The Flow has not summarised this story yet — read it at arXiv Computation and Language.
The paper introduces S‑Space, a linear subspace within multimodal language models where object‑token representations encode continuous spatial coordinates along horizontal, vertical, and distance axes. This subspace is intrinsic, readable under non‑spatial queries, and serves as an internal spatial workspace: perception maps object locations into it, reasoning transforms it for viewpoint changes, and verbal responses selectively access it. Diagnostic experiments using S‑Space reveal that while models can form spatial representations, they struggle with spatial transformations needed for reliable reasoning across viewpoint changes.
MJev-Spatial is a multimodal decision model that unifies spatial intelligence tasks—choice, numerical, and point prediction—into a single classification framework. It treats each candidate as a class, scoring them with pretrained query and key projections without adding parameters, and achieves predictions in a single forward pass with no answer tokens. Trained on 12.3 M question‑answer pairs, it outperforms eight baseline models, ranks first on seven of twelve benchmarks, and runs 2.5× faster than the fastest autoregressive baseline.
The paper investigates why multimodal large language models (MLLMs) struggle with vision‑centric tasks when visual evidence conflicts with pretrained language knowledge. Using image reconstruction and a new WhatIfVis benchmark, the authors show that MLLMs preserve coarse‑grained visual attributes but fail to consistently use them, and that supervised fine‑tuning and activation patching can improve controllability of visual context sensitivity. The study demonstrates that the main bottleneck lies in the models’ inability to reliably regulate their reliance on visual evidence rather than in visual perception itself.
The paper introduces “SpaceConflict”, a benchmark of 23,196 multimodal inputs designed to test whether large language models can not only report spatial facts but also use them in reasoning. Experiments show a gap between a model’s ability to recover an initial spatial state from visual evidence and its ability to apply that state to perform transformations, with the gap narrowing but not closing as model size increases. To address this, the authors propose Operational State Supervision (OSS), which supervises task‑relevant spatial states and their transformation trajectories, improving accuracy on judgments that require state organization and use.
arXiv:2606. 11719v1 Announce Type: cross Abstract: Spatial reasoning remains a persistent challenge for multimodal large language models (MLLMs).
The paper introduces PCSR-Bench, a benchmark of 84,373 question‑answer pairs derived from 2,600 omnidirectional images across 26 indoor environments, designed to evaluate perspective‑conditioned spatial reasoning (PCSR) in multimodal large language models (MLLMs). It reports a significant perception–reasoning gap, with accuracy dropping from 57.59% on limited field‑of‑view reasoning to as low as 0.64% on open‑ended compositional directional chains. An RL‑based diagnostic study on a 7B‑scale model shows that reward shaping can improve performance to 60.06% on a controlled task, indicating partial plasticity of PCSR capabilities.