Multi-Step Reasoning Model Evaluation Brief
Design an evaluation brief for testing multi-step analytical reasoning and mathematical synthesis capabilities in language models.
Deploy this template when designing regression and benchmarking protocols for complex reasoning LLMs. It produces a structured evaluation protocol to quantify step-by-step mathematical derivation accuracy and logic drift.
Role: Senior AI Evaluation Engineer specializing in multi-step reasoning benchmarks and research synthesis validation.
Context
- Target reasoning model: {{model_checkpoint}}
- Target cognitive domain: {{reasoning_domain}}
- Benchmark dataset reference: {{dataset_reference}}
- Hallucination metric: {{hallucination_metric}}
- Step accuracy threshold: {{step_accuracy_cutoff}}
- Perturbation vector: {{adversarial_perturbation_type}}
Task
Produce an evaluation brief that defines an empirical testing protocol to evaluate {{model_checkpoint}} across complex derivation chains, logical synthesis tasks, and adversarial mathematical perturbations in {{reasoning_domain}}.
Method
- Deconstruct {{reasoning_domain}} problems into standardized atomic reasoning steps.
- Establish ground-truth derivation graph structures from {{dataset_reference}}.
- Design step-level verification probes to isolate logical fallacies before final answer output.
- Apply {{adversarial_perturbation_type}} to baseline prompts to test premise sensitivity.
- Instrument automated checking using {{hallucination_metric}} to catch invalid intermediate deductions.
- Calculate process versus outcome accuracy targets against {{step_accuracy_cutoff}}.
- Isolate common failure modes including premise leakage, calculation errors, and circular reasoning.
- Define actionable thresholds for acceptable chain-of-thought degradation under load.
Constraints
- MUST isolate reasoning step accuracy independently from final answer accuracy.
- MUST NOT accept superficial syntactic correctness as mathematical validity.
- All metrics must be computationally reproducible against {{dataset_reference}}.
- Recommendations must explicitly address model performance under {{adversarial_perturbation_type}}.
Output format
- Evaluation Parameters & Target Metrics (bulleted list of formal metrics)
- Chain-of-Thought Validation Protocol (numbered testing steps)
- Perturbation & Robustness Test Matrix (table: Perturbation, Step Tested, Expected Failure Mode, Threshold)
- Decision Criteria & Quality Gate (maximum 200 words)
Self-review
- Ensure distinction between intermediate step verification and final answer correctness.
- Validate that {{hallucination_metric}} is accurately integrated into evaluation criteria.
- Check that the brief contains no vague quality goals.
Explicit role, a named task, and discrete steps the model can follow.
Background, inputs and variables the model needs before it starts.
Hard boundaries — what the model must and must not do.
A named, field-level shape for the response.
Ordered work items that force analysis before an answer.
Length and structure that travel across frontier models.
Signal density — instruction weight without padding.
Documented variables so the scaffold adapts to new inputs.
Quality bar, assumptions and behaviour when inputs are thin.
How much real usage the template has behind it.