preprint open access

A Pooled Score Is Not a Missing Binding: What Compositional Failures of Dual-Encoder Vision-Language Models Locate, Why Probes, Readouts and an Impossibility Proof Disagree About Where the Binding Is Lost, and the Controls That Would Assign It

Abstract

Life of Research

Dual-encoder vision-language models such as CLIP fail compositional tests that ask them to tell "a red cube and a blue sphere" from "a blue cube and a red sphere", and the failure has two rival diagnoses. One says the embeddings never encode the binding, because contrastive training on web captions does not reward it. The other, argued in an April 2026 study, says the binding is present and the global cosine readout throws it away: localised matching over frozen patch and token features raises out-of-domain scores where fine-tuning and five compositional training methods do not. In this analysis the author reviews and synthesises published results to show that the choice between "representation" and "inference" is posed one level too coarsely. A binding can be lost at five separable points: in the encoders' internal states, in pooling to one vector, in the cross-modal readout, in a benchmark that does not require it, or in a fix that does not travel. The evidence places losses at more than one point. Linear probes recover attribute-object bindings from pooled CLIP embeddings at 0.95 to 1.00 accuracy on synthetic scenes, while the same model matches across modalities at chance; this contradicts an impossibility proof only if the proof's premises are read as a description of CLIP. The April study's own tables tell a narrower story than its conclusion: against a stated chance level of 16.7, the alignment gains are concentrated on colour and, under structure-guided inference, material, while on size and quantity most variants stay at or below chance. Models that fuse tokens and patches, from cross-encoders to multimodal language models of more than 70 billion parameters, also fail binding tests whose distractors recombine the same attributes, so a global readout cannot be the whole cause. The paper consolidates the measured values in Table 1 and Figure 1, separates the five loss points in Table 2, states a procedure for locating a compositional failure as Algorithm 1, and specifies the controls that would assign it.

Pranay Mahendrakar, AI specialist

About the author

Pranay Mahendrakar is an ai specialist and large language model engineer based in Bengaluru, India. He builds production artificial intelligence systems and publishes open-access research on how those systems fail. See all 161 papers by Pranay Mahendrakar, or his ORCID record.

Questions about this paper

Who wrote "A Pooled Score Is Not a Missing Binding"?

Pranay Mahendrakar wrote "A Pooled Score Is Not a Missing Binding: What Compositional Failures of Dual-Encoder Vision-Language Models Locate, Why Probes, Readouts and an Impossibility Proof Disagree About Where the Binding Is Lost, and the Controls That Would Assign It", published 13 Nov 2026. Pranay Mahendrakar is an Indian AI specialist and LLM engineer based in Bengaluru, India. He is the Managing Director of SonyTech, Nodal Coordinator at IIRS-ISRO, and an instructor at Tutorials Point. His work covers large language models, natural language processing, computer vision and retrieval-augmented generation. He publishes open-access research papers and is the author of three books: Just AI With Pranay, Multiverse of AI and It's Me LLM.

Is "A Pooled Score Is Not a Missing Binding" free to read?

Yes. "A Pooled Score Is Not a Missing Binding" by Pranay Mahendrakar is open access under a Creative Commons Attribution 4.0 licence, with the full PDF available from Zenodo at https://doi.org/10.5281/zenodo.24198393. There is no paywall and no account required.

How do I cite "A Pooled Score Is Not a Missing Binding"?

Cite the DOI: Mahendrakar, P. (2026). A Pooled Score Is Not a Missing Binding: What Compositional Failures of Dual-Encoder Vision-Language Models Locate, Why Probes, Readouts and an Impossibility Proof Disagree About Where the Binding Is Lost, and the Controls That Would Assign It. Zenodo. https://doi.org/10.5281/zenodo.24198393 A BibTeX entry is provided on this page.

Related research by Pranay Mahendrakar

← All papers by Pranay Mahendrakar