Sam's News — tech-research — 2026-10-08¶
AI Research¶
6.5 OnlineQAT: On-Policy Distillation for Ultra-Low-Bit LLM Quantization¶
OnlineQAT applies on-policy distillation during quantization-aware training to recover accuracy in large language models compressed below four bits.
Sources: arXiv — Computation and Language RSS
6 ToolRACER: Benchmark for Robust Agentic Conversation Under Non-Cooperative User Behavior¶
ToolRACER provides a new benchmark for training and evaluating task-oriented conversational agents that remain robust when users deviate from expected scripts.
Sources: arXiv — Computation and Language RSS
6 Multi-Objective Aligned Small Language Model for SUD Patient Dialogue¶
Researchers develop a small language model framework aligned to generate realistic patient dialogue for substance use disorder counseling that reflects cognitive states and readiness for change.
Sources: arXiv — Computation and Language RSS
6 Expert Coupling in MoE Pretraining: Reducing All-to-All Communication Overhead¶
A technique using correlated expert placement and token shuffling reduces all-to-all communication overhead in Mixture-of-Experts pretraining across distributed GPUs.
Sources: arXiv — Computation and Language RSS
6 How Do LLMs Handle Negation in Predictions?¶
Evaluation of recent LLMs reveals that in 37–71% of cases they ignore negation and repeat predictions unchanged, indicating a fundamental weakness in processing this essential linguistic feature.
Sources: arXiv — Computation and Language RSS
6 SpikingVLA: Energy-Efficient Spiking Vision-Language-Action Models¶
SpikingVLA converts artificial neural networks to spiking neural networks for energy-efficient vision-language-action models, reducing required timesteps for inference.
Sources: arXiv — Computation and Language RSS
6 Catastrophic Forgetting in LLM Output Embeddings During Continual Learning¶
Analysis of catastrophic forgetting in continual pretraining and fine-tuning reveals that forgetting localizes to output embeddings of tokens absent from new training data.
Sources: arXiv — Computation and Language RSS
AI¶
6 Safety evaluation of open-weight LLMs on non-canonical inputs reveals gaps¶
Researchers assess the vulnerability of open-weight language models to obfuscated inputs including emojis, altered spellings, and encoded strings.
Sources: arXiv — Computation and Language RSS, arXiv — Cryptography and Security RSS
6 AI techniques improve energy efficiency of large data centers¶
Researchers propose AI-driven optimization methods to reduce the environmental impact and energy consumption of cloud computing infrastructure.
Sources: MIT — Computers RSS, MIT — Artificial Intelligence RSS
Security¶
6 Sigma-Hunter: domain-specific LLM for threat hunting and detection engineering¶
Researchers present a specialized language model trained to generate accurate detection rules and YAML configurations for security operations.
Sources: arXiv — Artificial Intelligence RSS, arXiv — Cryptography and Security RSS