New arXiv work proposes letting LLM agents decide per turn when reasoning is actually needed, using cross-turn likelihood signals instead of expensive verification, to cut redundant thinking across multi-step tasks.

#AI #LLMAgents #EfficientInference #Reasoning
https://arxiv.org/abs/2610.12061

arXiv paper shows LLMs can be trained to tolerate unreliable hardware, with error resilience growing alongside model size. Results from 40,000 GPU-hours suggest large models learn internal error-correcting…

#AIhardware #FaultTolerance #LLMs #EfficientInference
https://arxiv.org/abs/2610.10311

A new method called XOR-Trellis improves low-bit LLM weight quantization by combining a hardware-efficient dequantizer with curvature-aware path optimization, aiming to keep inference fast without costly codebooks.…

#LLMQuantization #AIHardware #EfficientInference
https://arxiv.org/abs/2610.00432