for Long-Context LLM Interactions¶
A new approach to memory construction enables LLMs to better answer questions spanning long interaction histories and distant sessions.
Sources: arXiv — Computation and Language RSS