Sam's News — tech-research — 2026-08-31¶
AI Security¶
6.5 Research reveals quantization-triggered backdoors in language models with cross-quantizer transferability¶
A study demonstrates that post-training quantization of large language models can introduce transferable backdoors that persist across different quantizers, creating a validation-deployment security gap.
Sources: arXiv — Computation and Language RSS, arXiv — Machine Learning RSS, arXiv — Cryptography and Security RSS
Technology¶
6.5 Julia programming language evolved from MIT research project to global tool with millions of users¶
Julia, a programming language that originated as an MIT research project, has grown to have millions of users worldwide and is used for drug design, jet engine optimization, and heat pump development.
Sources: MIT — Computers RSS, MIT — Artificial Intelligence RSS
Machine Learning¶
6.5 TimesFM-3: zero-shot multivariate forecasting foundation model¶
TimesFM-3 is a new foundation model capable of zero-shot multivariate time-series forecasting.
Sources: Google Research Blog RSS
AI¶
6 PersonaForge: Multi-Turn User Simulation for Agentic LLMs¶
A framework for realistic multi-turn user simulation shows that 75.9% of real-world LLM interactions involve multi-turn queries, not single-turn assumptions in existing benchmarks.
Sources: arXiv — Computation and Language RSS
5.5 Vector Index-Based Method Accelerates LLM Inference Decoding¶
Researchers propose reformulating output projection using vector indices to reduce memory bandwidth bottlenecks during LLM autoregressive decoding.
Sources: arXiv — Computation and Language RSS
5.5 Adaptive Framework for Detecting Implicit Hate Speech in Online Content¶
Researchers develop a fine-grained detection framework for implicit hate speech that hides malice through metaphors and contextual hints.
Sources: arXiv — Computation and Language RSS
5.5 Study: Emotional Context Influences LLM Decision-Making Advice¶
Research shows that emotional vulnerability in user prompts affects whether large language models shift their decision-making recommendations.
Sources: arXiv — Computation and Language RSS
5.5 INSPIRE: Internalize-Then-Improve Framework for Mathematical Reasoning in LLMs¶
INSPIRE is a training approach that tests whether LLMs truly internalize mathematical concepts rather than memorize solutions to improve reasoning.
Sources: arXiv — Computation and Language RSS
5.5 Survey on Rubric-Guided Reinforcement Learning for LLM Alignment¶
A survey of rubric-guided reinforcement learning methods that use interpretable, multi-dimensional reward signals instead of scalar rewards for LLM alignment.
Sources: arXiv — Computation and Language RSS
5.5 Decoding Methods for Reliable Retrieval-Augmented Generation with LLMs¶
Researchers develop decoding strategies to identify when retrieval-augmented generation has insufficient or conflicting information to answer reliably.
Sources: arXiv — Computation and Language RSS Update to: ReliableRAG Combats Misinformation in Question Answering via Reliability-Guided Reasoning