- Published on
Agent Efficiency, Memory and Confidence
- Authors
- Name
- Martin Andrews
- @mdda123
Presentation Link
The sixty-third MeetUp of the Machine Learning Singapore Group, was titled : "Agents, Experts and Extracting Structured Data" - and had talks covering a wide spectrum.
My Presentation
My talk was titled "Agent Efficiency, Memory and Confidenc", and had the following outline:
- Efficiency
- Memory
- Confidence
- Wrap-up & QR-code (the latter to reduce audience distractions)
The main driver behind the Efficiency section was the Oppo paper that detailed how they had explored the trade-offs in agent performance over a few important dimensions (such as how to use memory, parallel roll-outs, etc).
Following up on the Memory angle, I talked about 'Memento' (fka AgentFly), and then progressed to an overview of two 'Confidence' papers - ARPO (Agentic Reinforced Policy Optimization), and Deep Think with Confidence (a paper from Meta). These two papers finally formally published the key drivers of entropix (the first directly, with applications to Agent RL, the second through a proxy confidence measure).
The slides for my talk, which contain links to all of the reference materials and sources, are here :
If there are any questions about the presentation please ask below, or contact me using the details given on the slides themselves.
Video about 'Confidence'
Other Presentations
The MeetUp included two other speakers: Sam Witteveen and Florian Kowarsch:
- Sam's talk, "LangExtract + New Gemma", covered these two interesting new releases by Google.
- Florian's talk, "Mixture of Experts Routing", covered the increasingly dominant LLM model architectural choice of using Mixture of Experts. He described how these methods worked, and pointed to some potential future directions.
Acknowledgements
Many thanks to the Google team, who allowed us to use Google's Developer Space, and generously provided food for our audience.