Welcome to the 10th issue of Agents in Practice!
📬 Agents in Practice is a weekly newsletter on agentic AI research and applications. Subscribe here →
Fragility of Self-Improving Agents
Memory-based self-improving agents are agents that autonomously improve their performance over time by learning from experience. The authors study the reliability of such systems (Agent Workflow Memory and ReasoningBank) on 3 web browsing benchmarks (WebArena, VisualWebArena, SCUBA). The authors find that the self-improvement loops increase the already-large variance across agent runs. They also show that the system is sensitive to task order, showing that an agent that achieved a performance gain of 1.5% on the default task order showed a performance loss of 4.5% with a randomly shuffled task order.
Personal Thoughts
Self-improving agents have always been a topic of interest, perhaps due to the attractive idea of agents getting smarter and smarter and eventually reaching “singularity”. For now, though, human guidance seems necessary.
Read more
What does the benchmark measure?
Researchers from Ai2 propose a new method called BenchMIRT for understanding whether a benchmark tests general reasoning ability or safety. The method involves running 100 LLMs across 16 reasoning and safety benchmarks. The LLM responses are used to fit a multidimensional item response theory (MIRT) model with two dimensions, which shows two latent dimensions for each benchmark. The two latent dimensions are learned without being predefined to measure general reasoning and safety, but are validated experimentally. The authors show that the majority of the benchmarks are aligned with the stated goal. However, the authors do show that some safety benchmarks have at least some misaligned subsets that test general reasoning instead of safety.
Personal Thoughts
While preparing for this newsletter, I see a lot of news related to agents, and nearly half of it is about new benchmarks. There are a lot of new benchmarks on every aspect of agents, claiming both similar and different capabilities. An objective analysis of those benchmarks like this could help us better understand what performance on a benchmark reveals.
Read more
One-liners
- OpenAI claims it has solved the Navier-Stokes problem, one of the seven Millennium Prize Problems, but faced controversy over whether it could have used usage logs from mathematicians Tristan Buckmaster and Levent Alpöge who had been working on their own results.
- Perplexity added a feature to its Mac app that natively splits tasks between cloud models and a local model.
- GameXpert-Bench from Tencent tests the agents’ abilities to create a game from scratch, fix defects, and refine the game.
- The authors of Thinkingbox provide a reusable sandbox for tool-agent-user interaction and claim that Pass@1 is not a reliable metric due to considerable variance in agent success across runs.
- ASPIRE is a benchmark that tests whether an agent can self-improve when given only a vague goal.
Subscribe to Agents in Practice
Get new issues of Agents in Practice — a newsletter summarizing exciting new research and applications in agentic AI — delivered to your inbox.