CoVeR proposes coverage-based token pruning for multi-view 3D reasoning in VLMs
reasoning training
| Source: HF Papers | Original article
A new paper titled **CoVeR: Coverage‑Based Token Pruning for Multi‑View 3D Reasoning in VLMs** proposes a training‑free method to slash the visual token load of multi‑view vision‑language models while keeping most of their 3‑D reasoning power.
The authors observe that representing a scene through dozens of 2‑D views lets existing 2‑D VLMs reuse their massive pre‑training, but it also generates thousands of redundant visual tokens per query. CoVeR tackles this by selecting a spatial subset of tokens that guarantees full scene coverage and respects a strict token budget. In experiments on three standard 3‑D VLM benchmarks, the technique trims the input to roughly **8 % of the original visual tokens** yet retains **about 93.5 % of the full‑token accuracy**. Crucially, the approach requires no additional training, making it a drop‑in optimizer for any multi‑view pipeline.
The development matters because token explosion has become a bottleneck for scaling 3‑D reasoning with large VLMs. As we reported on September 9, 2026, methods such as BeaconKV have shown that cache‑aware token compression can accelerate inference for massive reasoning models. CoVeR extends that line of work to the visual domain, offering a practical way to deploy 3‑D capable VLMs on limited hardware or at lower cloud cost, and potentially widening access to applications like robotics, AR/VR, and spatial analytics that rely on multi‑view inputs.
Going forward, the community will watch for integration of CoVeR into open‑source VLM stacks and its impact on downstream tasks that demand real‑time 3‑D understanding. Further validation on larger, more diverse scene collections and on emerging multimodal models will reveal whether the coverage‑based pruning principle can become a standard component of efficient 3‑D AI pipelines.
Sources
Back to AIPULSEN