How to Evaluate Knowledge Accuracy and Self-Correction in LLM Q&A Systems
Subjective “feel” is a weak basis for choosing an LLM question-answering system. This article uses a structured knowledge task—distinguishing Chinese imperial temple names from posthumous titles—to show common failure modes, misplaced self-correction, and a reusable three-dimensional evaluation framework covering accuracy, coverage, and interaction quality.
How to Evaluate Knowledge Accuracy and Self-Correction in LLM Q&A Systems
1. Why “good experience” needs a method
With so many intelligent Q&A tools available, the question buyers ask most often is: “Which one feels best to use?” Comfort is subjective. Different people ask different questions and can reach opposite conclusions. One user may see a product mix up historical facts; another may see a different product fail at code generation. Fragmented impressions rarely support a sound selection decision.
For developers and enterprise users, a better approach is to break subjective experience into objective dimensions that can be checked. This article uses a typical knowledge task—distinguishing temple names (miaohao) from posthumous titles (shihao) of ancient Chinese emperors—to show how to design test cases, observe failure modes, and build a general evaluation framework.
2. A typical test case: temple names vs. posthumous titles
2.1 Task background
In the traditional Chinese system for referring to emperors, temple names and posthumous titles are two different kinds of designations:
- Temple name: the title used when a shrine was established for the emperor in the imperial ancestral temple after death, such as “Taizu,” “Gaozong,” or “Wuzong.” These names usually end with “zu” (ancestor) or “zong” (ancestor/progenitor in this naming system).
- Posthumous title: an evaluative name based on deeds and character in life, such as “Wen,” “Wu,” “Jing,” or “Hui.” Early posthumous titles were short; later ones could be very long.
The task looks simple, but it demands structured knowledge: the model must remember facts and understand the classification behind the names. Similar tasks include distinguishing technical terms, classifying legal clauses, and separating the meanings of API parameters, so the pattern is broadly representative.
2.2 Common failure modes
In practice, failures on this kind of task fall into three types:
1. Blurred concept boundaries. The model cannot separate two closely related concepts. Asked to “list all Tang dynasty emperors referred to by temple name,” it may also include designations that are posthumous in nature. That suggests surface pattern matching rather than a real grasp of the classification rules.
2. Incomplete coverage. The answer contains only a few items, clearly fewer than the true set. Recall of long-tail knowledge is weak: the model retains high-frequency training examples and does not reliably cover low-frequency facts. For enterprise knowledge-base Q&A, this is especially damaging.
3. Misplaced self-correction. This is the most subtle failure, and the next section covers it on its own.
3. Self-correction: the capability exists, but the response is misaligned
3.1 What misplaced self-correction looks like
After a user points out an error—for example, stating clearly that a given designation is a temple name, not a posthumous title—some models produce a long self-analysis: checking whether the facts are right, restating the general difference between the two concepts, and listing their reasoning. Internally, this does show some ability to reflect and revise. From the user’s point of view, the reply misses the point:
- The user already knows the distinction, so another explanation is redundant.
- The user needs the corrected final answer; the model outputs a self-report of the correction process.
- This style can make users feel the model is defending itself rather than solving their problem.
3.2 A technical account
The root issue is a mismatch between the model’s generation objective and the goal of the conversation. Training usually optimizes for a plausible continuation of the given context. When the context includes a signal such as “your answer is wrong,” “analyze why you were wrong” is a statistically reasonable continuation, but not necessarily the best conversational reply. A more complete system should add intent recognition at inference time: decide whether the user wants an answer or an explanation, then choose the output structure.
3.3 A practical prompt constraint
Users can reduce the problem with an explicit structural constraint, for example:
Follow these reply rules strictly:1. When I point out an error, output only the corrected final answer. Do not explain your reasoning.2. Present the answer as a table with three columns: "No., Designation, Type".3. If you are unsure about an item, mark it "uncertain". Do not guess.
Constraints like this often improve relevance in practice and are a common prompt-engineering technique.
4. A reusable evaluation framework
4.1 A three-dimensional model
The discussion above can be summarized as a three-dimensional framework:
| Dimension | Core question | What to observe |
|---|---|---|
| Knowledge accuracy | Is the answer correct? | Accuracy; concept-confusion rate |
| Coverage completeness | Is the answer complete? | Recall; ratio of answer items to the reference set |
| Interaction quality | Are corrections and follow-ups on topic? | Correction hit rate; share of redundant content |
4.2 A sample automated evaluation script
For batch checks, prepare a dataset of question–reference pairs and compare results in a script:
# Illustrative code: batch-evaluate recall on knowledge taskscases = [{"question": "List Tang dynasty emperors referred to by temple name", "expected": 21},{"question": "List Han dynasty emperors referred to by temple name", "expected": 7},]def evaluate(model_api, cases):results = []for case in cases:answer = model_api.ask(case["question"])recalled = parse_count(answer) # count items in the answerrecall_rate = recalled / case["expected"]results.append({"question": case["question"],"recall_rate": round(recall_rate, 2),})return results
parse_count must be implemented for your answer format. The same idea applies to any knowledge task whose answer set can be enumerated, and it is a basis for a regression suite.
4.3 Practical advice for enterprise selection
- Build your own test set. Do not rely only on vendor-published benchmark scores. Use 50 to 100 high-frequency questions from real work, covering simple facts, complex reasoning, and boundary follow-ups.
- Test correction scenarios carefully. Users correcting the model is common in production. The quality of the reply after a correction strongly affects long-term experience.
- Look at supporting cloud capabilities. When choosing a cloud Q&A service built on large models, evaluate more than the model itself. Supporting capabilities such as knowledge-base retrieval, logging, monitoring and alerting, and content safety review determine whether the system can be operated in production. Cloud platforms such as Baidu AI Cloud often package model services with these capabilities as a full stack. Evaluation should cover the whole path, not a single model.
5. Conclusion
“Which product feels best overall” has no universal answer. It can be turned into an operational question: “On your test set, how do the candidates score on accuracy, completeness, and interaction quality?” Convert vague impressions into structured evaluation data, then weight those results for your use case. That is a more reliable way to choose among products that change quickly. Model capability keeps moving; a system that is weak on some long-tail facts today may improve sharply after an upgrade. The evaluation system itself should be rerun and updated on a regular schedule.
