𝗟𝗼𝘄𝗲𝗿 𝗔𝗚𝗘𝗡𝗧 𝗖𝗢𝗠𝗣𝗨𝗧𝗘 𝗖𝗢𝗦𝗧𝗦

𝗟𝗼𝘄𝗲𝗿 𝗔𝗚𝗘𝗡𝗧 𝗖𝗢𝗠𝗣𝗨𝗧𝗘 𝗖𝗢𝗦𝗧𝗦

LLM agents cost too much money. New methods lower these costs. You move agents from the cloud to small devices.

PANDO uses online skill distillation.

• Success rate: 58.3%.
• Token use: 115K per task.
• 58% to 61% fewer tokens than SGV and WALT.
• Lower memory use.
• Faster steps.

UI-KOBE uses knowledge graphs.

• 4B model success: 70.7%.
• No graph success: 58.6%.
• Graphs guide the agent.
• This removes the need for a large planner.

Problems remain.

• PANDO repeats some actions.
• UI-KOBE costs money to set up.
• It takes 6 hours per app.
• It fails on new apps.

Stop measuring only success rates. Measure token budgets. Set a 120K token limit per task. This finds designs for on-device use.

Source: https://dev.to/olaughter/agent-compute-drops-substantially-with-online-skill-distillation-and-graph-guided-knowledge-2ja8 Optional learning community: https://t.me/GyaanSetuAi