Explicit rules help an AI interviewer follow a plan, but controlling tone remains difficult

An AI interviewer needs more than plausible questions. It may also need to follow a sequence, adjust its language and respond with a particular emotional tone. Barbara Gendron, Gaël Guibon and Mathieu d’Aquin investigate whether these requirements can be expressed in an ontology: a formal description of concepts, relationships and rules that software can reason over.

Their study, submitted to arXiv on 6 November 2026, is a preprint; the supplied source does not establish peer-reviewed publication. Its clearest finding is a distinction: the agent followed the broad interview structure, but precise control over its wording and emotional characteristics remained limited.

A conversation plan outside the model

The researchers built an ontology describing conversations, user and agent utterances, conversation phases and utterance characteristics. Unlike instructions embedded only in a prompt, this representation makes the strategy explicit and separate from the language model.

Their demonstration used five interview phases: small talk, the candidate’s background, a technical discussion, debate about a controversial topic, and a concluding summary. Phase boundaries depended on the number of user utterances. Each phase supplied its own system prompt.

The system also tracked three characteristics: language proficiency level on the six-level Common European Framework of Reference scale, whether an utterance was emotionally charged, and whether its emotional polarity was negative, neutral or positive. These last two describe properties of the text, not a direct measurement of the speaker’s internal feelings.

Two previously trained BERT-based classifiers estimated emotional load and polarity. For language level, the team extracted 44 linguistic features and trained a decision tree. They translated its classification rules into the ontology, allowing logical reasoning to assign a level.

The ontology then derived instructions for the next response using the current phase and characteristics of user utterances. One example rule requests an A2-level response during the background phase when recent user language meets a specified B-level condition.

Teaching the model to follow the signals

Writing a rule does not guarantee that a language model will obey it. The researchers therefore fine-tuned Llama3.1-8B-Instruct and Qwen2.5-7B-Instruct to respond to prefixes specifying requested characteristics.

Training paired randomly selected open-ended questions with randomly sampled descriptor instructions. A reinforcement-learning method called Group Relative Policy Optimization rewarded generated responses for matching those instructions. LoRA adapters allowed training to update a small fraction of the model’s parameters.

The authors describe this as a data-free approach because it avoids assembling a dedicated, manually annotated dataset for the generation task. It does not mean the whole system uses no training data: the descriptor predictors were trained on existing corpora, and the language models already had extensive prior training.

In standalone generation tests, Llama’s exact language-level accuracy rose from 18% to 29%, while polarity accuracy increased from 42% to 62%. Qwen’s corresponding results rose from 18% to 21% and from 46% to 57%. These gains show some responsiveness to the control signals, but exact language-level matching remained low.

Emotional load was harder to control. Llama’s accuracy barely changed, from 50% to 51%, while Qwen’s fell from 49% to 46%. Fine-tuning therefore did not improve every requested characteristic.

What happened in full interviews

For the conversation evaluation, three people with academic computer-science backgrounds role-played interview candidates. The paper reports 30 conversations, each containing 12 user–agent exchanges. The comparison used the original and fine-tuned Llama models.

The authors report coherent transitions through the five interview phases. However, automatic agreement with requested emotional load fell from 24% to 18% after fine-tuning, while polarity agreement rose from 34% to 40%.

Human judgments differed from the automatic measurements. Evaluators generally perceived better emotional-load compliance than the classifiers detected. Two of three annotators recorded improved polarity compliance after fine-tuning, but two recorded worse emotional-load compliance. The authors suggest that frequent polite or encouraging opening phrases can push automatic classifications toward emotionally charged, positive labels.

What the results do not establish

The evaluation is small and does not demonstrate that using the agent improves interview skills or employment outcomes. A broader external human evaluation remained future work. Applying the framework elsewhere also requires designing an appropriate ontology, which can involve substantial knowledge-modeling effort.

The contribution is an explicit way to specify and inspect conversational strategy—not a guarantee of compliant output. In these experiments, organizing an interview was more reliable than making every response match a prescribed linguistic or emotional profile.