Prior work calls the high attention on the first and last action tokens of a flow-based VLA policy a radial action sink, and uses these tokens to set the execution horizon. In language models, a sink is a token with high incoming attention, a small value, and little effect on the output. We test whether that profile holds, not the horizon rule, in one $π_{0.5}$ checkpoint on five RoboCasa tasks. The two endpoints receive 5.4% of action-key mass, 1.4 percentage points above their uniform share, and rotary positional encoding produces this excess: directly in shallow layers and through upstream action-token RoPE deeper in the network. The endpoints do not fit the sink profile. Their value norms stay near the action-token mean in all 18 layers. On 250 paired episodes, masking them or zeroing their values changes success by less than 4 points, and replacing their value directions with those of another observation lowers it by 3.2 to 7.2 points across three donors, whereas the same replacement on two non-endpoint tokens in the executed prefix lowers it by 28.4 points. The token that does fit the sink profile on mass and value norm is the prefix BOS; on this checkpoint, zeroing its value raises success in two paired runs, by +13.2 points (95% CI [+7.0, +19.4]) and +6.0 points ([+0.2, +11.8]). Inside the action chunk, a relative-offset routing profile estimated on other tasks keeps 75% of action-token spread and keeps closed-loop success on three held-out tasks (42.7% against a 38.7% baseline), while uniform or shuffled routing drops success to 3-4%. High attention mass alone does not tell what a token does.
Multimodal Large Language Models (MLLMs) have shown strong performance on video question answering, but their application to long-form videos is constrained by limited context length and computational cost, making keyframe sampling essential. Existing approaches typically rely on semantic relevance or reinforcement learning, which either fail to capture evidential clues or suffer from inefficient combinatorial optimization. In this work, we propose an evidence-driven keyframe sampling framework motivated by an information-theoretic perspective. We consider maximizing the conditional mutual information between the selected frames and the answer given the query as an idealized objective, and use a modular singleton-gain surrogate to motivate a tractable frame-wise selection criterion. Guided by this motivation, we train a query-conditioned evidence scoring network using annotated answer-supporting segments. A contrastive objective encourages greater aggregate score mass on frames from evidence regions than on background frames, providing a practical evidence-based criterion beyond semantic similarity. Experiments on long-form video understanding benchmarks show that our method consistently outperforms prior sampling strategies under comparable frame budgets. With Qwen2.5-VL-7B under the same 32-frame budget, our method improves over uniform sampling by 10.1 points on LVBench and 4.4 points on the long-video subset of VideoMME.
Multimodal large language models (MLLMs) have made substantial advancements in video understanding, yet the reliability of their responses remains underexplored. This work presents a diagnostic study of absent answer detection for MLLMs in video understanding, where the correct answer is deliberately excluded from the candidate set and a reliable model is expected to recognize that no valid option exists. We evaluate the absent answer detection behavior under three settings: multiple-choice questions augmented with a ``None of the Above'' option, open-ended generation with a detection instruction, and standard evaluation without any guidance. Across a diverse set of models and benchmarks, we find that MLLMs overwhelmingly select plausible distractors rather than detecting the absent answer. This failure is more pronounced in temporal reasoning tasks and worsens with denser frame sampling. We further explore chain-of-thought prompting as a mitigation strategy and find that while it substantially improves detection rates, performance remains unsatisfactory, suggesting that prompting-based strategies alone are insufficient to fully address this limitation. These findings expose a systematic failure in absent answer detection and highlight the need for explicit detection mechanisms in multimodal systems.