Taming LLMs with Negative Samples: A Reference-Free Framework to Evaluate Presentation Content with Actionable Feedback

Ananth Muppidi; Dharun D A; Sambaran Bandyopadhyay; Tarak Das; Tripti Shukla

REVIEW 4 major objections 5 minor 12 references

Taming LLMs with Negative Samples: A Reference-Free Framework to Evaluate Presentation Content with Actionable Feedback

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read REFLEX trains an LLM on deliberately broken slides and beats GPT-4o-based scoring for reference-free presentation evaluation.

desk verdict A useful new dataset and a plausible training recipe, but the headline superiority claim rests on a circular automatic evaluation and a small human study that partly rescues it. read the letter →

arxiv 2505.18240 v1 pith:DVOV5PJB submitted 2025-05-23 cs.CL cs.AI

Ananth Muppidi , Tarak Das , Sambaran Bandyopadhyay , Tripti Shukla , Dharun D A This is my paper

classification cs.CLcs.AI

keywords reference-freeevaluationpresentationqualitymetricsnegativesamplegenerationmultimodalLLMfine-tuningactionablefeedbackRefSlidesbenchmarkcontrastivelearning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that an LLM can judge presentation quality—coverage, redundancy, text-image alignment, and flow—without needing a reference presentation at inference time, and can explain its scores in actionable terms. The trick is to manufacture the training signal: take good human-made decks, corrupt them in controlled metric-specific ways at four intensity levels, assign pseudo scores of 5 minus the degree, and fine-tune a small LLM to produce explanations and scores. The authors build a new benchmark, RefSlides, with 8,111 filtered SlideShare decks, and report that REFLEX beats heuristic embeddings, a same-size prompted LLM, and the GPT-4o-based G-Eval on correlation with pseudo ground truth, with human ratings, and on ROUGE for explanations. If this holds, automatic slide generators can be checked and improved without human references or expensive annotations.

What carries the argument

The load-bearing object is the perturbation function h^m_d for each metric, which turns positive presentations into negative samples with a pseudo ground-truth score of 5-d and a templated explanation (e.g., removing d*20% of topic-related slides for coverage, duplicating d*20% of slides for redundancy, permuting images across d*20% of slides for alignment, and permuting d*20% of slide positions for flow). These five-way contrastive sets train two LoRA adapters on Phi3-Mini: one generation head for actionable explanations and one single-neuron regression head for scores. The mechanism is that the degree-graded corruptions teach the model to separate fine quality differences and to verbalize what went wrong.

What would settle it

Collect a set of real presentations with naturally occurring quality variation (e.g., decks from different authors without synthetic corruption), obtain human ratings on the four metrics, and compute the Spearman correlation of REFLEX scores with those ratings; if REFLEX no longer beats the G-Eval baseline on this natural set, the synthetic-perturbation assumption fails.

Watch

Extended reading notes

Core claim

REFLEX establishes that synthetically generated negative samples—slides with deliberately removed topics, duplicated slides, swapped images, or shuffled orders—carry enough signal to train a reference-free evaluator that outperforms prompt-based LLM evaluators. For each metric, a perturbation function h^m_d maps a good presentation to a negative one of degree d, with pseudo ground-truth score 5-d and a templated explanation. Fine-tuning Phi3-Mini with LoRA on these five-way training sets (one positive, four negative degrees) yields a scoring model and an explanation model that, at inference, take only the presentation (plus the source document for coverage) and return a score in [0,1] and concrete feedback. The central claim is that this contrastive training transfers to real human judgments better than heuristic cosine-similarity metrics or in-context LLM evaluation.

Load-bearing premise

The assumption that each hand-designed perturbation degrades exactly the intended metric, and that the perturbation degree d corresponds linearly to true quality via score = 5 - d, is the load-bearing premise; if real presentation flaws differ from these synthetic corruptions, the trained scores and feedback may not reflect actual quality.

Editorial extensions

If this is right

Presentation generators can be evaluated on the fly, without reference decks, enabling closed-loop refinement of auto-generated slides.
The negative-sample recipe extends to new quality metrics whenever a plausible corruption operator can be defined, making the framework modular.
A small fine-tuned model can beat a much larger prompted LLM on this structured evaluation task, lowering the compute barrier for deployment.
Feedback is generated alongside scores, so users see not just a number but which slides to fix and why.
Human annotation cost for evaluator training is largely replaced by synthetic labeling, though human validation remains necessary.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond presentations, the same perturbation-and-score recipe could be applied to other structured multimodal outputs—posters, infographics, report layouts—where quality flaws can be synthesized.
The linear score mapping (5-d) is a design choice; a learned mapping from perturbation degree to quality, or probabilistic labels, could improve calibration and is directly testable.
The real-world validity of REFLEX hinges on whether natural presentation flaws resemble the synthetic corruptions; a natural-corruption benchmark with human ratings would settle that.
The method's sensitivity to small perturbation degrees suggests it could double as a diagnostic tool, pinpointing the exact slides responsible for a low score.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The core claim is only half-supported as presented. The automatic correlations in Table 1 are circular, but the human study gives real, if modest, independent signal that REFLEX beats G-Eval on three of four metrics. That is enough to take the work seriously, but not enough to accept the central claim without changes.

What is actually new: RefSlides, an 8,111-presentation cross-domain dataset curated from SlideShare, and REFLEX, a reference-free evaluator that fine-tunes Phi3-Mini with LoRA on synthetic negative samples to output per-metric scores and template-based feedback. The perturbation scheme for coverage, redundancy, text-image alignment, and flow is plausible and well described. The human evaluation covers 50 SciDuet decks with three annotators each and reports inter-annotator agreement, which is good practice. REFLEX beats G-Eval on coverage, redundancy, and text-image alignment, though not on flow.

The soft spots are serious but not fatal. Section 5.2 defines the pseudo-ground-truth score as 5 minus perturbation degree, for the same perturbations used in training, and Table 1 measures correlation against those labels. That is fitting the training target, not independent validation; any baseline that detects the perturbation type would score well. The human study is the only independent check, and it has two weaknesses: only 50 samples, and per Appendix D annotators saw both original and perturbed slides, so they may be rating perturbation degree rather than intrinsic presentation quality. The explanation comparison via ROUGE is weak, since the fine-tuned model is trained to reproduce template strings, and the few-shot baselines are at a structural disadvantage. Flow underperforming G-Eval is buried in Table 3. On the citation pattern: the related work is reasonable for the subfield, and the self-citations to prior Adobe work are contextually appropriate. The paper lists limitations, but not the circularity of the synthetic-label evaluation, which is the load-bearing issue.

Who this is for: people building document-to-slide systems or reference-free multimodal evaluators. RefSlides, if released, would be a solid benchmark. The method is a reasonable baseline, but the evaluation needs rework.

Recommendation: send to peer review, but the authors should be pushed to release the dataset and code, add a transfer test on naturally occurring or differently-perturbed errors, and either expand the human study or reframe the automatic metrics as calibration checks rather than proof of superiority.

Referee Report

4 major / 5 minor

Summary. The paper introduces REFLEX, a reference-free framework for scoring and explaining presentation quality along four metrics: coverage, redundancy, text-image alignment, and flow. The authors curate RefSlides, a dataset of 8,111 SlideShare presentations, generate negative training samples by applying metric-specific perturbations at four degrees with pseudo-ground-truth scores defined as s = 5 - d, and fine-tune Phi-3-Mini with LoRA to produce both scores and actionable explanations. They compare REFLEX with heuristic baselines, G-Eval, and Phi3-Eval, reporting higher correlations with pseudo ground truth (Table 1), higher ROUGE F1 for explanations (Table 2), and generally higher correlations with human ratings on 50 SciDuet presentations (Table 3).

Significance. If the evaluation evidence is validated, REFLEX would be a useful contribution to automatic multimodal presentation evaluation, particularly because it combines score prediction with actionable feedback and does not require a reference presentation at inference time. The paper is also commendable for formalizing four presentation-quality metrics, for providing a modular LoRA-based architecture that can be extended to new metrics, and for including a human evaluation component. However, the central evidence is largely generated by the same perturbation schedule used to construct the training labels, so the real-world value of the method currently rests on a small and partly synthetic human study.

major comments (4)

[Section 5.2 and Table 1] The automatic evaluation in Table 1 is circular as a validation of quality estimation. Section 5.2 defines the training target as sid = 5 - d and generates negative samples from degree-d perturbations, and Table 1 then reports correlations with exactly these pseudo ground-truth labels. Because REFLEX is trained to minimize MSE against these labels, high Spearman and Kendall-Tau correlations on a held-out split mainly demonstrate that the model reproduces the label-generating function. This does not provide independent evidence that REFLEX is better than the baselines at estimating true presentation quality. The claim in Section 6.3.1 that the model 'correlates the highest with the Pseudo Ground Truth' should be reframed as an internal consistency check, and the headline comparison should rest on external validation.
[Section 6.4 and Appendix D, Table 3] The human evaluation, which is the main independent evidence, is limited in two ways. First, it uses only 50 SciDuet presentations with three annotators per sample, and Appendix D instructs annotators that 'Both original and perturbed slides are provided for comparison,' where the perturbed slides are generated by the same metric-specific perturbations used for training. The study therefore partly measures sensitivity to the synthetic corruptions rather than alignment with real-world quality judgments. Second, the paper's abstract claims that REFLEX outperforms state-of-the-art baselines, but Table 3 shows that for Flow, G-Eval is better (rho = 0.36 vs. 0.32; tau = 0.33 vs. 0.29). This discrepancy must be acknowledged, and a realistic transfer test on naturally occurring, non-perturbed presentations is needed to support the central claim.
[Section 5.2.1] The perturbation-to-quality mapping is assumed rather than validated. The paper sets x = d * 20 and sid = 5 - d for all metrics, but the perturbation functions are not metric-isolating: removing topic slides can also alter flow and redundancy, duplicating slides can alter flow, and permuting images can affect the visual narrative. The linear relationship between perturbation degree and quality is never checked against human judgments. The Limitations section acknowledges model size, non-exhaustive metrics, and lack of design evaluation, but it does not address this validity threat. A calibration study in which human raters score presentations at each perturbation degree, or a justified argument for metric isolation, is necessary before the pseudo-labels can anchor the method's claims.
[Section 6.3.3 and Section 6.4.3] The explanation evaluation is also tied to the synthetic training signal. Table 2 computes ROUGE F1 against pseudo-ground-truth explanations produced by the same template-based perturbation functions used to create REFLEX's training targets, so high ROUGE is expected to at least some degree. The human evaluation of explanations in Section 6.4.3 reports only a score distribution and does not provide per-method means, standard deviations, or statistical tests. Please report per-method summary statistics and a significance test to support the claimed superiority in explanation quality.

minor comments (5)

[Throughout] There are several typos and inconsistent notations: 'Actionalble' in Section 1, 'Kendal-Tau' in Section 6.3.1, 'constrative' in the Limitations section, and inconsistent use of 'Reflex' vs. 'REFLEX'. These should be corrected.
[Table 2] The table formatting for 'ROUGE Metric' is misaligned, and the table would benefit from reporting sample sizes and confidence intervals for the ROUGE scores.
[Appendix F] The hyperparameter search space is reported, but the selected final hyperparameter values are not. Listing the chosen configuration would improve reproducibility.
[Section 3] RefSlides is described as a new benchmark dataset, but no availability statement, data release URL, or licensing information is provided. The paper should state how the dataset will be made accessible.
[Section 6.4.3] The explanation human evaluation should include a table with per-method mean scores, standard deviations, and inter-annotator agreement by metric, rather than a single aggregate distribution.

Circularity Check

2 steps flagged · score 6.0 of 10

Automated evaluation is circular: REFLEX is trained and tested on pseudo-scores and pseudo-explanations generated by its own perturbation functions, so the main automated claims reduce to a fit.

fitted input called prediction [Section 5.2, Section 5.2.3, Table 1]
"We also set the pseudo ground truth score sid = 5−d for degree perturbation d. ... The target output are the corresponding scores from ˆTm. ... Comparison of correlation coefficients between baselines and pseudo ground truths."

The numeric labels used as the evaluation gold standard in Table 1 are exactly the labels on which the REFLEX scoring model was trained. Section 5.2 defines the pseudo-ground-truth score as sid = 5 − d for perturbation degree d, and Section 5.2.3 sets the training target to 'the corresponding scores from ˆTm'. Measuring Spearman and Kendall-Tau correlation against these same constructed labels therefore shows only that the model learned the paper's own score-by-perturbation-degree mapping; it does not show independent prediction of presentation quality. The baselines were never given these labels, so the comparison is asymmetrical and the claimed advantage is partly forced by training on the evaluation target.
fitted input called prediction [Section 5.2.2, Section 6.3.3, Table 2]
"We use standard CrossEntropy loss over the tokens generated by the model and the tokens in the pseudo ground truth explanations. ... We compute the ROUGE-F1 scores between the ground truth explanations and the model-generated explanations to evaluate their quality."

The 'ground truth explanations' used in the ROUGE evaluation are the pseudo explanations produced by the perturbation functions, i.e., template strings such as 'The following topics from the source document should be added: ...'. The explanation module is fine-tuned with cross-entropy to reproduce exactly these token sequences. Computing ROUGE-F1 against the same training-target templates therefore measures how well the model fits the template distribution, not whether the generated feedback is independently actionable or correct. The result gives no external evidence about explanation quality for naturally occurring presentation flaws.

full rationale

REFLEX's central claim of outperforming heuristic and LLM-based evaluators rests mainly on Table 1 and Table 2, and both are circular in the same way: the gold-standard labels are manufactured by the paper's own perturbation pipeline. Scores are defined as sid = 5 − d, and the scoring head is trained with MSE on exactly those labels; explanations are template-generated by h_m_d, and the explanation module is trained with cross-entropy on exactly those templates. High correlation with pseudo-scores in Table 1 and high ROUGE against pseudo-explanations in Table 2 therefore reduce to fit-to-training-labels rather than independent evaluation quality. The human study in Table 3 and the human explanation ratings provide some external grounding, and they are genuinely informative, but they are small (50 samples, 3 annotators per sample), and Appendix D tells annotators that 'Both original and perturbed slides are provided for comparison,' so the human ratings may also track the same synthetic perturbation degree rather than naturally occurring presentation defects. No load-bearing self-citation or imported uniqueness theorem appears in the paper; the circularity is of the fitted-input-called-prediction kind, concentrated in the automated evaluation sections. Overall score 6: the central automated benchmark is partially circular, while the human study keeps the result from being fully equivalent to its inputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The framework depends on hand-set perturbation schedules and score mappings that generate the pseudo ground truth, plus domain assumptions about how well extracted summaries and synthetic corruptions represent real presentation quality. No new physical or conceptual entities are introduced.

free parameters (3)

Perturbation degree-to-score mapping = score = 5 - d for d in {1,2,3,4}
Hand-set linear mapping defines pseudo ground truth labels for training and evaluation; Section 5.2.
Perturbation fraction = x = 20*d percent of slides or topics
Hand-set schedule determining how much content is removed, duplicated, swapped, or shuffled per degree; Section 5.2.1.
Dataset curation thresholds = 5-35 slides, >80% consecutive overlap rejected, 16:9 aspect ratio, English
Hand-set filters that define which SlideShare decks count as high-quality positives; Section 3 and Appendix A.

assumptions (4)

domain assumption Filtered human-made presentations are high-quality positive samples
Training treats all RefSlides and SciDuet decks as score 5; Section 3 and Section 5.2.
domain assumption Each perturbation affects only its target metric and degrades quality monotonically with degree
The pseudo-labels assume removing topics, duplicating slides, swapping images, and shuffling slides linearly lower the intended metric; Section 5.2.
domain assumption Slide text and image summaries from Phi-3-vision preserve enough information for evaluation
All downstream scoring uses extracted JSON summaries, not raw slide images; Section 5.1.1.
domain assumption ROUGE-F1 is a valid automatic measure of explanation quality
Used to compare generated explanations with template pseudo-explanations; Section 6.3.3.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Taming LLMs with Negative Samples: A Reference-Free Framework to Evaluate Presentation Content with Actionable Feedback." pith.science (2026). https://pith.science/paper/DVOV5PJB

@misc{pith2026250518240,
  author       = {Pith},
  title        = {Pith review of: Taming LLMs with Negative Samples: A Reference-Free Framework to Evaluate Presentation Content with Actionable Feedback},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DVOV5PJB}},
  note         = {Machine review of arXiv:2505.18240}
}

read the original abstract

The generation of presentation slides automatically is an important problem in the era of generative AI. This paper focuses on evaluating multimodal content in presentation slides that can effectively summarize a document and convey concepts to a broad audience. We introduce a benchmark dataset, RefSlides, consisting of human-made high-quality presentations that span various topics. Next, we propose a set of metrics to characterize different intrinsic properties of the content of a presentation and present REFLEX, an evaluation approach that generates scores and actionable feedback for these metrics. We achieve this by generating negative presentation samples with different degrees of metric-specific perturbations and use them to fine-tune LLMs. This reference-free evaluation technique does not require ground truth presentations during inference. Our extensive automated and human experiments demonstrate that our evaluation approach outperforms classical heuristic-based and state-of-the-art large language model-based evaluations in generating scores and explanations.

Figures

Figures reproduced from arXiv: 2505.18240 by the authors.

**Figure 1.** Overview of the architecture of document pre [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 3.** Correlations between the pseudo ground truth and metric for each degree. Reflex shows a significantly [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Distribution of scores for explanations gener [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: explains the inference pipeline of REFLEX [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

**Figure 6.** Figure 6: Sample output of actionable feedback from REFLEX [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 10 canonical work pages

[1]

Length: We worked with presentations of length 5 to 35 slides only. This is because pre- sentations less than 5 slides long were found to be lacking any considerable information and those longer than 35 slides were rejected due to context-length limitations of LMMs

work page
[2]

Heuristics: Based on examination of ran- domly sampled downloaded presentations, we came up with a set of qualitative criteria to reject low-quality samples: • Introductory and Concluding Slides: We assumed a presentation can be called ‘complete’ only if it has an introductory and a concluding slide. • Duplicate Slides: We removed all the presentations fr...

work page
[3]

No Text". Make it concise

Aspect Ratio: We strictly limit all the presen- tations to be of16 : 9aspect ratio. For step-2, we convert each presenta- tion Pi into a set of images of slides Pi = [Si,1, Si,2,· · ·, Si,mi]∀i∈[N], P i ∈ P. This array is then parsed into batched of 5 slide-images concatenated vertically and fed into GPT-4o using a suitable prompt to evaluate for each of ...

work page
[4]

Flow Flow assesses the logical progression of the slides. Consider: • Does the presentation have a smooth and logical flow? • Are transitions between slides clear and cohesive? Score: High if the flow is logical; low if the content feels disjointed. 14 And the following instructions were given to score the explanations generated by the model and the basel...

work page 2019
[8]

If you feel that there is no redundancy in the slide and it is concise, respond with the following sentence:

T opic 2" If you feel that there is no redundancy in the slide and it is concise, respond with the following sentence: "Slides are concise and there is little to no redundant information”." Similarly, as explained in section 6 5, models for other metrics are prompted appropriately. Notice that the prompting follows the exact same pattern as the pseudo exp...

work page
[9]

Coverage Coverage assesses how well the slides cap- ture key information from the source doc- ument. Consider: • Are key topics from the document included in the slides? •Are important topics missing? Score: High if most key topics are cov- ered; low if important content is missing

work page
[10]

Redundancy Redundancy evaluates the presence of repetitive information. Consider: • Are slides free of unnecessary repe- tition? • Could any slides be combined or re- moved? Score: High if the presentation is concise; low if content is repeated unnecessarily

work page
[11]

Text-Image Alignment This metric measures the relevance of im- ages to the text. Consider: • Do images complement and clarify the text? • Are any images irrelevant to the slide’s content? Score: High if images and text are well- aligned; low if images are irrelevant

work page

Show all 12 references

[2019]

InProceedings of the 57th Annual Meeting of the Association for Computa- tional Linguistics, pages 2748–2760, Florence, Italy

Sentence mover’s similarity: Automatic evalu- ation for multi-sentence texts. InProceedings of the 57th Annual Meeting of the Association for Computa- tional Linguistics, pages 2748–2760, Florence, Italy. Association for Computational Linguistics. Daniel Deutsch, Tania Bedrax-...
[2021]

Daniel Deutsch, Rotem Dror, and Dan Roth

Towards question-answering as an automatic metric for evaluating the content quality of a sum- mary.Transactions of the Association for Computa- tional Linguistics, 9:774–789. Daniel Deutsch, Rotem Dror, and Dan Roth. 2022. On the limitations of reference-free evaluations of g...

2022 arXiv
[2023]

Himanshu Maheshwari, Sambaran Bandyopadhyay, Aparna Garimella, and Anandhavelu Natarajan

Chatgpt as a factual inconsistency evaluator for text summarization.Preprint, arXiv:2303.15621. Himanshu Maheshwari, Sambaran Bandyopadhyay, Aparna Garimella, and Anandhavelu Natarajan. 2024. Presentations are not always linear! gnn meets llm for text document-to-presentation ...

2024 arXiv
[2024]

Hung-Ting Su, Ya-Ching Hsu, Xudong Lin, Xiang-Qian Shi, Yulei Niu, Han-Yuan Hsu, Hung-yi Lee, and Winston H

Fenice: Factuality evaluation of summariza- tion based on natural language inference and claim extraction.Preprint, arXiv:2403.02270. Hung-Ting Su, Ya-Ching Hsu, Xudong Lin, Xiang-Qian Shi, Yulei Niu, Han-Yuan Hsu, Hung-yi Lee, and Winston H. Hsu. 2024. Unveiling narrative rea...

2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Length: We worked with presentations of length 5 to 35 slides only. This is because pre- sentations less than 5 slides long were found to be lacking any considerable information and those longer than 35 slides were rejected due to context-length limitations of LMMs

work page

[2] [2]

Heuristics: Based on examination of ran- domly sampled downloaded presentations, we came up with a set of qualitative criteria to reject low-quality samples: • Introductory and Concluding Slides: We assumed a presentation can be called ‘complete’ only if it has an introductory and a concluding slide. • Duplicate Slides: We removed all the presentations fr...

work page

[3] [3]

No Text". Make it concise

Aspect Ratio: We strictly limit all the presen- tations to be of16 : 9aspect ratio. For step-2, we convert each presenta- tion Pi into a set of images of slides Pi = [Si,1, Si,2,· · ·, Si,mi]∀i∈[N], P i ∈ P. This array is then parsed into batched of 5 slide-images concatenated vertically and fed into GPT-4o using a suitable prompt to evaluate for each of ...

work page

[4] [4]

Flow Flow assesses the logical progression of the slides. Consider: • Does the presentation have a smooth and logical flow? • Are transitions between slides clear and cohesive? Score: High if the flow is logical; low if the content feels disjointed. 14 And the following instructions were given to score the explanations generated by the model and the basel...

work page 2019

[5] [8]

If you feel that there is no redundancy in the slide and it is concise, respond with the following sentence:

T opic 2" If you feel that there is no redundancy in the slide and it is concise, respond with the following sentence: "Slides are concise and there is little to no redundant information”." Similarly, as explained in section 6 5, models for other metrics are prompted appropriately. Notice that the prompting follows the exact same pattern as the pseudo exp...

work page

[6] [9]

Coverage Coverage assesses how well the slides cap- ture key information from the source doc- ument. Consider: • Are key topics from the document included in the slides? •Are important topics missing? Score: High if most key topics are cov- ered; low if important content is missing

work page

[7] [10]

Redundancy Redundancy evaluates the presence of repetitive information. Consider: • Are slides free of unnecessary repe- tition? • Could any slides be combined or re- moved? Score: High if the presentation is concise; low if content is repeated unnecessarily

work page

[8] [11]

Text-Image Alignment This metric measures the relevance of im- ages to the text. Consider: • Do images complement and clarify the text? • Are any images irrelevant to the slide’s content? Score: High if images and text are well- aligned; low if images are irrelevant

work page