Research

# Research, explained

The papers behind modern multimodal search: retrieval, vector indexing, embeddings, segmentation, and personalization: each explained in about a minute, with what it actually means for Mixpeek.

All (17)

AlphaWiSE(1)

Apple(1)

Baidu(1)

Carnegie Mellon & Princeton(1)

EvoGraph-R1 (CVPR)(1)

Google(1)

Google Research(1)

Human–AI Study(1)

Meta(4)

Meta AI(1)

Meta AI (FAISS)(1)

Microsoft(1)

Stanford & Illuin (ColPali)(1)

University of Tokyo(1)

[

![How Reverse Video Search Works: Embeddings, Time, and Copy Detection](https://cdn.mixpeek.com/videos/thumbs/85-reverse-explainer.jpg)75s

Meta AI

## How Reverse Video Search Works: Embeddings, Time, and Copy Detection

Meta's Video Similarity Challenge (VSC23) is the benchmark that defines reverse video search as a research problem: given a query video, detect whether any segment of it was copied from a reference corpus, and localize exactly which segment. It is the reason modern copy detection matches a sequence of per-second descriptors rather than a single whole-video embedding.



](/research/video-similarity-challenge)[

![Google's MUVERA: The One-Vector Trick Behind Fast Multi-Vector Search](https://cdn.mixpeek.com/videos/thumbs/78-muvera.jpg)79s

Google Research

## Google's MUVERA: The One-Vector Trick Behind Fast Multi-Vector Search

MUVERA collapses a multi-vector (ColBERT-style) representation into a single Fixed Dimensional Encoding whose dot product provably approximates the multi-vector score: so late-interaction quality can ride an ordinary single-vector ANN index.



](/research/muvera)[

![Meta's SAM 3: Type a Phrase, Segment Every Instance](https://cdn.mixpeek.com/videos/thumbs/79-sam3.jpg)81s

Meta

## Meta's SAM 3: Type a Phrase, Segment Every Instance

SAM 3 introduces promptable concept segmentation: a short noun phrase or exemplar image makes one model detect, segment, and track every instance of that concept across images and video.



](/research/sam3)[

![Baidu's Unlimited-OCR: The Model Whose Memory Never Grows](https://cdn.mixpeek.com/videos/thumbs/77-unlimited-ocr.jpg)80s

Baidu

## Baidu's Unlimited-OCR: The Model Whose Memory Never Grows

Baidu's Unlimited-OCR replaces decoder attention with Reference Sliding Window Attention, turning the KV cache into a fixed-size queue so 40+ page PDFs parse in one forward pass instead of being chopped at page breaks.



](/research/unlimited-ocr)[

![Mamba Explained: The LLM Architecture With No Attention (5x Faster, Million-Token Context)](https://cdn.mixpeek.com/videos/thumbs/69-mamba.jpg)65s

Carnegie Mellon & Princeton

## Mamba Explained: The LLM Architecture With No Attention (5x Faster, Million-Token Context)

Mamba replaces attention with a selective state space model, one fixed-size state updated per token with a learned keep-or-forget decision, reporting linear scaling and 5x Transformer inference throughput with quality holding to million-token sequences.



](/research/mamba)[

![LoRA Fine-Tuning Explained on a Real Model: Train 1% of Qwen2.5-1.5B](https://cdn.mixpeek.com/videos/thumbs/75-lora.jpg)76s

Microsoft

## LoRA Fine-Tuning Explained on a Real Model: Train 1% of Qwen2.5-1.5B

LoRA freezes the base model and trains a small low-rank adapter beside each weight matrix, roughly 1% of the parameters, merged in afterward with zero added inference latency.



](/research/lora)[

![Late Interaction: How ColBERT and ColPali Search Documents as Images](https://cdn.mixpeek.com/videos/thumbs/82-late-interaction.jpg)77s

Stanford & Illuin (ColPali)

## Late Interaction: How ColBERT and ColPali Search Documents as Images

Late interaction keeps one embedding per token (ColBERT) or per image patch (ColPali) and scores query-document pairs with MaxSim, recovering fine-grained matching that single-vector embeddings flatten away: ColPali retrieves documents as images, no OCR.



](/research/colbert-colpali)[

![Cluster with Auctions: The FAISS Team Teaches Vector Indexes to Learn From Queries](https://cdn.mixpeek.com/videos/thumbs/70-cwa.jpg)61s

Meta AI (FAISS)

## Cluster with Auctions: The FAISS Team Teaches Vector Indexes to Learn From Queries

Most IVF vector indexes assign queries with the same function that partitioned the data. Cluster with Auctions jointly learns a balanced partition and a neural probing function, reassigning vectors by a large-scale auction while the probe learns the live query distribution.



](/research/cluster-with-auctions)[

![Learned Indexes Explained: When a Model Replaces the B-Tree](https://cdn.mixpeek.com/videos/thumbs/46-learned-indexes.jpg)106s

Google

## Learned Indexes Explained: When a Model Replaces the B-Tree

The classic case for learned index structures: a model that has learned the data distribution can predict where a key lives, replacing B-trees and hash maps with smaller, faster learned models, the idea that seeded a decade of learned-index research.



](/research/learned-indexes)[

![How Meta Cut Its Ads Model Serving Cost From 24.6x to 3.9x (HILL, Explained)](https://cdn.mixpeek.com/videos/thumbs/63-meta-hill.jpg)74s

Meta

## How Meta Cut Its Ads Model Serving Cost From 24.6x to 3.9x (HILL, Explained)

Meta's teardown of MoNN (the retrieval model behind Facebook/Instagram ads) plus HILL, the hierarchical index that made it affordable: organizing every item into a learned tree keeps +6.0% recall at 3.9x cost instead of 24.6x.



](/research/meta-hill-ads)[

![Meta SiRA: The Retrieval System That Beats Dense Embeddings With Zero Training](https://cdn.mixpeek.com/videos/thumbs/30-meta-sira-retrieval.jpg)87s

Meta

## Meta SiRA: The Retrieval System That Beats Dense Embeddings With Zero Training

SIRA is a learned sparse retrieval approach: instead of a dense embedding, documents and queries expand into weighted term sets, keeping the exact-match precision and inverted-index efficiency of lexical search while learning which terms matter.



](/research/sira)[

![How Apple TV Ranks Search Results From Two Keystrokes (RecSys 2026)](https://cdn.mixpeek.com/videos/thumbs/71-appletv.jpg)63s

Apple

## How Apple TV Ranks Search Results From Two Keystrokes (RecSys 2026)

Apple's search-as-you-type system gives every title two embeddings, a text encoder fine-tuned on co-engagement and an ID embedding learned from co-watching, and fuses per-keystroke similarities with boosted trees for +2.99% NDCG@10.



](/research/appletv-search)[

![EvoGraph-R1: The Knowledge Graph That Edits Itself (CVPR 2026)](https://cdn.mixpeek.com/videos/thumbs/72-evograph.jpg)56s

EvoGraph-R1 (CVPR)

## EvoGraph-R1: The Knowledge Graph That Edits Itself (CVPR 2026)

EvoGraph-R1 treats a knowledge graph as a workspace agents edit (retrieve, reason, then write back new nodes, corrected edges, and multimodal hyperedges) so every query leaves the graph slightly smarter.



](/research/evograph)[

![Composed Image Retrieval With No Vision Model: 44% R@1 Zero-Shot](https://cdn.mixpeek.com/videos/thumbs/73-vfcir.jpg)63s

University of Tokyo

## Composed Image Retrieval With No Vision Model: 44% R@1 Zero-Shot

Composed image retrieval ('this jacket, but in red, with a hood') done with images represented purely as text: a caption plus an attribute checklist at ingest, an LLM to merge the modification, plain text retrieval, then an LLM reranker, 44.04% R@1 on CIRR zero-shot.



](/research/vision-free-cir)[

![Why Fine-Tuning Quietly Breaks Your Embedding Index (and AlphaWiSE's Fix)](https://cdn.mixpeek.com/videos/thumbs/74-alphawise.jpg)68s

AlphaWiSE

## Why Fine-Tuning Quietly Breaks Your Embedding Index (and AlphaWiSE's Fix)

Fine-tune a CLIP-style model and the cross-modal alignment that made old queries work shifts underneath your index: recall erodes with no warning. AlphaWiSE interpolates two frozen checkpoints with one learned coefficient per tensor to keep both, at zero added inference cost.



](/research/alphawise)[

![Brain2Qwerty: What Meta's Brain-to-Text Decoder Actually Does](https://cdn.mixpeek.com/videos/thumbs/57-brain2qwerty.jpg)87s

Meta

## Brain2Qwerty: What Meta's Brain-to-Text Decoder Actually Does

Brain2qwerty decodes text from non-invasive brain recordings as a person types: a research milestone in reading language from neural signals without implanted electrodes.



](/research/brain2qwerty)[

![Ten Minutes of AI Help Made People Worse: The Persistence Study](https://cdn.mixpeek.com/videos/thumbs/58-ai-persistence.jpg)79s

Human–AI Study

## Ten Minutes of AI Help Made People Worse: The Persistence Study

A controlled study found people who leaned on an AI assistant for ten minutes performed worse on their own afterward and gave up on hard problems faster — a caution about assistance that replaces understanding rather than supporting it.



](/research/ai-persistence)