arXiv:2602.22453v4 Announce Type: replace Abstract: Retrieval heads, a subset of attention heads in Transformers, were studied in English, showing its crucial role in retrieving information from the...
arXiv:2605. 21854v2 Announce Type: replace-cross Abstract: Vision-Language-Action (VLA) models have rapidly converged on a small set of architectural patterns: discrete-token autoregression (e.
The paper investigates how new concepts can be integrated into unified multimodal models (UMMs) by separating generation and understanding objectives through a novel visual entity bound to a single task direction. Experiments show that the effectiveness of cross‑task usability depends on where the concept is injected into the shared computation, with a mid‑stack alignment objective achieving high concept acquisition with minimal loss to overall performance. The study highlights that unified weights alone are insufficient; the two directions must share a semantic format at the entry point for efficient concept integration.
arXiv:2610.10526v1 Announce Type: cross Abstract: Vision-language-action models (VLAs) are strikingly sensitive to instruction phrasing and do not inherit the language robustness of the vision-langua...
arXiv:2609.00868v1 Announce Type: cross Abstract: Vision-language models are evaluated by aggregate accuracy on multimodal benchmarks, a practice that implicitly assumes the model uses its visual inp...
The paper argues that meaning identity—whether two sentences convey the same idea after wording changes—is not encoded in the geometry of independently produced sentence embeddings. Experiments on frozen off‑the‑shelf encoders and language models show that identity can only be reliably computed when both sentences are processed together in a single forward pass, yielding high accuracy (0.90–0.96) on PAWS‑X, whereas independent embeddings or simple fusion methods perform near chance. Even advanced bi‑encoder fine‑tuning improves performance on PAWS but fails to generalize to other similarity tasks, underscoring that identity is a cheap computed operator rather than a property of individual sentence vectors.
arXiv:2606. 03879v1 Announce Type: cross Abstract: As foundation models scale toward fusing more heterogeneous visual streams, understanding how diverse encoders interact under joint training becomes a prerequisite for principled design.
The paper introduces a perception interface that separates vision from language in vision‑language models. A frozen perception stack detects objects, a deterministic semantic serializer converts the perceived state into text, and a standard text‑only large language model (LLM) answers questions. Experiments on a campus‑robot benchmark show that this serialized interface outperforms a zero‑shot VLM of the same language‑model size, especially as the language model shrinks, and that the advantage persists under paraphrase and different supervision regimes.
The paper investigates how the modality gap— the separation between image and text representations in contrastive vision‑language models—affects different downstream tasks. By showing that a single dominant direction accounts for most of the image‑text mean separation, the authors explain why reducing or removing this gap can improve zero‑shot classification, degrade retrieval, or restore performance depending on the task. The study provides a geometric framework that clarifies when and why gap interventions should be applied in vision‑language systems.
arXiv:2608.23253v1 Announce Type: cross Abstract: Vision-language models typically encode an image into hundreds of visual tokens, incurring substantial inference latency and GPU memory overhead. Exi...
The paper investigates the projector training stage in multimodal large language models, showing that only a small fraction of data is needed and that projectors from different datasets can be interpolated with little loss. It proposes PORTAL, a training‑free initialization that uses optimal transport between Gaussian approximations of vision and language features, eliminating the need for paired image‑caption data. PORTAL matches or surpasses the standard pipeline across multiple backbones, encoders, and benchmarks.
Long-context evaluations often test whether a model can recover distant evidence, but recoverability does not guarantee behavioral influence. We test the prediction that a distant discourse anchor can...