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

(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. 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.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-13
DOI
https://doi.org/10.5281/zenodo.24198392
Primary Topic
Multimodal Machine Learning Applications
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

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

Pranay M. Mahendrakar
Zenodo (CERN European Organization for Nuclear Research)
Multimodal Machine Learning Applications
preprint

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

Pranay M. Mahendrakar
preprint en

Abstract

(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. 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.

Zenodo (CERN European Organization for Nuclear Research)
Multimodal Machine Learning Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon