AI Agent Evaluation Leaderboards Explained: Causal Tracing, Key Metrics, and Practical Guidance

1. Introduction: A Paradigm Shift in Agent Evaluation

As large language models (LLMs) see deeper use in automated task execution, AI agent systems have evolved from simple Q&A tools into complex systems capable of autonomous planning, tool calling, and multi-turn interaction. However, traditional evaluation approaches—such as asking users to vote on which of two model outputs they prefer—have significant limitations: voting results are heavily influenced by subjective judgment, sample coverage is uneven, and they fail to reflect an agent’s overall performance in real, long-horizon tasks.

Against this backdrop, evaluation leaderboards built on real agent task sessions have emerged. The core idea behind these platforms is to automatically mine evaluation signals from users’ real interaction behavior and use statistical inference to quantify each model’s net improvement rate relative to the average, thereby building a more objective and reproducible evaluation system. This article systematically breaks down this technical paradigm—its methodology, key metrics, and practical implications.

2. Core Methodology: From Voting to Causal Tracing

2.1 The Bottlenecks of Voting-Based Evaluation

Traditional A/B tests or crowdsourced voting rely on human judgment, typically following this flow:

User submits a task → Model A and Model B each generate a response → User votes for the better one → Win rates are tallied

The problems with this approach:

• Bias: users tend to vote for the answer that “looks better” rather than the one that “actually completes the task better”;
• Sparse samples: for long-horizon, multi-step agent tasks, users often cannot make accurate judgments within a short time;
• Poor scalability: as the number of models grows, the number of pairwise comparisons grows quadratically, and evaluation costs rise sharply.

2.2 Causal Tracing: A Randomized Inference Approach

The new generation of evaluation methods adopts a randomized inference framework known as “causal tracing.” Its core idea is analogous to randomized controlled trials (RCTs) in medicine: different models are randomly assigned to real users’ session requests, and statistical methods are then used to isolate the causal effect of the model itself on task outcomes—rather than mere correlation.

The technical flow can be abstracted as:

1. A real user initiates an agent session request
2. The system randomly routes the request to different models (randomized assignment)
3. Multi-dimensional signals are automatically collected during the session
4. A causal inference model estimates each model's net improvement rate relative to the mean

The advantages of this approach:

• Eliminates selection bias: randomization ensures that model assignment is independent of user characteristics;
• Automatic signal extraction: no manual annotation is required, reducing evaluation costs;
• Interpretable net improvement rate: positive values indicate the model outperforms the average, while negative values indicate it underperforms—intuitive and directly comparable.

2.3 Key Signal Dimensions

The signals automatically extracted from real sessions span multiple layers:

Signal Category	Specific Content	Evaluation Significance
Task confirmation	User explicitly confirms task completion or approves the result	Measures task completion quality
Praise and complaints	Positive feedback or negative complaints from users	Reflects user satisfaction
Correction instructions	User corrects or redirects the model’s output	Measures the model’s steerability
Shell command execution errors	Code or command failures, error messages	Measures tool-calling reliability

Used in combination, these signals characterize an agent’s overall capabilities in real-world scenarios from multiple angles—far more granular than a single “good or bad” vote.

3. Key Evaluation Metrics

3.1 Overall Net Improvement Rate

The overall net improvement rate is the headline aggregate metric on the leaderboard. It represents a model’s net improvement over the average of all evaluated models across all evaluation dimensions. For example, a model with an overall net improvement rate of +5.1% means that, under randomized conditions, its task performance is on average 5.1 percentage points better than the overall mean.

The calculation relies on estimating the average treatment effect (ATE) from causal inference:

Net Improvement = E[Y | Model=i] - E[Y | All Models]

where Y is a composite measure of task success.

3.2 Bash Reliability

For agents capable of code execution, file operations, and system command invocation, the reliability of shell command execution is a critical metric. This dimension measures:

• Command syntax correctness;
• The proportion of executions that complete without errors;
• Fault tolerance and recovery under exceptional conditions.

High Bash reliability means the agent is more trustworthy in scenarios such as automated operations, data processing, and development assistance.

3.3 Steerability

Steerability measures how effectively users can adjust an agent’s direction and correct its course. It shows up in three ways:

• Whether the model quickly understands correction instructions and adjusts its execution path;
• Whether the model maintains continuous tracking of user intent across multi-turn conversations;
• Whether the model proactively clarifies ambiguous instructions instead of blindly executing them.

This metric directly affects an agent’s usefulness in complex, open-ended tasks.

4. Scale and Coverage of Evaluation Data

High-quality evaluation rests on real data of sufficient scale and diversity. For example, current mainstream evaluation platforms typically base their rankings on tens of thousands of real, long-horizon agent sessions from global users. These sessions cover:

• Work scenarios: document writing, data analysis, email drafting, schedule planning, and more;
• Coding scenarios: code generation, debugging, refactoring, technical Q&A, and more;
• Chat scenarios: open-domain conversation, knowledge queries, creative writing, and more.

The inclusion of long-horizon sessions is especially critical—an agent’s true capabilities only emerge in multi-step tasks that require sustained context management. Single-turn Q&A cannot expose weaknesses in state management, error recovery, or task planning.

5. Why Open-Weight Model Rankings Matter

In the model ecosystem, the capability comparison between open-weight and closed-source models is a focal point for the industry. Leaderboards typically maintain a separate ranking dimension for open-weight models, which matters for several reasons:

1. Promotes technology democratization: developers and enterprises gain visibility into the real capability boundaries of locally deployable models;
2. Guides selection decisions: in scenarios where data security or compliance requirements restrict the use of external APIs, open models are a viable alternative;
3. Drives community progress: transparent evaluation data provides an objective reference for model iteration in the open-source community.

From a practical standpoint, when an open-weight model consistently ranks near the top in both overall and per-scenario rankings, it usually indicates a high level of maturity in parameter efficiency, training data quality, and alignment strategy.

6. Practical Recommendations for Developers and Enterprises

6.1 Using Leaderboards for Model Selection

• Don’t rely on overall rankings alone: a model with a high overall rank is not necessarily the best fit for your specific scenario—pay attention to sub-metrics;
• Focus on net improvement rate rather than absolute scores: net improvement rate better reflects relative advantage and is better suited to side-by-side comparison;
• Check the evaluation sample size: rankings based on small samples may carry substantial statistical noise and should be referenced with caution.

6.2 Improving Your Agent’s Evaluation Performance

Based on the metric system above, developers can optimize along the following directions:

# Illustrative optimization directions (not a concrete implementation)
optimization_targets = {
    "task_completion": "Increase the share of task confirmation signals; improve output format and completeness",
    "bash_reliability": "Strengthen syntax validation for generated code and fault tolerance in sandboxed execution",
    "steerability": "Enhance instruction-following; introduce multi-turn intent tracking",
    "error_recovery": "Provide automatic retries or alternative plans when command execution fails"
}

6.3 Leveraging Cloud Infrastructure for In-House Evaluation

Building an in-house agent evaluation system also requires underlying infrastructure. Baidu AI Cloud, for example, offers a large model platform, container services, log analysis, and related capabilities that can support such an effort: deploying different models on a container platform for A/B testing, collecting session behavior signals through log services, and tracking task execution status with monitoring and alerting. Combined with the evaluation methodology described above, these components form a closed-loop capability assessment process.

7. Summary and Outlook

Agent evaluation is moving from subjective voting toward a new paradigm that is automated, causal, and large-scale. By mining signals from real sessions and applying randomized inference, evaluation leaderboards can more objectively reflect model capabilities in real-world tasks. For developers and enterprises, understanding this paradigm not only supports scientifically informed model selection but also guides the continuous optimization of agent systems. As evaluation data scales up and causal inference methods mature, agent evaluation will become an indispensable piece of infrastructure in the LLM ecosystem.