# Qwen2.5-VL

Vision-language models lose accuracy when irrelevant images sit beside the target, and the loss persists even when the prompt names the target image, because middle-layer distractor-to-text connections carry the interference into reasoning. Using the single-image prediction as a reference, answer consistency regularization aligns answer distributions with and without distractors on Qwen2.5-VL, raising circular accuracy by an average of 16 percentage points across five perception benchmarks with five distractors while preserving single-image performance.

Vision-language models lose accuracy when irrelevant images sit beside the target, and the loss persists even when the prompt names the target image, because middle-layer distractor-to-text connections carry the interference into reasoning. Using the single-image prediction as a reference, answer consistency regularization aligns answer distributions with and without distractors on Qwen2.5-VL, raising circular accuracy by an average of 16 percentage points across five perception benchmarks with five distractors while preserving single-image performance.

In aerial drone imagery, a referred target often covers only tens of pixels and sits among visually similar vehicles or people, so a generative editor can draw a plausible mask around the wrong instance. DroneSegEdit freezes the open-weight LongCat-Image-Edit, trains only a lightweight LoRA to render candidate regions, then uses a language-grounded spatial verifier to strip foreground inside that candidate that the referring expression does not support. On the complete official RIS-LAD test set it reaches 52.78 mIoU, 7.16 points above the previous best ReSaP-8B, and stays competitive overall on UAVid-RIS and VDD-RIS, with the clearest gains on compact vehicles and humans.

In aerial drone imagery, a referred target often covers only tens of pixels and sits among visually similar vehicles or people, so a generative editor can draw a plausible mask around the wrong instance. DroneSegEdit freezes the open-weight LongCat-Image-Edit, trains only a lightweight LoRA to render candidate regions, then uses a language-grounded spatial verifier to strip foreground inside that candidate that the referring expression does not support. On the complete official RIS-LAD test set it reaches 52.78 mIoU, 7.16 points above the previous best ReSaP-8B, and stays competitive overall on UAVid-RIS and VDD-RIS, with the clearest gains on compact vehicles and humans.

The work proposes a "visual resolvability" design rule requiring that correct answers depend on visual evidence while the task stays learnable, and tests it on counterfactual coordinate scenes with standard GRPO and correctness-and-format rewards on Qwen2.5-VL-7B: on held-out 20-point scenes denser than any seen in training, discovery accuracy rises from 0.425 to 0.875; replacing test images with gray canvases drops accuracy to zero, while training on gray canvases at matched step 30 across four seeds yields essentially none of the gain, indicating the gain comes from training images; adding a caption that answers the training question, with the same images, reward and budget, cuts the gain by nearly two thirds.

The work proposes a "visual resolvability" design rule requiring that correct answers depend on visual evidence while the task stays learnable, and tests it on counterfactual coordinate scenes with standard GRPO and correctness-and-format rewards on Qwen2.5-VL-7B: on held-out 20-point scenes denser than any seen in training, discovery accuracy rises from 0.425 to 0.875; replacing test images with gray canvases drops accuracy to zero, while training on gray canvases at matched step 30 across four seeds yields essentially none of the gain, indicating the gain comes from training images; adding a caption that answers the training question, with the same images, reward and budget, cuts the gain by nearly two thirds.

The work proposes ReLaViS, which formulates GUI grounding as recurrent latent visual search within a single model interaction: at each step a spatial search head uses the hidden state to query the screenshot's visual tokens, producing a spatial search distribution that aggregates those tokens into latent visual evidence fed back as the next input embedding, supervised by GUI-aware coarse-to-fine trajectories built from flat element annotations; built on Qwen2.5-VL-7B, it outperforms matched single-step baselines on all five benchmarks, raising ScreenSpot-Pro accuracy from 53.2% to 56.3% with only 3.5% more inference FLOPs.

The work proposes ReLaViS, which formulates GUI grounding as recurrent latent visual search within a single model interaction: at each step a spatial search head uses the hidden state to query the screenshot's visual tokens, producing a spatial search distribution that aggregates those tokens into latent visual evidence fed back as the next input embedding, supervised by GUI-aware coarse-to-fine trajectories built from flat element annotations; built on Qwen2.5-VL-7B, it outperforms matched single-step baselines on all five benchmarks, raising ScreenSpot-Pro accuracy from 53.2% to 56.3% with only 3.5% more inference FLOPs.

The work proposes Behavior Pack Optimization (BPO), which replaces the single-response reward in video MLLM post-training with a jointly scored behavior pack of outputs across counterfactual views chosen by question type, requiring stability under irrelevant interventions, sensitivity when key evidence is removed, and abstention when no evidence remains, using the original-view response as an anchor-relative advantage; on TempCompass, MVBench, and NExT-QA, BPO improves Qwen2.5-VL-7B-Instruct macro accuracy by 4.7 pp, the temporal-hard subset by 7.8 pp, and abstention F1 by 20.0 pp over a budget-matched vanilla GRPO baseline from the same SFT checkpoint, transferring to Video-MME, LongVideoBench, and LLaVA-Video-7B.

The work proposes Behavior Pack Optimization (BPO), which replaces the single-response reward in video MLLM post-training with a jointly scored behavior pack of outputs across counterfactual views chosen by question type, requiring stability under irrelevant interventions, sensitivity when key evidence is removed, and abstention when no evidence remains, using the original-view response as an anchor-relative advantage; on TempCompass, MVBench, and NExT-QA, BPO improves Qwen2.5-VL-7B-Instruct macro accuracy by 4.7 pp, the temporal-hard subset by 7.8 pp, and abstention F1 by 20.0 pp over a budget-matched vanilla GRPO baseline from the same SFT checkpoint, transferring to Video-MME, LongVideoBench, and LLaVA-Video-7B.