What is G-Eval and how does it differ from a direct-score LLM judge?

**G-Eval** is a way to use an LLM as an evaluator that is more careful than simply asking it to rate an answer from 1 to 5. The correct option (A) captures the core idea: G-Eval first uses chain-of-thought to break a high-level rubric into explicit evaluation steps, and only then scores against those steps. Think of the difference between two graders. A direct-score judge glances at an answer and emits a number with no visible reasoning. G-Eval first writes down what to check (is it coherent, is it relevant, does it cover the key points) and then walks that checklist. Because the criteria are explicit, the score is steadier and you can audit why it came out the way it did. The wrong options describe things G-Eval does not do. It is not a multi-modal image and audio framework (B). It does not auto-calibrate its rubric from a labeled human set (C). And it does not run 100 random prompts and average them (D). The practical takeaway: G-Eval trades a little extra prompting work for scores that correlate noticeably better with human judgment.

**G-Eval** is a chain-of-thought, form-filling LLM evaluator. It differs from a naive direct-score judge in two specific moves. First, it does not ask for a number cold. It feeds the LLM a task description plus a rubric, and prompts the model to generate explicit evaluation steps via chain-of-thought. The judge writes down what to check before checking it. Second, it scores by form-filling against those generated steps, then aggregates using token probabilities. Instead of taking the single emitted integer, G-Eval reads the probability the model assigns to each score and computes a probability-weighted average. That yields finer-grained, continuous scores like 3.4 rather than a coarse 3. The explicit steps make the judgment auditable and more reproducible. The probability weighting recovers signal that a single integer throws away. The reported payoff is higher correlation with human ratings than direct scoring on summarization and dialogue benchmarks. The distractors are misconceptions. G-Eval is text-only, not multi-modal. It uses a task-derived rubric, not auto-calibration from labels. And it is a structured single pass, not a 100-prompt random ensemble.