Current Video Large Language Models (Video LLMs) excel in question answering (QA) but largely operate as black boxes, providing textual answers without verifiable visual grounding. Existing explainability efforts rely on textual rationales or sparse bounding boxes, which struggle to capture complex video dynamics such as occlusions and non-rigid deformations. We propose Evidence-Backed Video Question Answering (E-VQA), a novel task requiring models to jointly output a semantic answer and precise spatio-temporal evidence: temporal segments and dense, tracked object segmentation masklets. To support this, we introduce ST-Evidence, the first human-verified benchmark for both discriminative and generative pixel-level grounding. Evaluations of state-of-the-art models reveal a critical decoupling between QA accuracy and true visual perception that scaling alone fails to bridge. To address this, we develop scalable, automated generation pipelines to create ST-Evidence-Instruct, a 160k-scale dataset bridging high-level reasoning with fine-grained grounding. Fine-tuning grounded Video LLMs on this data yields substantial gains over the corresponding size-matched UniPixel baselines (e.g., +27.2 t-mean and +13.8 J&F on a 7B model), establishing a robust baseline for explainable, evidence-backed video understanding. Code and data are available at https://github.com/SalesforceAIResearch/EVQA.
Alex: Welcome to another episode of ResearchPod. Today, we're looking at a paper that tackles a significant limitation in how AI models watch and understand video.
Sam: I've noticed this. You can ask an AI about a video clip, and it gives a detailed summary — but how do we actually know it's seeing what it describes?
Alex: That's the central puzzle. Current models are fluent conversationalists but unreliable witnesses. They can describe a scene in detail while having no verifiable connection to what's actually on screen.
Sam: So the paper is asking: how do we force these models to prove their work, rather than just guessing based on patterns in language?
Alex: Exactly. The researchers propose a new task they call Evidence-Backed Video Question Answering. Instead of just producing a text answer, the model must also identify the specific time segment in the video where the answer occurs — and produce a pixel-level map of the relevant objects.
Sam: A pixel-level map? That sounds far more precise than just saying "the dog is in the kitchen."
Alex: It is. Think of it like a maths exam. You don't just want the final answer — you want to see every step of the working, so you can confirm the student didn't just guess. The researchers call these maps "masklets." A masklet is a sequence of shapes that traces the exact outline of a specific object as it moves through a video, frame by frame.
Sam: So instead of drawing a rough box around something, the model has to follow the actual contour of the object across time.
Alex: Right. And by requiring that output, the model is structurally prevented from relying on its internal language habits to fake an answer. It has to point to something real.
Sam: But why is that so difficult for current models? They seem capable enough in other areas.
Alex: The problem is that most of these models are what researchers call "black boxes." They take in video and produce text, but there's no internal mechanism that links those words to specific visual data. Think of it like a student who has read every book ever written about swimming — they can describe every stroke in perfect detail — but they've never actually been in a pool.
Sam: So they might just be predicting the most plausible sentence to follow a video, rather than genuinely analysing what's on screen.
Alex: That's the core concern. Researchers call this "language bias." The model leans on what it expects to happen — based on patterns from its training data — rather than what is actually happening in front of it.
Sam: That sounds genuinely risky for something like a self-driving car. If the system just predicts what it expects to see, it could miss a real hazard that doesn't fit the pattern.
Alex: Exactly. In high-stakes settings — medical imaging, autonomous vehicles, security systems — you need a verifiable audit trail for every decision. You need to be able to ask: what, precisely, did the model look at?
Sam: So how did the researchers actually test whether models could do this?
Alex: They built a new benchmark called ST-Evidence — the first of its kind to use human-verified, pixel-level data to grade how well a model connects its reasoning to the visual evidence. And then they ran a revealing experiment: they "corrupted" the video by pixelating the specific objects the model had identified as its evidence, and watched what happened to accuracy.
Sam: And when they hid the evidence, the model couldn't answer correctly anymore.
Alex: Precisely. Conversely, when they pixelated random background objects that weren't relevant, accuracy barely moved. That's a meaningful confirmation that the model was genuinely using those pixels to reason — not just picking up on visual noise.
Sam: But is the model doing the reasoning itself, or is it leaning on a separate tool to find the objects?
Alex: That's a critical distinction, and the paper addresses it directly. They compared their approach against a proxy tool — essentially a separate system that handles the visual search — and found that the bottleneck isn't the tool's ability to see. It's the model's ability to reason about what it sees. Even when the visual data is correct, the model often fails to connect it to the right answer.
Sam: So the "brain" is the limiting factor, not the "eyes."
Alex: That's the core finding. And it's why the fine-tuned model — trained on their new dataset, which they call ST-Evidence-Instruct — shows meaningful improvement. It's learning that visual grounding isn't optional. It's a requirement for earning the answer.
Sam: It's a shift from "predict the next word" to "prove the next fact."
Alex: That's a clean way to put it. And the practical upshot is something like live debugging. If a model makes a mistake, you can examine the masklets and see exactly which pixels it was attending to — and where its reasoning went wrong.
Sam: That feels like a meaningful step toward AI we can actually audit. Though I imagine the quality of that training data matters a lot.
Alex: The paper is clear on that point. Performance is still bounded by the quality of the data used to train the model. If the annotations are flawed, the model inherits those flaws. It's more verifiable, but it isn't perfect.
Sam: So it's not a complete solution — but it's a more honest framework. You can see where the model is looking, which means you can see where it goes wrong.
Alex: That's a fair summary. The broader implication is that scaling these models up with more data alone isn't sufficient. What's needed are fundamental changes in how they're trained — specifically, training them to treat visual evidence and language as genuinely integrated, rather than treating the visual input as optional context.
Sam: It's a more demanding standard, but probably the right one for systems we're being asked to trust.
Alex: Thanks for listening to ResearchPod.