Question about paper on Qwen3-VL

[zrli03]

Woderful work! Thank you for your contributions, though I have some questions to be solved.
I noticed that you mentioned :

We observed that directly applying RL training on Qwen3-VL fails to yield improvements, likely because Qwen3-VL has undergone large-scale multi-task RL training that includes VTG data, preventing the model from generating rollouts with sufficient diversity on VTG task during our continual RL.Therefore, we first perform a small SFT stage to, in a sense, revert the model back to the “base model” state before RL. This is merely a workaround specific to Qwen3-VL, a model that has already acquired strong VTG capabilities through an RL stage similar to that proposed in this paper. In the common scenario, our recipes are designed to enhance the VTG capabilities of a “base MLLM”, where this trick is not required.

in page 16 of your paper.
My question:

1. Is this strategy necessary when evaluating Qwen3-VL?
2. Could you briefly introduce the SFT configuration/settings for Qwen3-VL?
3. What do you mean by “common scenario”? Can I understand it as: this step may not be required for other tasks (e.g., VQA), but it is necessary when fine-tuning Qwen3-VL on our own VTG dataset?

These questions are intended purely for discussion. I will also run my own experiments to validate the points above. I hope we can have a pleasant, in-depth discussion, and thank you again for your contribution.
Wishing you all the best.

[jonasjunzhang]

Hello! These are excellent questions.

First, by "common scenario," we refer to the setting where we start with a base MLLM (pretrained and perhaps post-trained) that has general capabilities but yields suboptimal performance on Video Temporal Grounding (VTG), such as Qwen2.5-VL. Our goal is to design data and algorithms to enhance the VTG capabilities of such models, which is the motivation of TimeLens.

Qwen3-VL is not "common scenario" because, based on its strong performance on TimeLens-Bench and its technical report, it appears to have already undergone extensive optimization (during SFT and RL) specifically for VTG. It‘s like the Qwen team have developed a "closed-source competitor" to our data and method. Directly applying our data and method on top of a model trained with a "closed-source competitor" to ours can lead to unpredictable results. This is why we conducted our ablation studies on Qwen2.5-VL to ensure reliability.

Second, the role of the "small SFT stage":

• As mentioned in the paper, we observed that Qwen3-VL could not generate rollouts with sufficient diversity on the VTG task (likely due to entropy collapse from its RL stage). The SFT stage serves to revert the model closer to a "base model" state and restore entropy.
• Post-SFT performance improved slightly over the original Qwen3-VL but was negligible compared to the gains from the subsequent RL stage. Thus, we reported the original Qwen3-VL performance in the paper. However, I think you are right—reporting the post-SFT performance might be fairer. Thanks for pointing that out!

Third, advice for your experiments:

• For other tasks (e.g., VQA): I haven't experiment on this, so I cannot give a definitive conclusion on whether SFT is needed.
• For fine-tuning on your own VTG dataset: It is highly likely that a preceding SFT stage is beneficial. However, it is also possible that our specific data and RL recipe conflicts with their proprietary ones, necessitating this "SFT reset". I still recommend you to comparing performance with and without SFT to verify this. Sorry I cannot provide more useful suggestions. I would be very interested to hear your results if you are willing to share them!

I think if a "base" version of Qwen3-VL is available, it would benefit the research community and those who wants to fine-tune it. Unfortunately, only the final RL-tuned model is available.

Last, for our SFT Configuration on Qwen3-VL, we train for one epoch, with a batch size of 128 and a learning rate of 1e−5. We sampled approximately 30K data for this stage. Note that this data amount was not ablated; using more or less data might yield better results.

I hope this addresses your concerns and provides insights for your future experiments. Please feel free to ask follow-up questions or bring some further discussions!

[jonasjunzhang]

By the way, I think we don't need to close this issue. It might be helpful to others.

[zrli03]

Wow, this is a very comprehensive and detailed reply! I agree with your point, and I think this issue can be kept as a discussion thread on RL/SFT + Qwen3-VL.

Recently, I have been trying to sample hard cases that Qwen3-VL struggles with from multiple TVG datasets (e.g., TimeRFT, DiDeMo, and TimeLens-100K proposed by your team) for post-training. As you mentioned, Qwen3-VL already demonstrates solid TVG capability, but in specific real-world scenarios, directly applying a smaller-sized Qwen3-VL (e.g., 3B) may not perform well. I hope to explore some effective few-shot training strategies to help promote the practical adoption of MLLMs in real-world applications.

Wishing you all the best in your future research. Thank you again for your serious attitude and meticulous work.