AI agent benchmarks are deceptive, research warns

We need to hear from you! Take our fast AI survey and share your insights on the present state of AI, the way you’re implementing it, and what you count on to see sooner or later. Learn More


AI brokers have gotten a promising new analysis route with potential functions in the actual world. These brokers use basis fashions reminiscent of giant language fashions (LLMs) and imaginative and prescient language fashions (VLMs) to take pure language directions and pursue complicated targets autonomously or semi-autonomously. AI brokers can use varied instruments reminiscent of browsers, search engines like google and code compilers to confirm their actions and motive about their targets. 

Nonetheless, a recent analysis by researchers at Princeton University has revealed a number of shortcomings in present agent benchmarks and analysis practices that hinder their usefulness in real-world functions.

Their findings spotlight that agent benchmarking comes with distinct challenges, and we are able to’t consider brokers in the identical approach that we benchmark basis fashions.

Price vs accuracy trade-off

One main subject the researchers spotlight of their research is the dearth of price management in agent evaluations. AI brokers may be far more costly to run than a single mannequin name, as they typically depend on stochastic language fashions that may produce completely different outcomes when given the identical question a number of instances. 


Countdown to VB Remodel 2024

Be part of enterprise leaders in San Francisco from July 9 to 11 for our flagship AI occasion. Join with friends, discover the alternatives and challenges of Generative AI, and discover ways to combine AI functions into your trade. Register Now


To extend accuracy, some agentic techniques generate a number of responses and use mechanisms like voting or exterior verification instruments to decide on the perfect reply. Generally sampling a whole lot or 1000’s of responses can improve the agent’s accuracy. Whereas this method can enhance efficiency, it comes at a major computational price. Inference prices usually are not at all times an issue in analysis settings, the place the purpose is to maximise accuracy.

Nonetheless, in sensible functions, there’s a restrict to the price range accessible for every question, making it essential for agent evaluations to be cost-controlled. Failing to take action could encourage researchers to develop extraordinarily expensive brokers merely to high the leaderboard. The Princeton researchers suggest visualizing analysis outcomes as a Pareto curve of accuracy and inference price and utilizing strategies that collectively optimize the agent for these two metrics.

The researchers evaluated accuracy-cost tradeoffs of various prompting strategies and agentic patterns launched in numerous papers.

“For considerably comparable accuracy, the price can differ by nearly two orders of magnitude,” the researchers write. “But, the price of working these brokers isn’t a top-line metric reported in any of those papers.”

The researchers argue that optimizing for each metrics can result in “brokers that price much less whereas sustaining accuracy.” Joint optimization may also allow researchers and builders to commerce off the mounted and variable prices of working an agent. For instance, they’ll spend extra on optimizing the agent’s design however scale back the variable price through the use of fewer in-context studying examples within the agent’s immediate.

The researchers examined joint optimization on HotpotQA, a well-liked question-answering benchmark. Their outcomes present that joint optimization formulation gives a option to strike an optimum stability between accuracy and inference prices.

“Helpful agent evaluations should management for price—even when we in the end don’t care about price and solely about figuring out modern agent designs,” the researchers write. “Accuracy alone can’t establish progress as a result of it may be improved by scientifically meaningless strategies reminiscent of retrying.”

Mannequin improvement vs downstream functions

One other subject the researchers spotlight is the distinction between evaluating fashions for analysis functions and creating downstream functions. In analysis, accuracy is commonly the first focus, with inference prices being largely ignored. Nonetheless, when creating real-world functions on AI brokers, inference prices play a vital position in deciding which mannequin and method to make use of.

Evaluating inference prices for AI brokers is difficult. For instance, completely different mannequin suppliers can cost completely different quantities for a similar mannequin. In the meantime, the prices of API calls are usually altering and would possibly differ based mostly on builders’ choices. For instance, on some platforms, bulk API calls are charged in a different way. 

The researchers created a website that adjusts mannequin comparisons based mostly on token pricing to handle this subject. 

In addition they performed a case research on NovelQA, a benchmark for question-answering duties on very lengthy texts. They discovered that benchmarks meant for mannequin analysis may be deceptive when used for downstream analysis. For instance, the unique NovelQA research makes retrieval-augmented technology (RAG) look a lot worse than long-context fashions than it’s in a real-world state of affairs. Their findings present that RAG and long-context fashions have been roughly equally correct, whereas long-context fashions are 20 instances dearer.

Overfitting is an issue

In studying new duties, machine studying (ML) fashions typically discover shortcuts that permit them to attain effectively on benchmarks. One distinguished sort of shortcut is “overfitting,” the place the mannequin finds methods to cheat on the benchmark exams and gives outcomes that don’t translate to the actual world. The researchers discovered that overfitting is a major problem for agent benchmarks, as they are typically small, sometimes consisting of only some hundred samples. This subject is extra extreme than data contamination in coaching basis fashions, as information of take a look at samples may be instantly programmed into the agent.

To deal with this drawback, the researchers recommend that benchmark builders ought to create and hold holdout take a look at units which can be composed of examples that may’t be memorized throughout coaching and may solely be solved via a correct understanding of the goal process. Of their evaluation of 17 benchmarks, the researchers discovered that many lacked correct holdout datasets, permitting brokers to take shortcuts, even unintentionally. 

“Surprisingly, we discover that many agent benchmarks don’t embody held-out take a look at units,” the researchers write. “Along with making a take a look at set, benchmark builders ought to think about protecting it secret to forestall LLM contamination or agent overfitting.”

In addition they that several types of holdout samples are wanted based mostly on the specified stage of generality of the duty that the agent accomplishes.

“Benchmark builders should do their finest to make sure that shortcuts are unattainable,” the researchers write. “We view this because the duty of benchmark builders slightly than agent builders, as a result of designing benchmarks that don’t permit shortcuts is way simpler than checking each single agent to see if it takes shortcuts.”

The researchers examined WebArena, a benchmark that evaluates the efficiency of AI brokers in fixing issues with completely different web sites. They discovered a number of shortcuts within the coaching datasets that allowed the brokers to overfit to duties in ways in which would simply break with minor adjustments in the actual world. For instance, the agent may make assumptions concerning the construction of net addresses with out contemplating that it’d change sooner or later or that it might not work on completely different web sites.

These errors inflate accuracy estimates and result in over-optimism about agent capabilities, the researchers warn.

With AI brokers being a brand new area, the analysis and developer communities have but a lot to study easy methods to take a look at the bounds of those new techniques that may quickly turn into an vital a part of on a regular basis functions.

“AI agent benchmarking is new and finest practices haven’t but been established, making it onerous to differentiate real advances from hype,” the researchers write. “Our thesis is that brokers are sufficiently completely different from fashions that benchmarking practices must be rethought.”