Evaluation & AI Reliability · core
Coding-agent Benchmarks
Issue resolution tasks, repository setup, patch grading, test-based scoring, contamination, and benchmark validity.
Mental model
A coding benchmark is an environment plus task distribution and verifier. Scores are useful only when setup, contamination, flaky tests, and patch validity are controlled. Scope: this card owns one benchmark family and its validity — repository setup, patch grading, flaky tests, and contamination. The general practice of grading agent actions is `tool-use-evaluations`; grading text output is `llm-evals`.
How to study Coding-agent Benchmarks
Begin by restating the mental model in your own words, then connect it to a concrete system you have built or operated. Name the mechanism, the constraint it addresses, and the trade-off it introduces. Use Evaluating Large Language Models Trained on Code (HumanEval / Codex), LiveCodeBench: Holistic and Contamination Free Evaluation of LLMs for Code, The SWE-Bench Illusion: When State-of-the-Art LLMs Remember Instead of Reason to check details, but close the source before writing your explanation. Retrieval is the learning step; rereading is only preparation.
Next, compare Coding-agent Benchmarks with Model Evaluation, Tool-use Evaluations. Ask what changes in correctness, latency, resource use, operability, and failure recovery. Complete Design exercise: Coding-agent Benchmarks and preserve the command, input, output, and one failed attempt as evidence. Finish by explaining the idea without jargon to someone who has not studied the track.
Proof of understanding
- Explain the mechanism from first principles and identify the state it reads or changes.
- Give one situation where the concept is the right choice and one where it is not.
- Predict a realistic failure mode before running the drill, then compare the prediction with evidence.
- Connect the result to a roadmap or build artifact instead of treating the concept as isolated trivia.
Learn from primary sources
- Evaluating Large Language Models Trained on Code (HumanEval / Codex) (doc)
- LiveCodeBench: Holistic and Contamination Free Evaluation of LLMs for Code (doc)
- The SWE-Bench Illusion: When State-of-the-Art LLMs Remember Instead of Reason (doc)
- SWE-bench: Can Language Models Resolve Real-World GitHub Issues? (paper)
- SWE-bench (doc)
Practice and explain it back
Design exercise: Coding-agent Benchmarks
Issue resolution tasks, repository setup, patch grading, test-based scoring, contamination, and benchmark validity. Implement designOutline() returning non-empty values for: taskDistribution, environment, verifier. Each value must name a concrete mechanism or decision.
Expected evidence: A design outline with taskDistribution, environment, verifier plus an explicit failure mode or trade-off.
Open the interactive drill →Review prompts
- A model scores well on a public coding benchmark. Name the two threats to that score's validity.
Build evidence
Synthesize: Evaluation & AI Reliability
Build an evidence-backed evaluation and observability system for models, tools, and agents. Produce one working system, benchmark, or evidence-backed design that integrates the path.
- Implements or precisely models the core mechanisms from all three milestones
- Includes at least one injected failure or adversarial case and demonstrates recovery
- Reports quality, latency, resource, reliability, or usability measurements relevant to the domain
- Ships a concise architecture note explaining decisions, trade-offs, and remaining risks