PEFT/agents evals + fine-tuning techniques + agent memory/skills
Key Questions
What alternatives to LoRA are explored in PEFT research?
The Beyond LoRA article examines various parameter-efficient fine-tuning methods beyond standard LoRA techniques. These expand options for adapting models with lower resource use.
How does HyperTool improve agent performance?
HyperTool boosts tool-augmented agent accuracy from 15.69% to 35.29% on Kimi-k2.5 benchmarks. It focuses on enhancing agent capabilities through targeted fine-tuning.
What updates are provided on fine-tuning LLMs in 2026?
An overview article shares updated best practices for fine-tuning at scale, including tools like Databricks MLflow and Spark. New tutorials and guides complement these recommendations.
How are agent harnesses being used in evaluations?
HarnessX compiles existing agent harnesses to simplify evaluation instead of building custom ones. This streamlines testing for agent memory, skills, and performance.
What practical approaches exist for agent skills and standards?
Vercel's method uses skills, linters, and evals to align coding agents with design standards. Persistent memory techniques like Elasticsearch integration achieve high recall rates.
PEFT landscape expands with Beyond LoRA article exploring alternatives. HyperTool improves tool-augmented agent accuracy from 15.69% to 35.29% on Kimi-k2.5. Fine-Tuning LLMs in 2026 overview provides updated best practices. HarnessX compiles agent harnesses instead of hand-building. Maxime Labonne talk on training frontier small models. Persistent agent memory on Elasticsearch (0.89 recall). New fine-tuning guides and tutorials published. FORCE paper on VLA RL fine-tuning addresses imitation ceiling, relevant for open VLA models. Hybrid local-frontier agentic development workflow validated. Vercel's approach to imbue coding agents with design standards via skills, linters, evals – practical for agentic AI development.