New Benchmark Reveals AI Agents Struggle with Long-Term Decision Making

Researchers from Microsoft and City University of Hong Kong have established a new framework to evaluate AI agents' decision-making capabilities over long time horizons. The findings indicate that even the most advanced models reach an accuracy of just 60 percent on a series of complex decision-making tasks.

AI agents are increasingly relied upon for multi-step operations, such as drafting research papers or developing software projects. These agents must navigate a series of choices, deciding on which hypotheses to test or which modifications to implement. The repercussions of these decisions often manifest only after the agents have expended significant resources. As a result, a poor choice may not be immediately apparent, leading to difficulties in evaluating long-term strategies.

Introducing Taste-Bench: A New Evaluation Method

An agent's ability to make sound decisions is referred to as its "taste." Traditional benchmarks gauge whether an agent successfully completes a task but do not assess the integrity of the steps taken throughout the process. To address this gap, the researchers introduced Taste-Bench. This benchmark allows for the evaluation of an agent's decision-making quality in a structured manner.

Long-term decisions cannot be assessed in real-time; a clean implementation and a flawed one may yield identical outputs in the short term. The advantages of a well-structured approach typically surface only when modifications must be made. Assessing these decisions thoroughly often requires significant human expertise, which is not scalable. Thus, the researchers opted to utilize hindsight as a measure of decision quality.

In practice, AI systems frequently revisit tasks, resulting in divergent paths due to various choices made during execution. To quantify an agent's decision-making taste, researchers identify "decision forks," where agents must choose between two potential paths. By analyzing the agent's trajectory up to this point, researchers can conceal future results and assess the agent's ability to select the preferable option.

Understanding Decision Forks and Prediction Challenges

Taste-Bench involves a total of 502 decision-making scenarios, including the original task, the trajectory up to the decision fork, and the two divergent paths. The first scenario type comes from independent attempts at a task where the agent has not encountered the unsuccessful path. The second scenario involves an agent correcting its course within a single run by exploring a previously disregarded option.

The robustness of this methodology is bolstered by utilizing thousands of recorded agent runs from coding and AI research tasks, such as SWE-bench Pro and METR's public transcript release MALT. A filtering generator model curated the questions to focus on meaningful decision forks, and human reviewers validated the auto-generated labels.

Current Models Underperform Even Against Simple Benchmarks

The research team evaluated 14 state-of-the-art models, including those from the Claude, GPT, and Mistral families, among others. Each model faced questions in their natural order and again with their options reversed. A question was counted only if the model answered both iterations accurately. The baseline of random guessing provided a score of 25 percent.

The top-performing model in this study, GPT-5.6 Sol, achieved an average score of 59.7 percent, closely followed by GPT-5.5 at 59.5 percent. However, other models scored significantly lower, showcasing a consistent struggle with binary decision-making tasks.

The Impact of Time on Decision-Making

The analysis conducted by the research team emphasized the requirement for foresight in decision-making. They measured the progression of accuracy based on how far ahead an observer needed to look to discern the superior path. When clues were present in prior steps, models achieved an average accuracy of 62.3 percent. However, when the critical information appeared only after considerable further effort, accuracy plummeted to 21 percent, which is close to random guessing.

Interestingly, increasing the reasoning time did not yield significant improvements. The researchers experimented with three different reasoning settings for two models, discovering that even switching to a higher complexity level did not enhance accuracy. The models often used the most reasoning tokens on the harder tier, indicating that crucial deciding evidence lacked presence during those initial forks.

Training for Better Decision Making

Encouragingly, the research demonstrates that taste can be cultivated. They employed a teacher model with known paths to refine a student model focused solely on decision-making. Using Qwen3.6-27B as a base, the student model’s performance improved from 30 percent to 47.9 percent on unseen tasks.

Furthermore, incorporating the student's guidance at decision forks allowed a separate agent to tackle problems with increased efficiency. In a series of 41 SWE-bench Pro tasks, success rates rose from 14.6 to 33.7 percent—significantly closer to the theoretical upper limit of 39 percent.

The research concludes that while AI agents currently struggle with long-term decision-making, progress can be made in both measurement and improvement of this critical capability. The Taste-Bench benchmark and associated code are publicly accessible for further exploration.

As AI agents continue to operate autonomously for extended periods, identifying and overcoming their decision-making limitations will be crucial. Evidence from the MirrorCode benchmark suggests that many AI-generated solutions may appear functional yet be fundamentally flawed upon closer inspection, emphasizing the need for enhanced decision-making frameworks like Taste-Bench.