The thesis goals are moderate-to-hard, since it deals with recent advances in speech-LLMs. The problem of 'spoken question answering' however is not new. 

The primary work is reasearch in nature and the experiments and findings from the thesis reflect this clearly.

- The student plans to submit the work to an upcoming conference or workshop. 

- Consultations were regular through out the work.

Evaluation level: zadání splněno

The assignment was fullfiled completely and without exception. The student also presents additional results and findings related to a crucial issue of shortcut learning for audio-LLMs, where the input audio documents are not attended to when generating the answers.

I found that the theoretical introduction (chapter 2) of the thesis a little hard to digest, as while the student provides a comprehensive overview of spoken question-answering (QA) and retrieval-augmented generation (RAG) as a whole, some of the presented concepts were not exactly directly related to the experimental work itself. On the other hand, audio-LLMs receive only limited attention though they are a core topic of the thesis. Some of the section names would also benefit from more specifc wordings for better reader orientation. Lastly, sometimes, certain pieces of information are unnecessarily repeated between neighbouring sections. Despite these shortcomings, the overall structure of the thesis makes sense for the reader and the experimental chapters of the thesis are of significantly higher quality in terms of presentation and flow, and I find that they offset the shortcomings of the theoretical parts.

The student cites relevant prior works in reasonable breadth with respect to the theis topic. When covering certain LALM concepts, survey articles are cited where it would be perhaps more appropriate to cite the original works in addition. Also, some of the bibliography entries are arxiv preprints, where the said article was published at a conference/in a journal and should instead be cited as such.

Overall, the thesis presents valuable results and analysis in a currently highly-relevant domain of spoken QA. I find that the methodology is well controlled, where every part of both the baselines and the final system is well-ablated. The experimental results on adapting the Qwen audio-LLM for spoken QA are valuable for the wider spoken QA community, providing insights on the current shortcomings of such modles and how they can be addressed. I commend the presentation of evidence of shortcut learning, where the model does not attend to the spoken inputs, as such potentially hidden issues of audio-LLMs should be more widely understood in the audio/speech-LLM domain than they currently are. In my overall opinion, the work presented in the thesis warrants a publication.