Flamingo: a Visual Language Model for Few-Shot Learning

Fitting a new task from a handful of examples required fine-tuning on thousands of times more task data.

Interleaves vision and text in one sequence and inserts PERCEIVER RESAMPLER blocks between the frozen vision backbone and the frozen language model. The resampler maps a variable-length visual token sequence to a small fixed number of vectors via learned latent queries with cross-attention, so the LM's sequence length grows by a constant regardless of image count or video length.

NeurIPS 2022 (arXiv comment 'In Proceedings of Neural Information Processing Systems (NeurIPS)'; the NeurIPS 2022 proceedings index contains the title). Weights were never released, hence research-only and open_weights false. THE KEY INFERENCE MECHANISM FOR THIS REPO IS THE PERCEIVER RESAMPLER: it makes vision-token count a tunable constant independent of input resolution and modality count. Every later resampler-based VLM (Q-Former in BLIP-2, SALMONN's audio resampler) is a variation on this, and it is the direct alternative to pruning, with the tradeoff that resampling is lossy but fixed-cost while pruning keeps full fidelity only where it matters. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. This is STRUCTURAL rather than researched: neurips has no Findings volume and no workshop track in which a paper could appear, so 'main' is the only possible honest value and no per-paper evidence was needed. The field is annotated here so that 'findings' and 'workshop' remain MEANINGFUL - a filter for findings=true must return only the papers that actually are findings, which requires every other paper to be explicitly NOT findings.