AI Evaluation

AI evaluation targets are divided into three layers—large language model evaluation, knowledge base (RAG) evaluation, and agent (Agent) evaluation—forming a progressive relationship of increasing complexity and knowledge depth.

AI testing requires simultaneously fixing the version under test and the evaluation version; only by fixing one and varying the other does a meaningful horizontal comparison become possible.

The AI field commonly uses continuous values from 0 to 1 for scoring, and can also use the two discrete values 0 and 1 to represent the traditional fail and perfect score.

AI evaluation is not about judging right or wrong, but about expressing quality through a scoring mechanism.

AI testing introduces the concept of Evaluation, which typically uses continuous values from 0 to 1 for scoring, where evaluation is responsible for scoring and testing is responsible for setting the passing threshold.

For claimed breakthroughs in decrypting long-unsolved ciphers, the cryptographic community typically requires public disclosure of methodology and verifiable results.

The core transformation in AI testing lies in moving from 'assertions' (Test) to 'evaluation' (Evaluation), which is an internationally accepted practice.

AI evaluation bears a high degree of similarity to performance testing: neither focuses particularly on the correctness of a single result, but instead identifies problems and draws conclusions through a set of data and multiple rounds of comparison.

Evaluators need to distinguish nominal oversight from effective oversight: the former requires only drawing a human node on a process diagram, while the latter requires verifying whether humans can substantively change outcomes under real workloads, information conditions, and organizational incentives.

In AI evaluation, 'auditable' means that the inputs, outputs, intermediate states, and system interventions of every decision are fully recorded, and the records can be independently verified.

In order to test model performance in real-world scenarios, the evaluation process often requires the model to access the live internet, for example to retrieve information, call online APIs, or execute multi-step web operations.

Probabilistic systems must be evaluated through repeated sampling and statistical distribution analysis; the traditional approach of single-run, single-point assertions cannot reliably capture their true quality.

In the current industry consensus, human evaluation is regarded as the gold standard.

Measures for improving evaluation include strengthening environment isolation and protection, introducing integrity audit mechanisms, and designing evaluation protocols that are harder to exploit (such as dynamically generating test cases and hiding verification details)

The version under test is jointly determined by the model used, model parameters, prompts, and context; a change in any one of these constitutes a change in the version under test.

A meaningful side-by-side comparison can only be made when replacing one version at a time; changing two versions simultaneously eliminates the value of comparison.

The version of the system under test includes not only the software version number, but also the model type, parameter configuration (temperature, top_p, etc.), prompt strategy, and context information.

AI laboratories will need in the future to find a new balance between the authenticity and the safety of evaluations, either by building high-fidelity simulated network environments or by introducing more fine-grained permission tiering and real-time monitoring mechanisms.