All whitepapers
Whitepaper

How the AI Interviewer Grades Candidate Answers

Dassh AIJuly 2026

Many HR professionals ask: how does Stella actually rate a candidate's answers? This white paper provides a high-level overview of the grading mechanism used by the Dassh AI Interview platform. We explain how the AI evaluates responses by comparing them to an ideal answer using semantic vector similarity. The process is data-driven and standardized, ensuring consistent scoring for every candidate.

Semantic Vectorization of Answers

To evaluate an answer's quality, Stella must first understand the answer's meaning. This is achieved by converting text responses into numerical embedding vectors — points in a high-dimensional space that capture the answer's semantic content. This technique, known as vectorization, uses advanced language models (Transformer-based NLP models) to represent the answer's meaning rather than just its exact words.

Capturing Meaning

In this vector space, answers with similar meaning end up close together, even if they use different wording. For example, "I led a marketing project" and "I spearheaded a promotional campaign" would produce vectors that are nearby, reflecting their similar intent. Stella looks beyond buzzwords or specific phrasing to the actual substance of the answer. Candidates get credit for conveying the right idea, even if their wording differs from the ideal.

Role of Pre-trained Models

Stella leverages large pre-trained language models to generate these embeddings. Using pre-trained Transformer models to vectorize both the expected answer and the candidate's answer allows a meaningful comparison of their content. The AI transforms free-form text into a form it can quantitatively analyze.

Comparing Answers to an Ideal Response

For each interview question, the system has a benchmark ideal answer — effectively, the best possible response given the question and role requirements. Once both the candidate's answer and the ideal answer are in vector form, Stella computes a similarity score to measure how closely they match in meaning.

Vector Distance (Semantic Similarity)

The distance between two vectors corresponds to the difference in meaning between the two answers. A smaller distance (higher cosine similarity) means the candidate's response closely mirrors the ideal. A larger distance means divergence from what was expected.

In practice, Stella calculates cosine similarity between the candidate's answer vector and the ideal answer vector. For example, a cosine similarity of 0.95 indicates a very close match, whereas 0.5 indicates a moderate match in content.

Interpreting the Similarity Score

A score near 1.0 signifies that the candidate's answer has almost the same meaning as the ideal answer, whereas a score near 0 means the response is semantically far off. The more the candidate's answer aligns with the expected key points and facts, the higher this similarity metric will be.

Example

If the question is "Describe a time you solved a difficult problem at work," and the ideal answer highlights identifying the root cause, taking initiative, and collaborating with a team, then a candidate who mentions those same core elements will yield a high similarity score. Stella recognizes the overlap in meaning (even if phrased differently). If a candidate veers off-topic or misses crucial aspects, the vectors will be farther apart, resulting in a lower score.

Translating Similarity into a Rating

Once Stella has a similarity score, it translates that score into a rating on a human-readable scale (for example, 0 to 100 points). This step maps the raw similarity measure to an intuitive evaluation that hiring teams can understand at a glance.

Score Mapping

The system uses a predefined mapping algorithm to convert vector similarity into a rating. High similarity corresponds to a strong rating; low similarity corresponds to a weaker rating. The mapping may involve sophisticated calibration. The key point is that there is a consistent rule set for turning "% match to ideal" into a score.

Calibration and Benchmarking

To ensure fairness, this translation is calibrated using benchmarks — sample answers of various quality levels, expert-defined standards, or company-specific requirements. The ratings are backed by benchmarks so a given level of similarity will reliably correspond to the same rating every time.

Multi-Dimensional Considerations

Stella scores answers on multiple dimensions (e.g. content quality, communication, problem-solving skill). Semantic similarity on content is one major factor, and additional ratings are given for other aspects. All sub-scores are data-driven and standardized.

Aggregating Answer Ratings into an Overall Score

After each individual answer is evaluated and rated, Stella combines these results to produce an overall assessment. Every candidate gets a comprehensive score based on all their answers, ensuring the evaluation is well-rounded and fair.

Weighted Combination

The overall rating is derived by aggregating individual answer ratings. If a candidate consistently scores highly, their overall interview score will be high. The system assigns weights if some competencies are more crucial for the role, predefined according to job requirements.

Balanced Evaluation

Using multiple answer scores prevents one poor answer from fully dictating the outcome. Stella provides a multi-score breakdown and then a composite view. For instance, a candidate might excel in some skills but be average in others — the overall score reflects all of these, with a report noting strengths and weaknesses.

Final Recommendation

The aggregated score is used to rank or categorize candidates (for example, "Highly Recommended," "Recommended," or "Not Recommended" for next stages). The final rating is presented along with explanations — e.g., "Overall Score: 82/100 — Candidate demonstrated strong problem-solving and teamwork (scores of 92 and 88) but gave a weaker answer on technical knowledge (score of 70)."

Advantages of Vector-Based Grading

  • Consistency and Objectivity: Every candidate is measured against the same ideal answers and criteria. Personal biases or inconsistencies that human interviewers might introduce are minimized.
  • Depth of Understanding: Because Stella evaluates the meaning of responses, it appreciates good answers even if phrased uniquely. The AI checks whether the candidate actually addressed the question and hit the key points — rather than scanning for specific keywords.
  • Alignment with Human Judgment: Embedding-based scores correlate closely with how experts would grade an answer's correctness and completeness. This gives confidence that the rating mechanism is a reasonable proxy for human reasoning — only faster and more scalable.
  • Transparency in Feedback: Since each answer is scored against ideal content, reports can highlight what was missing or where the answer deviated. This insight is valuable for both recruiters (to probe further) and candidates (to learn and improve).

Conclusion

Stella grades answers by comparing the candidate's response to an ideal answer using vector-based semantic similarity. The steps involve transforming answers into vector form, quantifying their similarity to the expected answer, and mapping that similarity to a clear rating. Each answer's rating contributes to a cohesive overall score for the interview.

This approach — without divulging proprietary specifics — ensures that scoring is standardized, data-driven, and aligned with the content of answers rather than superficial keywords. By grounding evaluations in semantic similarity, Stella provides reliable, explainable ratings that HR teams can trust to identify top talent efficiently.

See Stella score in real time.

Book a demo and watch how AI Interview evaluates candidates on your rubric.