· 3 items
arXiv ↗
Pre-Compiled Pipeline Shards for Distributed LLM Inference on Intel A…
- What
- Pre-Compiled Pipeline Shards enable distributed LLM inference across Intel AI PC fleets.
- Why
- It allows running large 70B models locally on modest unified memory hardware.
- Impact
- Run large local models on standard hardware fleets
arXiv ↗
Harness Continual Learning: Continual Adaptation Beyond Model Paramet…
- What
- Harness Continual Learning extends agent adaptation beyond weights to prompts, tools, and memory.
- Why
- It formalizes non-weight architectural adaptation for modern agents.
- Impact
- Upgrade agent scaffolding and tool-routing patterns
arXiv ↗
Training-Free Inference-Time Self-Reflection and Cost-Bounded Early S…
- What
- EvoResearcher adds training-free, inference-time self-reflection to frozen LLMs.
- Why
- It bypasses expensive reinforcement learning training loops for reasoning models.
- Impact
- Build cheaper reflection loops without training models