- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper generates tennis commentary with a multimodal LLM (InternVL3.5-8B). The model, TennisRAM, adds four parts:
- Auxiliary heads that predict ball position, player pose and strokes. They are trained on labels from off-the-shelf detectors.
- A token-compression step guided by the predicted ball and pose heatmaps.
- A GRU summarizes the visual features every second in order to predict which player holds the advantage and tag the supporting factors.
- Commentary generation, conditioned on the visual tokens and these predictions.
The authors also annotate 1,200 TennisVL rallies with advantage and point labels.
Critical strengths
- Tracking who holds the advantage during a rally is a sensible, well-motivated second-order (tactical) step toward commentary that explains how a point develops, not just what happened.
- Using perception to guide token compression is an interesting idea. In the appendix it beats feeding the heatmaps as extra tokens (Table 4). Table 2 also shows large gains from guidance, but the baseline it compares against is weak (see W2).
Critical weaknesses
- Some main claims do not appear to be validated:
- The paper says the auxiliary heads learn “fine-grained representations”. But the InternVL encoder is frozen in all stages (Stage III only adds LoRA to the language decoder), so the heads cannot change its features. They work as small detectors, which are never evaluated.
- The paper says the compression improves efficiency, but no token counts, FLOPs or latency are reported.
- High-frame-rate sampling is presented as a contribution, but the frame rate is never given or ablated.
- Section 6.4 says advantage conditioning helps commentary, but Table 3 (which is not mentioned in the main text) shows it alone lowers BLEU and CIDEr. (Not relevant: I assume in this section when you mention Score you refer to Point(-R) ).
- The gains in recall quoted in Section 6.3 (pose over the ball) don’t match Table 2.
- The experiments and method description are not solid enough to trust or reproduce.
- The baseline without auxiliary heads in Table 2 looks too weak. Its advantage accuracy is near the majority-class rate, and its CIDEr (5.49) is far below the no-condition model in Table 3 (37.91). A fair compression baseline without perception guidance is needed.
- The general MLLMs are zero-shot, and some score below random on a 3-class task, which suggests parsing problems. The TennisExpert commentary scores are copied from its preprint and probably come from a different test set.
- The test set is small (219 rallies, 2 matches), with a single run and no seeds.
- Multi-label metrics are undefined, and there’s no annotator agreement for a subjective label.
- It is unclear whether the reference commentary is TennisVL’s original text or LLM-generated text. A generated sample in Fig. 1 mentions a soccer “penalty kick”.
- The method lacks tensor shapes and head definitions, and the Q-Former is not cited.
- Design choices (layers 12/6/1, 4×4 grid, 10% bound) are never justified.
Questions for authors
Q1. Is the commentary target TennisVL’s original commentary? If so, what are the GLM-5.2 descriptions in Fig. 1 used for? If they are used as targets or inputs, conditioning on the labels they were generated from would be circular. Q2. The TennisExpert BLEU/ROUGE/CIDEr scores are copied from its paper. Were they computed on the same test rallies and reference texts as TennisRAM? Q3. What is the inter-annotator agreement for the advantage labels? Q4. Adv.-R and Point-R are multi-label tasks, and Table 1 reports “accuracy” for them. Is this exact match (all labels correct), per-label accuracy, or something else? Can you also report macro-F1 and a majority-class baseline? Q5. The paper says high-frame-rate sampling and token compression help, but gives neither value. How many frames per second do you sample, and how many visual tokens reach the LLM per second (the number of Q-Former queries)? How do FLOPs or latency compare with the default InternVL3.5 input? Q6. Can you report results over several seeds? Q7. At test time, the commentary is conditioned on predicted advantage and point labels. During Stage III training, does it see ground-truth labels or predicted ones? If ground truth, how much does commentary quality drop when predicted labels replace them at test time? Q8. In the first row of Table 2, with no auxiliary branches, what does PGSP compute? Is it plain region averaging?
Suggestions:
- Fine-tune plain InternVL3.5-8B on the same data, to show the gains come from the new modules and not just from fine-tuning.
- Feed the ball & pose outputs from the third party detectors directly. This will show whether the heads add anything.