Large Language Models (LLMs)

2026-11-02

RESEARCH · 1 SOURCE · arXiv cs.AI

Calibration methods for HiFloat4 quantization: HiGPTQ, HiAutoRound, HiFlatQuant

This arXiv paper (2610.41112v1) evaluates three calibration approaches—HiGPTQ, HiAutoRound, and HiFlatQuant—for the HiFloat4 (HiF4) 4-bit block floating-point format. The authors report that these methods narrow the accuracy gap to BF16 to under 1% and outperform NVFP4 across a range of models and downstream tasks compared with direct HiF4 conversion.

6.0

2026-10-27

RESEARCH · 1 SOURCE · arXiv cs.AI

Meta-analysis: LLMs show modest loss aversion, mixed diminishing sensitivity, human-like probability weighting

The paper (arXiv:2610.33103v1) meta-analyzes 102 fitted estimates from 11 studies covering 32 language models to test prospect-theory components in LLMs. It finds modest loss aversion (pooled λ = 1.26, 95% CI 1.06–1.46), value functions that are near-linear or method-dependent for diminishing sensitivity, and human-like inverse-S probability weighting (γ ≈ 0.81–0.84) though based on fewer studies.

6.0

2026-10-07

RESEARCH · 1 SOURCE · arXiv cs.AI

AegisFlow: agentic multi-agent framework for autonomous pipeline remediation (arXiv:2610.06971v1)

arXiv preprint arXiv:2610.06971v1 introduces AegisFlow, a multi-agent framework that uses a Watchdog agent for telemetry and an LLM-powered Repair agent to generate, test and deploy code patches in digital-twin environments via a Parallel Shadow Patching (MAPE-K) execution model. The authors report a 98.1% reduction in mean time to repair (from ~170 minutes to 3.2 minutes), a 92% overall patch success rate, and scenario rates such as 96% for JSON schema drift, 98% for punctuation drift and 85% for Shadow DOM failures; the system is described as deployment-agnostic and plugin-friendly.

6.0

2026-09-23

RESEARCH · 1 SOURCE · arXiv cs.AI

ProcessLight and STeP-PO: LLM-based step-wise policy optimization for traffic signal control (arXiv:2609.22746v1)

The paper presents ProcessLight, an LLM-centered framework that decomposes traffic-signal decisions into verifiable semantic steps, and STeP-PO, a reinforcement learning method that assigns step-level advantages using step quality and importance to optimize reasoning tokens rather than only final outcomes; experiments on multiple real-world datasets reportedly show improved performance and the authors released code on GitHub.

6.0

2026-09-15

RESEARCH · 1 SOURCE · arXiv cs.AI

Carbon-Aware Routing for Function Calling in Edge-Cloud LLM Systems

An arXiv paper proposes a carbon-aware routing framework that distributes function-calling queries across a three-tier edge-cloud architecture using edge and cloud LLMs. A lightweight k-NN predictor in a unified semantic-lexical embedding space estimates per-query accuracy, delay, and power, and combines these with real-time grid carbon intensity to route queries to the lowest-emission tier; evaluations report matching cloud-level accuracy while cutting operational carbon emissions by about 4× on average.

7.0

2026-08-28

RESEARCH · 1 SOURCE · Apple Machine Learning Research

LLMs Are Not (Consistently) Bayesian — paper quantifies deviations from Bayes updates

The paper treats LLMs as information-processing rules and introduces the 'information processing gap' metric to quantify deviations from Bayesian updates when LLMs update probabilistic beliefs in light of new evidence. The authors run extensive experiments to evaluate internal (in)consistencies of LLMs' belief-updating behavior, with relevance to high-stakes domains like medicine, science and law.

7.0