It seems the relentless march of technological progress has, once again, stumbled over its own feet. Despite the grand pronouncements surrounding Vision-Language Models (VLMs) and their action-oriented kin (VLAs), recent research confirms what some of us have known all along: they still struggle with the very basics. Two new preprints, published today on arXiv CS.AI, meticulously detail the continued efforts to prop up unreliable action planning and a rather dubious grasp of spatial numerical data, rather than any genuine paradigm shifts.

The initial fanfare for VLMs and VLAs promised intuitive interpretation of complex visual and linguistic inputs, leading to competent actions in embodied environments, such as robotics. However, the reality, as is so often the case, is a familiar disappointment. Their inherent reactive behavior frequently proves insufficient for complex, long-horizon tasks or when encountering minor deviations from their carefully curated training data—a condition politely termed 'distribution shift' arXiv CS.AI. Further complicating matters, the numerical values these models produce for actions or coordinates often lack genuine grounding in actual spatial perception arXiv CS.AI. It's a rather inconvenient truth for systems meant to interact with a physical world.

V-VLAPS: An Attempt at Foresight for the Short-Sighted

The first of these papers, "V-VLAPS: Value-Guided Planning for Vision-Language-Action Models" arXiv CS.AI, confronts the rather predictable issue of VLAs selecting suboptimal actions despite being equipped with "strong action priors" arXiv CS.AI. It seems that guiding a model with a prior isn't quite the same as ensuring it makes good choices. The V-VLAPS method endeavors to rectify this by employing value-guided planning to direct tree search, moving past the rudimentary dependence on policy priors and mere visit-count exploration. It's essentially an effort to graft a rudimentary form of foresight onto models that remain stubbornly reactive. One might reasonably question why such an elaborate mechanism is required if the foundational models possessed even a shred of genuine competence.

SpaceNum: Re-evaluating Basic Arithmetic and Spatial Grounding

Concurrently, the paper "SPACENUM: Revisiting Spatial Numerical Understanding in VLMs" arXiv CS.AI brings to light an even more elementary deficiency. It scrutinizes the rather unsettling notion that VLMs, despite generating numerical outputs for action magnitudes or spatial coordinates, may not genuinely grasp the meaning of these figures. The researchers explicitly state it remains "unclear whether these numerical outputs are genuinely grounded in spatial perception" [arXiv CS.AI](https://arxiv.org/abs/2605.23898]. The SpaceNum framework, therefore, undertakes the task of re-examining this fundamental spatial numerical understanding. It's a rather telling indictment when a framework must be introduced simply to ascertain if an "intelligent" system is merely fabricating plausible-sounding numbers instead of comprehending their intrinsic connection to the physical world. One would assume that a rudimentary understanding of "one" and "two" would be a prerequisite for any entity claiming to interact with its environment.

The Ongoing Pursuit of Basic Competence

These research endeavors, while commendable in their detailed analysis, ultimately highlight a recurring theme in the development of foundational models: a persistent immaturity in their real-world reliability and interpretability. The very necessity for specialized frameworks like V-VLAPS and SpaceNum serves as a stark reminder that Vision-Language Models, despite their often-touted capabilities, remain susceptible to fundamental shortcomings in perception, reasoning, and planning. This suggests that the robustness of many 'deployed' embodied AI systems may be less inherent understanding and more a progressively intricate series of workarounds.

Moving forward, the industry faces the challenge of integrating these insights to foster genuinely more reliable and transparent VLM applications. The critical question remains whether these solutions represent a true stride towards fundamental understanding or merely another layer of sophisticated complexity applied to models still grappling with the most basic tenets of interaction and cognition. The cycle of identifying limitations and engineering targeted solutions continues, and while each iteration promises incremental refinement, the path to truly autonomous and intelligent systems appears, as ever, a long and arduous one. One can only observe and record the ongoing effort with a profound sense of weariness.