Improving Model Performance

There are several common ways to improve performance:

• Prompting and instructions: Give the model clearer directions, examples, constraints, or context.
• Retrieval and context: Give the model access to the information it needs, such as company documents, policies, databases, or customer data.
• Tools and workflow design: Give the model tools it can use to complete tasks, or change the steps it follows within the application.
• Model selection and routing: Use a stronger model for difficult tasks or route different types of work to different models.
• Inference settings: Give the model more reasoning time or computation when the provider supports it.
• Post-training: Train the model to develop or strengthen behaviors and capabilities that are difficult to achieve reliably through the surrounding system alone.

These approaches are frequently combined. An enterprise agent might use retrieval for internal knowledge, tools for taking actions, routing for controlling cost, and a post-trained model for the specialized reasoning required by the workflow.

Start with the failures in your evaluation.

Look for patterns in when and why the system breaks down. A model that lacks access to the right information has a different problem from one that has the information but consistently reasons about it incorrectly.

Common failure sources include:

• Missing or unclear instructions
• Missing knowledge or context
• Weak reasoning
• Poor tool selection or tool use
• Failure to follow a multi-step process
• Weak domain knowledge
• Inconsistent application of policies or standards
• Difficulty recognizing when to ask for help or escalate
• Poor output formatting or structure

Breaking performance into specific failure modes makes it easier to choose an intervention and measure improvement.

Surge can help teams design evaluations that identify these failure modes, analyze model behavior, and turn the results into a concrete improvement plan.

Prompting is often a useful first intervention when the model already appears capable of doing the task but does not consistently understand what is expected.

Changes might include clearer instructions, better examples, explicit decision rules, more structured inputs, or a better division of responsibilities between the model and surrounding software.

System design can also have a large effect. A workflow may improve by breaking a complex task into stages, adding verification steps, giving the model better tools, or requiring approval before high-risk actions.

These approaches are relatively easy to test because you can compare the new system against the previous version on the same evaluation.

Retrieval is useful when performance depends on information the model does not reliably have in its parameters.

That might include:

• Internal policies
• Product documentation
• Customer information
• Proprietary research
• Frequently changing data
• Large collections of business documents

A retrieval system finds relevant information and provides it to the model as context when it handles a request.

Retrieval works particularly well for knowledge problems. It is less likely to solve a capability problem. If the model receives the correct information but still cannot analyze it properly or complete the required workflow, another intervention may be needed.

It is important to keep in mind that these retrieval systems should also be rigorously evaluated to ensure that the model both has access to the necessary data and also can utilize the information effectively.

Sometimes the right improvement is simply selecting a stronger model or routing harder requests to one. See Comparing and Choosing Models for more.

Post-training is the process of improving a pretrained model after its initial training is complete.

It can be used to strengthen existing capabilities, teach new behaviors, improve instruction following, specialize a model for a domain, or make an agent more effective at completing particular kinds of work.

Post-training can use several methods, including supervised fine-tuning, preference-based training, reinforcement learning, and training in interactive environments.

The appropriate method depends on the behavior you want to change and the kind of feedback you can provide.

Post-training becomes especially useful when you need to improve a recurring behavior or capability that prompting, retrieval, and workflow changes do not address reliably.

Examples might include:

• Following complicated instructions
• Performing specialized professional work
• Planning across long workflows
• Using tools correctly
• Applying organization-specific standards
• Producing consistently high-quality outputs
• Recovering from mistakes during multi-step tasks

Before investing in post-training, you should have a clear picture of the behavior you want to improve and an evaluation that can measure it as it can be a very expensive endeavor.

Surge can work with your teams to identify which capabilities are good candidates for post-training, design the evaluation program, determine what training signal is needed, and test whether improvements generalize beyond the training examples.

Supervised fine-tuning, often called SFT, trains a model on examples of the behavior you want it to produce.

Each training example typically contains an input and a high-quality response or trajectory. The model learns from those demonstrations.

For an enterprise application, this might include examples of strong financial analyses, code changes, customer interactions, research tasks, or multi-step workflows.

The quality of the demonstrations matters substantially. For specialized work, strong examples often need to come from people who understand the domain and can demonstrate how an expert would approach the task.

Reinforcement learning (RL) improves a model using feedback about the quality of its behavior.

The feedback might come from an automated verifier, a rubric-based grader, a reward model, the outcome of an environment, or another reliable signal of success.

This can be especially useful for tasks with many possible valid solutions or for agents completing multi-step workflows. Instead of teaching the model one exact response, the training process rewards behavior that leads to successful outcomes.

Reward design matters. The signal used during training should closely reflect the behavior you ultimately care about, which makes evaluation design and training design closely connected.

A training environment gives a model a setting where it can take actions, observe the results, and receive feedback.

This is particularly useful for agents that work with software, browsers, files, terminals, APIs, or other tools.

For example, an agent might be asked to resolve a customer account issue inside a simulated enterprise application. It can inspect records, use available tools, take actions, and receive a score based on whether it completed the task correctly.

High-quality environments make it possible to train full workflows and behaviors that are difficult to capture in isolated question-and-answer examples.

Surge has extensive experience in designing RL training environments for many advanced domains with highly discriminatory reward signals. We’re happy to discuss how RL environments might fit into your post-training workflows.

Good training data represents the behavior you want the model to learn and the situations where it needs to perform.

Useful datasets often contain:

• Realistic tasks and workflows
• High-quality demonstrations
• Difficult examples
• Important edge cases
• Examples targeting known failure modes
• Reliable feedback about what constitutes a good result
• Enough diversity for the behavior to generalize beyond the training set

For complex enterprise work, data quality often depends on the people creating it. A fluent answer can still contain a subtle legal, financial, scientific, or technical error. Subject-matter experts can create demonstrations and feedback that reflect how the work should actually be done.

Evaluation and model improvement should operate as a continuous loop:

1. Evaluate the current system.
2. Identify important failure modes.
3. Determine what is causing them.
4. Choose an intervention.
5. Create any data or feedback needed.
6. Improve the system or model.
7. Evaluate again on held-out tasks.
8. Look for improvements and regressions.

This keeps training focused on measurable problems and makes it possible to determine whether an intervention actually improved the underlying capability.

We can support this full process, from designing the initial evaluations and diagnosing failures to building expert training data and environments and measuring the results of each iteration.

Run the improved system against evaluations that were not directly used to make the change.

Look at overall performance, but also examine the specific capabilities and failure modes you intended to improve.

Teams should ask:

• Did the targeted failures become less common?
• Did the improvement generalize to new examples?
• Did performance improve across relevant task categories?
• Did any previously strong capabilities get worse?
• Did cost or latency change?
• Did the severity of the remaining failures change?

Held-out evaluation is particularly important after post-training because models can improve on the training distribution without developing a capability that generalizes to new tasks.

Surge works with teams across the model improvement lifecycle.

Depending on the problem, that can include:

• Designing evaluations around real enterprise workflows
• Diagnosing model failures and identifying capability gaps
• Determining which failures are best addressed through system changes or training
• Designing post-training experiments
• Creating expert demonstrations and preference data
• Building detailed rubrics and reward signals
• Developing realistic training and reinforcement learning environments
• Evaluating trained models on held-out tasks
• Measuring whether improvements transfer to new workflows and benchmarks

The goal is to connect model improvement to measurable outcomes, so teams can understand what changed, why it changed, and whether the investment was worthwhile.

We regularly publish research on model evaluation, post-training, reinforcement learning, training data, and how different capabilities can be improved.

Our research explores questions such as which training signals produce meaningful gains, how capabilities transfer across tasks, how expert rubrics can support training, and how realistic environments can improve agent performance.

Visit surgehq.ai/research to read the latest work from the Surge research team and our research collaborations.