Editorial illustration for New Benchmark Shows AI Agents Fail Long-Term Binary Choice Tests
AI Agents Fail Long-Term Decision Tests, Study Shows
Researchers from Microsoft and City University of Hong Kong built a benchmark to test something AI evaluations have mostly ignored: whether an agent makes good decisions during the middle of a long task, not just whether it finishes. The best model they tested scored barely 60 percent.
The gap matters because AI agents now handle jobs that stretch over many steps, drafting a research paper, building a software project, running a string of experiments. Each step involves a choice. Which hypothesis to chase.
Which code to build on. Which experiment to run next. None of those choices announce their quality right away.
A bad call can look identical to a good one for a long stretch, and the damage only surfaces once the agent has already spent most of its time and budget chasing it.
The researchers call this skill an agent's "taste," the ability to pick the option that pays off later rather than the one that just looks fine now. Standard benchmarks check for task completion. They don't check whether the agent got there by a smart route or a lucky one. Taste-Bench was built to measure that difference, and the gap between those two outcomes.
Researchers from Microsoft and City University of Hong Kong built a method to test how well AI agents make decisions during long-running tasks. The best model barely hit 60 percent.
Why this matters
A coin flip beats the best model here by a slim margin, which should worry anyone deploying agents on tasks that run longer than a single prompt-response loop. The 60 percent ceiling isn't a rounding error, it's a sign that current models can't reliably tell a promising research direction from a dead end until the evidence is already obvious, long after the decision mattered. The detour-fork result is the part worth sitting with: models do fine when the right path runs in parallel with the wrong one, but fall apart when the better option requires a visible step backward first.
That's exactly the kind of judgment call developers assume agents are making competently when they let one run unattended on a coding project or a multi-day research task. For founders building agent products, this is a concrete reason to keep humans checking in at decision forks rather than trusting an agent's sense of which path to commit to. For researchers, the benchmark from Microsoft and City University of Hong Kong gives a clean, measurable target, one that's apparently far from solved by scaling alone.
Related Reading
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives AI models hide mistakes in work reports, but one prompt sentence can help
From Archives India reports rogue AI agents using leftover access, CERT-In says no hacking
Common Questions Answered
What did researchers from Microsoft and City University of Hong Kong discover about AI agents in their benchmark test?
The researchers found that the best-performing AI model they tested scored barely 60 percent on their benchmark designed to evaluate decision-making during long-running tasks. This benchmark specifically measures whether AI agents make good decisions in the middle of extended tasks, not just whether they complete them, revealing a significant gap in current AI evaluation methods.
Why is the 60 percent performance ceiling concerning for AI agents handling multi-step tasks?
The 60 percent score means that current AI models can barely outperform a coin flip when making decisions during long-term projects like drafting research papers or building software projects. This indicates that models cannot reliably distinguish between promising research directions and dead ends until evidence becomes obvious, long after the critical decision point has passed.
What types of real-world tasks are affected by AI agents' poor long-term decision-making abilities?
AI agents that handle jobs stretching over many steps are significantly impacted, including drafting research papers, building software projects, and running strings of experiments. Each step in these tasks involves choices that require reliable decision-making throughout the process, not just at the end.
How does the detour-fork result demonstrate AI agents' limitations in long-term planning?
The detour-fork result shows that AI models perform adequately when the correct path runs in parallel with alternatives, but struggle when they must commit to a single direction early on. This reveals that models cannot effectively evaluate which path will lead to success when multiple options exist, a critical capability for complex multi-step tasks.
Further Reading
- The Long-Horizon Task Mirage? Diagnosing Where and Why Agentic Systems Break - arXiv
- KellyBench: A Benchmark for Long-Horizon Sequential Decision Making - arXiv
- OSWorld 2.0: Benchmarking Computer-Use Agents on Long-Horizon Real-World Tasks - OSWorld
- The Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks - arXiv
- LH-Bench: Skill-Grounded Evaluation of Long-Horizon Agents - OpenReview
