Mamba-caption: Long-range sequence modelling for efficient and accurate image captioning
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Image captioning has been a problem in vision–language research for a long time. Long-range dependencies and efficiency are challenges for the standard models, such as recurrent neural networks (RNNs) and Transformers. To overcome this, we present Mamba-Caption, an efficient sequence processing model that replaces attention mechanisms with selective state-space modelling. The core novelty is a Mamba-based decoder that substitutes self-attention with selective state-space updates, enabling linear-time caption generation while preserving long-range token dependencies; this decoder is a drop-in language-side component that conditions on a convolutional neural network (CNN) image embedding without domain-specific heuristics. Our model utilizes a CNN encoder, a token embedding layer, and a Mamba-based decoder; the decoder is trained using teacher forcing with a cross-entropy objective. Our model outperforms baselines on all standard metrics when evaluated on the Flickr30k dataset, achieving a Bilingual Evaluation Understudy (BLEU-1) score of 0.83, a Metric for Evaluation of Translation with Explicit ORdering (METEOR) score of 0.79, a Recall-Oriented Understudy for Gisting Evaluation—Longest Common Subsequence (ROUGE-L) score of 0.73, and a Consensus-based Image Description Evaluation (CIDEr) score of 1.30. We further contextualize efficiency via a qualitative/complexity discussion and ablation framing that isolates decoder-side design choices, reinforcing that the gains in efficiency do not sacrifice accuracy. Mamba-Caption can be applied to real-world captioning tasks due to its high efficiency and generalizability.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it