What counts as visual evidence?

A video can look convincing while violating a task rule. An answer can be correct without using the video. An evaluator can write a fluent explanation without making its judgment correct. These are different problems, and they need different checks.

I coauthored VidGround, VideoScore2, VGI-Bench, Structured Defect Grounding (SDG), and MedClaw. This guide connects their research questions; it does not combine their scores or claim that one system solves all five problems.

Five questions, different evidence

1. VidGround: did the model need to watch?

VidGround finds that text cues alone can answer many questions in the video benchmarks and post-training datasets it studies. It curates questions that require visual grounding and examines their use in post-training. This makes the training data—not just the optimization algorithm—a central part of the evidence. A useful reading question is whether improvements also survive evaluation that controls for text-only shortcuts. Read the paper.

2. VideoScore2: what does the score mean?

VideoScore2 evaluates visual quality, text-to-video alignment, and physical/common-sense consistency, with explanatory rationales. It uses human-annotated VideoFeedback2 data and supervised training followed by GRPO. These dimensions help separate a good-looking clip from one that follows its prompt. My caution is that an explanation remains a model output: inspect agreement with human judgments and performance beyond the training distribution, not only the fluency of the rationale. Read the paper · Project.

3. VGI-Bench: was the process valid?

VGI-Bench tests visual reasoning expressed through generated video. Its task design considers inputs compatible with video models, evolving processes, and difficulty that is challenging but partly feasible. A plausible final frame alone is insufficient. The paper analyzes failure modes, sensitivity to input conditions, and limits of correction during denoising. My takeaway is to inspect the transitions that make an outcome valid, not to use visual polish as a substitute for task success. Read the paper.

4. SDG: which defect are we discussing?

SDG represents each defect with a location, type, reason, and importance. It introduces SDG-30K and SDG-Eval, then uses predicted defect sets to derive spatial rewards for alignment. This connects localized diagnosis to feedback, rather than compressing every failure into one scalar. The question for a reader is whether a stated reason actually refers to the marked region and the intended requirement. Read the paper · Code · Dataset page.

5. MedClaw: where is the temporal evidence?

MedClaw separates a text-only planner from frozen vision tools that inspect surgical video. It develops retrieval skills through a gradient-free, validation-reward-gated loop. Its benchmark includes long neurosurgery recordings and a held-out public lecture-video split. Here, the useful object to inspect is the evidence-gathering trace: what was viewed, what was revisited, and how it supports a before/after claim. This is research on video reasoning, not evidence of clinical deployment or patient benefit. Read the paper · Project.

A reading checklist, not a combined leaderboard

For a visual AI result, I would first ask:

1. Evidence: could the same answer come from text alone, and which frames or regions matter?
2. Criterion: are we judging appearance, prompt alignment, a valid process, or a temporal claim?
3. Validation: what independent reference checks the model’s explanation or score?
4. Transfer: which conclusions hold beyond the data, task conditions, and feedback used during development?

This checklist is my synthesis, not a new experiment or an evaluation protocol shared by the five papers. Their datasets, outputs, and scoring rules differ. Choose the paper that matches the failure you need to understand, then inspect its examples, protocol, and limitations.

Continue with the full publication list or the self-development evaluation guide.

Prepared with AI assistance and checked against the linked public papers and project pages on September 6, 2026. This article reproduces no paper figures, experimental tables, or patient images. Shared authorship is disclosed above; the questions and cautions are editorial interpretation, not additional measured results.