Gemma 4 Technical Report
Paper • 2607.02770 • Published • 57
Distillation Hugs
community
AI & ML interests
None defined yet.
Recent Activity
View all activity
osanseviero
authored a
paper 3 days ago
osanseviero
submitted a
paper to Daily Papers 4 days ago
Post
1219
🤗 Announcing the Ettin Reranker family: six new state-of-the-art CrossEncoder rerankers for search from 17M to 1B parameters, plus the full training data and the ~150-line recipe. Built on the Ettin ModernBERT encoders, Apache 2.0. Details:
All six were trained with the same single-stage pointwise MSE distillation recipe, with mixedbread-ai/mxbai-rerank-large-v2 (1.54B) as the teacher. Only the learning rate and per-device batch size change between sizes. The 1B student matches the teacher within 0.0001 NDCG@10 on MTEB(eng, v2) Retrieval, the 150M is the strongest reranker I tested in the under-600M range, and the 17M beats the 33M ms-marco-MiniLM-L12-v2 by +0.051 NDCG@10 at roughly half the parameter count.
Speed matters as much as quality for a reranker, since it determines whether the model fits the latency budget between retrieval and showing results. Our 17M is the fastest reranker in the whole comparison at 7517 pairs/sec on an H100. Our 150M runs 2.3x faster than the two other 150M ModernBERT-base rerankers (gte-reranker-modernbert-base and granite-embedding-reranker-english-r2) because the modular Transformer module propagates unpadded inputs through every layer rather than just the FA2 attention kernel. And our 1B is 2.4x faster than its 1.5B teacher while matching it on quality.
I bootstrapped the training recipe with the new train-sentence-transformers Agent Skill shipped in Sentence Transformers v5.5.0. Install it with
I wrote a blog post walking through usage, results across six embedder pairings, the speed story, and the complete training script. Check it out, or just point your Agent to the URL:
https://huggingface.co/blog/ettin-reranker
Collection: https://huggingface.co/collections/cross-encoder/ettin-rerankers
All six were trained with the same single-stage pointwise MSE distillation recipe, with mixedbread-ai/mxbai-rerank-large-v2 (1.54B) as the teacher. Only the learning rate and per-device batch size change between sizes. The 1B student matches the teacher within 0.0001 NDCG@10 on MTEB(eng, v2) Retrieval, the 150M is the strongest reranker I tested in the under-600M range, and the 17M beats the 33M ms-marco-MiniLM-L12-v2 by +0.051 NDCG@10 at roughly half the parameter count.
Speed matters as much as quality for a reranker, since it determines whether the model fits the latency budget between retrieval and showing results. Our 17M is the fastest reranker in the whole comparison at 7517 pairs/sec on an H100. Our 150M runs 2.3x faster than the two other 150M ModernBERT-base rerankers (gte-reranker-modernbert-base and granite-embedding-reranker-english-r2) because the modular Transformer module propagates unpadded inputs through every layer rather than just the FA2 attention kernel. And our 1B is 2.4x faster than its 1.5B teacher while matching it on quality.
I bootstrapped the training recipe with the new train-sentence-transformers Agent Skill shipped in Sentence Transformers v5.5.0. Install it with
hf skills add train-sentence-transformers --claude and ask Claude Code (or Codex / Cursor / Gemini CLI) to fine-tune a SentenceTransformer, CrossEncoder, or SparseEncoder model on your data.I wrote a blog post walking through usage, results across six embedder pairings, the speed story, and the complete training script. Check it out, or just point your Agent to the URL:
https://huggingface.co/blog/ettin-reranker
Collection: https://huggingface.co/collections/cross-encoder/ettin-rerankers
Post
545
🤖 I've just published Sentence Transformers v5.5.0, headlined by a new train-sentence-transformers Agent Skill that lets your AI coding agent (Claude Code, Codex, Cursor, Gemini CLI, ...) train and finetune embedding, reranker, and sparse encoder models for you. Plus training losses & fixes. Details:
The skill bundles curated guidance for the whole training workflow across all three model types: base model selection, loss and evaluator choice, hard-negative mining, distillation, LoRA, Matryoshka, multilingual training, static embeddings, etc. It also ships production-ready training template scripts the agent can adapt. Install it with
On the loss side: EmbedDistillLoss is a new embedding-level distillation loss for SentenceTransformer. Instead of distilling teacher scores like MarginMSELoss, it aligns the student's embeddings directly with pre-computed teacher embeddings, wtih an optional learnable projection for when the student and teacher dimensions differ. Second, ADRMSELoss is a new listwise learning-to-rank loss for CrossEncoder from the Rank-DistilLLM paper, aimed at the LLM-distillation reranking setting.
encode() and predict() also gained a per-call processing_kwargs override, so you can change processor settings like max_length, a vision-language model's image resolution, or a video's fps, for a single call without rebuilding the model.
The Agent Skill is the part of this release I'm most keen for people to try. Curious to hear how it works for you. I've been using it myself a lot to quickly set up some training runs that immediately use a bunch of best practices.
> pip install sentence-transformers==5.5.0
> hf skills add train-sentence-transformers
The full release notes: https://github.com/huggingface/sentence-transformers/releases/tag/v5.5.0
The skill bundles curated guidance for the whole training workflow across all three model types: base model selection, loss and evaluator choice, hard-negative mining, distillation, LoRA, Matryoshka, multilingual training, static embeddings, etc. It also ships production-ready training template scripts the agent can adapt. Install it with
hf skills add train-sentence-transformers, then just describe what you want, e.g. "finetune a reranker on my (question, answer) pairs, mine hard negatives, and push it to the Hub".On the loss side: EmbedDistillLoss is a new embedding-level distillation loss for SentenceTransformer. Instead of distilling teacher scores like MarginMSELoss, it aligns the student's embeddings directly with pre-computed teacher embeddings, wtih an optional learnable projection for when the student and teacher dimensions differ. Second, ADRMSELoss is a new listwise learning-to-rank loss for CrossEncoder from the Rank-DistilLLM paper, aimed at the LLM-distillation reranking setting.
encode() and predict() also gained a per-call processing_kwargs override, so you can change processor settings like max_length, a vision-language model's image resolution, or a video's fps, for a single call without rebuilding the model.
The Agent Skill is the part of this release I'm most keen for people to try. Curious to hear how it works for you. I've been using it myself a lot to quickly set up some training runs that immediately use a bunch of best practices.
> pip install sentence-transformers==5.5.0
> hf skills add train-sentence-transformers
The full release notes: https://github.com/huggingface/sentence-transformers/releases/tag/v5.5.0
- 3 replies
Post
1082
🌐 I've just published Sentence Transformers v5.4 to make the project fully multimodal for embeddings and reranking. The release also includes a modular CrossEncoder, and automatic Flash Attention 2 input flattening. Details:
You can now use SentenceTransformer and CrossEncoder with text, images, audio, and video, with the same familiar API. That means you can compute embeddings for an image and a text query using model.encode(), compare them with model.similarity(), and it just works. Models like Qwen3-VL-Embedding-2B and jinaai/jina-reranker-m0 are supported out of the box.
Beyond multimodal, I also fully modularized the CrossEncoder class. It's now a torch.nn.Sequential of composable modules, just like SentenceTransformer has been. This unlocked support for generative rerankers (CausalLM-based models like mxbai-rerank-v2 and the Qwen3 rerankers) via a new LogitScore module, which wasn't possible before without custom code.
Also, Flash Attention 2 now automatically skips padding for text-only inputs. If your batch has a mix of short and long texts, this gives you a nice speedup and lower VRAM usage for free.
I wrote a blog post walking through the multimodal features with practical examples. Check it out if you want to get started, or just point your Agent to the URL: https://huggingface.co/blog/multimodal-sentence-transformers
This release has set up the groundwork for more easily introducing late-interaction models (both text-only and multimodal) into Sentence Transformers in the next major release. I'm looking forward to it!
You can now use SentenceTransformer and CrossEncoder with text, images, audio, and video, with the same familiar API. That means you can compute embeddings for an image and a text query using model.encode(), compare them with model.similarity(), and it just works. Models like Qwen3-VL-Embedding-2B and jinaai/jina-reranker-m0 are supported out of the box.
Beyond multimodal, I also fully modularized the CrossEncoder class. It's now a torch.nn.Sequential of composable modules, just like SentenceTransformer has been. This unlocked support for generative rerankers (CausalLM-based models like mxbai-rerank-v2 and the Qwen3 rerankers) via a new LogitScore module, which wasn't possible before without custom code.
Also, Flash Attention 2 now automatically skips padding for text-only inputs. If your batch has a mix of short and long texts, this gives you a nice speedup and lower VRAM usage for free.
I wrote a blog post walking through the multimodal features with practical examples. Check it out if you want to get started, or just point your Agent to the URL: https://huggingface.co/blog/multimodal-sentence-transformers
This release has set up the groundwork for more easily introducing late-interaction models (both text-only and multimodal) into Sentence Transformers in the next major release. I'm looking forward to it!
lewtun
submitted 2
papers to Daily Papers 5 months ago
Post
4487
🐦🔥 I've just published Sentence Transformers v5.2.0! It introduces multi-processing for CrossEncoder (rerankers), multilingual NanoBEIR evaluators, similarity score outputs in mine_hard_negatives, Transformers v5 support and more. Details:
- CrossEncoder multi-processing: Similar to SentenceTransformer and SparseEncoder, you can now use multi-processing with CrossEncoder rerankers. Useful for multi-GPU and CPU settings, and simple to configure: just
- Multilingual NanoBEIR Support: You can now use community translations of the tiny NanoBEIR retrieval benchmark instead of only the English one, by passing
- Similarity scores in Hard Negatives Mining: When mining for hard negatives to create a strong training dataset, you can now pass
- Transformers v5: This release works with both Transformers v4 and the upcoming v5. In the future, Sentence Transformers will only work with Transformers v5, but not yet!
- Python 3.9 deprecation: Now that Python 3.9 has lost security support, Sentence Transformers no longer supports it.
Check out the full changelog for more details: https://github.com/huggingface/sentence-transformers/releases/tag/v5.2.0
I'm quite excited about what's coming. There's a huge draft PR with a notable refactor in the works that should bring some exciting support. Specifically, better multimodality, rerankers, and perhaps some late interaction in the future!
- CrossEncoder multi-processing: Similar to SentenceTransformer and SparseEncoder, you can now use multi-processing with CrossEncoder rerankers. Useful for multi-GPU and CPU settings, and simple to configure: just
device=["cuda:0", "cuda:1"] or device=["cpu"]*4 on the model.predict or model.rank calls.- Multilingual NanoBEIR Support: You can now use community translations of the tiny NanoBEIR retrieval benchmark instead of only the English one, by passing
dataset_id, e.g. dataset_id="lightonai/NanoBEIR-de" for the German benchmark.- Similarity scores in Hard Negatives Mining: When mining for hard negatives to create a strong training dataset, you can now pass
output_scores=True to get similarity scores returned. This can be useful for some distillation losses!- Transformers v5: This release works with both Transformers v4 and the upcoming v5. In the future, Sentence Transformers will only work with Transformers v5, but not yet!
- Python 3.9 deprecation: Now that Python 3.9 has lost security support, Sentence Transformers no longer supports it.
Check out the full changelog for more details: https://github.com/huggingface/sentence-transformers/releases/tag/v5.2.0
I'm quite excited about what's coming. There's a huge draft PR with a notable refactor in the works that should bring some exciting support. Specifically, better multimodality, rerankers, and perhaps some late interaction in the future!
andito
authored a
paper 9 months ago
Post
4667
🤗 Sentence Transformers is joining Hugging Face! 🤗 This formalizes the existing maintenance structure, as I've personally led the project for the past two years on behalf of Hugging Face! Details:
Today, the Ubiquitous Knowledge Processing (UKP) Lab is transferring the project to Hugging Face. Sentence Transformers will remain a community-driven, open-source project, with the same open-source license (Apache 2.0) as before. Contributions from researchers, developers, and enthusiasts are welcome and encouraged. The project will continue to prioritize transparency, collaboration, and broad accessibility.
Read our full announcement for more details and quotes from UKP and Hugging Face leadership: https://huggingface.co/blog/sentence-transformers-joins-hf
We see an increasing wish from companies to move from large LLM APIs to local models for better control and privacy, reflected in the library's growth: in just the last 30 days, Sentence Transformer models have been downloaded >270 million times, second only to transformers.
I would like to thank the UKP Lab, and especially Nils Reimers and Iryna Gurevych, both for their dedication to the project and for their trust in myself, both now and two years ago. Back then, neither of you knew me well, yet you trusted me to take the project to new heights. That choice ended up being very valuable for the embedding & Information Retrieval community, and I think this choice of granting Hugging Face stewardship will be similarly successful.
I'm very excited about the future of the project, and for the world of embeddings and retrieval at large!
Today, the Ubiquitous Knowledge Processing (UKP) Lab is transferring the project to Hugging Face. Sentence Transformers will remain a community-driven, open-source project, with the same open-source license (Apache 2.0) as before. Contributions from researchers, developers, and enthusiasts are welcome and encouraged. The project will continue to prioritize transparency, collaboration, and broad accessibility.
Read our full announcement for more details and quotes from UKP and Hugging Face leadership: https://huggingface.co/blog/sentence-transformers-joins-hf
We see an increasing wish from companies to move from large LLM APIs to local models for better control and privacy, reflected in the library's growth: in just the last 30 days, Sentence Transformer models have been downloaded >270 million times, second only to transformers.
I would like to thank the UKP Lab, and especially Nils Reimers and Iryna Gurevych, both for their dedication to the project and for their trust in myself, both now and two years ago. Back then, neither of you knew me well, yet you trusted me to take the project to new heights. That choice ended up being very valuable for the embedding & Information Retrieval community, and I think this choice of granting Hugging Face stewardship will be similarly successful.
I'm very excited about the future of the project, and for the world of embeddings and retrieval at large!
- 1 reply
Post
3024
Finally, our new paper is out! "𝗙𝗶𝗻𝗲𝗩𝗶𝘀𝗶𝗼𝗻: 𝗢𝗽𝗲𝗻 𝗗𝗮𝘁𝗮 𝗜𝘀 𝗔𝗹𝗹 𝗬𝗼𝘂 𝗡𝗲𝗲𝗱"! 🥳
FineVision: Open Data Is All You Need (2510.17269)
If you've ever trained a VLM, you know this problem: nobody shares their data mixtures. It's a black box, making replicating SOTA work impossible.
We wanted to change that.
FineVision unifies 200 sources into 24 million samples. With 17.3 million images and 9.5 billion answer tokens, it's the largest open resource of its kind.
In the paper, we share how we built it:
🔍 finding and cleaning data at scale
🧹 removing excessive duplicates across sources
🤗 decontaminating against 66 public benchmarks
My favorite part is Figure 6 (in the video!). It's our visual diversity analysis. It shows that FineVision isn't just bigger; it's more balanced and conceptually richer than other open datasets.
NVIDIA's Eagle 2 paper highlighted just how critical this visual diversity is, and our results confirm it: models trained on FineVision consistently outperform those trained on any other open dataset on 11 benchmarks!
🎉 To celebrate the paper, I’m also releasing a concatenated and shuffled version of the full dataset! 👉
It’s ready to stream, so you can start training your own models right away:
from datasets import load_dataset
d = load_dataset("HuggingFaceM4/FineVision_full_shuffled", split="train", streaming=True)
print(next(iter(d)))
A big shoutout to the first authors: Luis Wiedmann and Orr Zohar. They are rockstars!
FineVision: Open Data Is All You Need (2510.17269)
If you've ever trained a VLM, you know this problem: nobody shares their data mixtures. It's a black box, making replicating SOTA work impossible.
We wanted to change that.
FineVision unifies 200 sources into 24 million samples. With 17.3 million images and 9.5 billion answer tokens, it's the largest open resource of its kind.
In the paper, we share how we built it:
🔍 finding and cleaning data at scale
🧹 removing excessive duplicates across sources
🤗 decontaminating against 66 public benchmarks
My favorite part is Figure 6 (in the video!). It's our visual diversity analysis. It shows that FineVision isn't just bigger; it's more balanced and conceptually richer than other open datasets.
NVIDIA's Eagle 2 paper highlighted just how critical this visual diversity is, and our results confirm it: models trained on FineVision consistently outperform those trained on any other open dataset on 11 benchmarks!
🎉 To celebrate the paper, I’m also releasing a concatenated and shuffled version of the full dataset! 👉
HuggingFaceM4/FineVision_full_shuffled It’s ready to stream, so you can start training your own models right away:
from datasets import load_dataset
d = load_dataset("HuggingFaceM4/FineVision_full_shuffled", split="train", streaming=True)
print(next(iter(d)))
A big shoutout to the first authors: Luis Wiedmann and Orr Zohar. They are rockstars!
- 4 replies
osanseviero
authored a
paper 10 months ago
Post
5898
ModernBERT goes MULTILINGUAL! One of the most requested models I've seen, The Johns Hopkins University's CLSP has trained state-of-the-art massively multilingual encoders using the ModernBERT architecture: mmBERT.
Model details:
- 2 model sizes:
- jhu-clsp/mmBERT-small
- jhu-clsp/mmBERT-base
- Uses the ModernBERT architecture, but with the Gemma2 multilingual tokenizer (so: flash attention, alternating global/local attention, unpadding/sequence packing, etc.)
- Maximum sequence length of 8192 tokens, on the high end for encoders
- Trained on 1833 languages using DCLM, FineWeb2, and many more sources
- 3 training phases: 2.3T tokens pretraining on 60 languages, 600B tokens mid-training on 110 languages, and 100B tokens decay training on all 1833 languages.
- Both models are MIT Licensed, and the full datasets and intermediary checkpoints are also publicly released
Evaluation details:
- Very competitive with ModernBERT at equivalent sizes on English (GLUE, MTEB v2 English after finetuning)
- Consistently outperforms equivalently sized models on all Multilingual tasks (XTREME, classification, MTEB v2 Multilingual after finetuning)
- In short: beats commonly used multilingual base models like mDistilBERT, XLM-R (multilingual RoBERTa), multilingual MiniLM, etc.
- Additionally: the ModernBERT-based mmBERT is much faster than the alternatives due to its architectural benefits. Easily up to 2x throughput in common scenarios.
Check out the full blogpost with more details. It's super dense & gets straight to the point: https://huggingface.co/blog/mmbert
Based on these results, mmBERT should be the new go-to multilingual encoder base models at 300M and below. Do note that the mmBERT models are "base" models, i.e. they're currently only trained to perform Mask Filling. They'll need to be finetuned for downstream tasks like semantic search, classification, clustering, etc.
Model details:
- 2 model sizes:
- jhu-clsp/mmBERT-small
- jhu-clsp/mmBERT-base
- Uses the ModernBERT architecture, but with the Gemma2 multilingual tokenizer (so: flash attention, alternating global/local attention, unpadding/sequence packing, etc.)
- Maximum sequence length of 8192 tokens, on the high end for encoders
- Trained on 1833 languages using DCLM, FineWeb2, and many more sources
- 3 training phases: 2.3T tokens pretraining on 60 languages, 600B tokens mid-training on 110 languages, and 100B tokens decay training on all 1833 languages.
- Both models are MIT Licensed, and the full datasets and intermediary checkpoints are also publicly released
Evaluation details:
- Very competitive with ModernBERT at equivalent sizes on English (GLUE, MTEB v2 English after finetuning)
- Consistently outperforms equivalently sized models on all Multilingual tasks (XTREME, classification, MTEB v2 Multilingual after finetuning)
- In short: beats commonly used multilingual base models like mDistilBERT, XLM-R (multilingual RoBERTa), multilingual MiniLM, etc.
- Additionally: the ModernBERT-based mmBERT is much faster than the alternatives due to its architectural benefits. Easily up to 2x throughput in common scenarios.
Check out the full blogpost with more details. It's super dense & gets straight to the point: https://huggingface.co/blog/mmbert
Based on these results, mmBERT should be the new go-to multilingual encoder base models at 300M and below. Do note that the mmBERT models are "base" models, i.e. they're currently only trained to perform Mask Filling. They'll need to be finetuned for downstream tasks like semantic search, classification, clustering, etc.
Post
4569
😎 I just published Sentence Transformers v5.1.0, and it's a big one. 2x-3x speedups of SparseEncoder models via ONNX and/or OpenVINO backends, easier distillation data preparation with hard negatives mining, and more:
1️⃣ Faster ONNX and OpenVINO backends for SparseEncoder models
Usage is as simple as
2️⃣ New
This new output format is immediately compatible with the MarginMSELoss and SparseMarginMSELoss for training SentenceTransformer, CrossEncoder, and SparseEncoder losses.
3️⃣ Gathering across devices
When doing multi-GPU training using a loss that has in-batch negatives (e.g. MultipleNegativesRankingLoss), you can now use
4️⃣ Trackio support
If you also upgrade
5️⃣ MTEB Documentation
We've added some documentation on evaluating SentenceTransformer models properly with MTEB. It's rudimentary as the documentation on the MTEB side is already great, but it should get you started.
Plus many more smaller features & fixes (crash fixes, compatibility with datasets v4, FIPS compatibility, etc.).
See the full release notes here: https://github.com/UKPLab/sentence-transformers/releases/tag/v5.1.0
Big thanks to all of the contributors for helping with the release, many of the features from this release were proposed by others. I have a big list of future potential features that I'd love to add, but I'm
1️⃣ Faster ONNX and OpenVINO backends for SparseEncoder models
Usage is as simple as
backend="onnx" or backend="openvino" when initializing a SparseEncoder to get started, but I also included utility functions for optimization, dynamic quantization, and static quantization, plus benchmarks.2️⃣ New
n-tuple-scores output format from mine_hard_negativesThis new output format is immediately compatible with the MarginMSELoss and SparseMarginMSELoss for training SentenceTransformer, CrossEncoder, and SparseEncoder losses.
3️⃣ Gathering across devices
When doing multi-GPU training using a loss that has in-batch negatives (e.g. MultipleNegativesRankingLoss), you can now use
gather_across_devices=True to load in-batch negatives from the other devices too! Essentially a free lunch, pretty big impact potential in my evals.4️⃣ Trackio support
If you also upgrade
transformers, and you install trackio with pip install trackio, then your experiments will also automatically be tracked locally with trackio. Just open up localhost and have a look at your losses/evals, no logins, no metric uploading.5️⃣ MTEB Documentation
We've added some documentation on evaluating SentenceTransformer models properly with MTEB. It's rudimentary as the documentation on the MTEB side is already great, but it should get you started.
Plus many more smaller features & fixes (crash fixes, compatibility with datasets v4, FIPS compatibility, etc.).
See the full release notes here: https://github.com/UKPLab/sentence-transformers/releases/tag/v5.1.0
Big thanks to all of the contributors for helping with the release, many of the features from this release were proposed by others. I have a big list of future potential features that I'd love to add, but I'm
Post
3108
Many VLMs claim to process hours of video. But can they follow the story?🤔
Today, we introduce TimeScope: The benchmark that separates true temporal understanding from marketing hype. Let's see how much VLMs really understand!⏳
We test three skills that matter for real-world use:
🔎 Localized Retrieval: Find a specific action.
🧩 Information Synthesis: Piece together scattered clues.
🏃 Fine-Grained Perception: Analyze detailed motion (e.g., count how many times a person swings an axe).
The results are in, and they're revealing. Only Gemini 2.5 pro handles 1-hour-long videos.
Performance drops sharply with duration, proving that long video understanding is still challenging. We've found the breaking points—now the community can start fixing them.📈
Want to learn more? TimeScope is 100% open-source. Benchmark your model and help us build the next generation of video AI.
📖 Blog:
https://huggingface.co/blog/timescope-video-lmm-benchmark
👩💻 Leaderboard & Demo: Apollo-LMMs/TimeScope
📊 Dataset: Apollo-LMMs/TimeScope
⚙️ Eval Code: https://github.com/EvolvingLMMs-Lab/lmms-eval
Today, we introduce TimeScope: The benchmark that separates true temporal understanding from marketing hype. Let's see how much VLMs really understand!⏳
We test three skills that matter for real-world use:
🔎 Localized Retrieval: Find a specific action.
🧩 Information Synthesis: Piece together scattered clues.
🏃 Fine-Grained Perception: Analyze detailed motion (e.g., count how many times a person swings an axe).
The results are in, and they're revealing. Only Gemini 2.5 pro handles 1-hour-long videos.
Performance drops sharply with duration, proving that long video understanding is still challenging. We've found the breaking points—now the community can start fixing them.📈
Want to learn more? TimeScope is 100% open-source. Benchmark your model and help us build the next generation of video AI.
📖 Blog:
https://huggingface.co/blog/timescope-video-lmm-benchmark
👩💻 Leaderboard & Demo: Apollo-LMMs/TimeScope
📊 Dataset: Apollo-LMMs/TimeScope
⚙️ Eval Code: https://github.com/EvolvingLMMs-Lab/lmms-eval
Post
4116
🧠👁️ Can AI visualize solutions?
Humans often solve visual problems by sketching ideas in our minds. What if Vision-Language Models (VLMs) could do something similar, not by generating full images, but by using internal “mental sketches”?
That’s the idea behind Mirage, a new framework that empowers VLMs to reason using latent visual tokens. Instead of just thinking in words, Mirage mixes in abstract visual representations that help the model solve complex tasks.
These aren't photorealistic images. They're compact, internal representations optimized purely to support reasoning.
🔧 Mirage is trained in two phases:
1) Grounding: It learns to produce latent tokens anchored in real images.
2) Refinement: The model drops the images and learns to generate visual tokens on its own.
📈 And yes, it works!
On challenging benchmarks like Visual Spatial Planning, Jigsaw puzzles, and Spatial Attention Tasks, Mirage clearly outperforms GPT-4o and other strong baselines.
Smart sketches > empty words.
By mimicking the way humans visualize solutions, Mirage gives AI a new kind of imagination, one that’s faster, more efficient, and more human-like.
Kudos to the teams at UMass Amherst and MIT behind this exciting work.
Check the paper: Machine Mental Imagery: Empower Multimodal Reasoning with Latent Visual Tokens (2506.17218)
Humans often solve visual problems by sketching ideas in our minds. What if Vision-Language Models (VLMs) could do something similar, not by generating full images, but by using internal “mental sketches”?
That’s the idea behind Mirage, a new framework that empowers VLMs to reason using latent visual tokens. Instead of just thinking in words, Mirage mixes in abstract visual representations that help the model solve complex tasks.
These aren't photorealistic images. They're compact, internal representations optimized purely to support reasoning.
🔧 Mirage is trained in two phases:
1) Grounding: It learns to produce latent tokens anchored in real images.
2) Refinement: The model drops the images and learns to generate visual tokens on its own.
📈 And yes, it works!
On challenging benchmarks like Visual Spatial Planning, Jigsaw puzzles, and Spatial Attention Tasks, Mirage clearly outperforms GPT-4o and other strong baselines.
Smart sketches > empty words.
By mimicking the way humans visualize solutions, Mirage gives AI a new kind of imagination, one that’s faster, more efficient, and more human-like.
Kudos to the teams at UMass Amherst and MIT behind this exciting work.
Check the paper: Machine Mental Imagery: Empower Multimodal Reasoning with Latent Visual Tokens (2506.17218)
- 4 replies
Post
3214
‼️Sentence Transformers v5.0 is out! The biggest update yet introduces Sparse Embedding models, encode methods improvements, Router module for asymmetric models & much more. Sparse + Dense = 🔥 hybrid search performance! Details:
1️⃣ Sparse Encoder Models
Brand new support for sparse embedding models that generate high-dimensional embeddings (30,000+ dims) where <1% are non-zero:
- Full SPLADE, Inference-free SPLADE, and CSR architecture support
- 4 new modules, 12 new losses, 9 new evaluators
- Integration with @elastic-co , @opensearch-project , @NAVER LABS Europe, @qdrant , @IBM , etc.
- Decode interpretable embeddings to understand token importance
- Hybrid search integration to get the best of both worlds
2️⃣ Enhanced Encode Methods & Multi-Processing
- Introduce encode_query & encode_document automatically use predefined prompts
- No more manual pool management - just pass device list directly to encode()
- Much cleaner and easier to use than the old multi-process approach
3️⃣ Router Module & Advanced Training
- Router module with different processing paths for queries vs documents
- Custom learning rates for different parameter groups
- Composite loss logging - see individual loss components
- Perfect for two-tower architectures
4️⃣ Comprehensive Documentation & Training
- New Training Overview, Loss Overview, API Reference docs
- 6 new training example documentation pages
- Full integration examples with major search engines
- Extensive blogpost on training sparse models
Read the comprehensive blogpost about training sparse embedding models: https://huggingface.co/blog/train-sparse-encoder
See the full release notes here: https://github.com/UKPLab/sentence-transformers/releases/v5.0.0
What's next? We would love to hear from the community! What sparse encoder models would you like to see? And what new capabilities should Sentence Transformers handle - multimodal embeddings, late interaction models, or something else? Your feedback shapes our roadmap!
1️⃣ Sparse Encoder Models
Brand new support for sparse embedding models that generate high-dimensional embeddings (30,000+ dims) where <1% are non-zero:
- Full SPLADE, Inference-free SPLADE, and CSR architecture support
- 4 new modules, 12 new losses, 9 new evaluators
- Integration with @elastic-co , @opensearch-project , @NAVER LABS Europe, @qdrant , @IBM , etc.
- Decode interpretable embeddings to understand token importance
- Hybrid search integration to get the best of both worlds
2️⃣ Enhanced Encode Methods & Multi-Processing
- Introduce encode_query & encode_document automatically use predefined prompts
- No more manual pool management - just pass device list directly to encode()
- Much cleaner and easier to use than the old multi-process approach
3️⃣ Router Module & Advanced Training
- Router module with different processing paths for queries vs documents
- Custom learning rates for different parameter groups
- Composite loss logging - see individual loss components
- Perfect for two-tower architectures
4️⃣ Comprehensive Documentation & Training
- New Training Overview, Loss Overview, API Reference docs
- 6 new training example documentation pages
- Full integration examples with major search engines
- Extensive blogpost on training sparse models
Read the comprehensive blogpost about training sparse embedding models: https://huggingface.co/blog/train-sparse-encoder
See the full release notes here: https://github.com/UKPLab/sentence-transformers/releases/v5.0.0
What's next? We would love to hear from the community! What sparse encoder models would you like to see? And what new capabilities should Sentence Transformers handle - multimodal embeddings, late interaction models, or something else? Your feedback shapes our roadmap!
conceptofmind
authored 2
papers about 1 year ago
andito
authored a
paper about 1 year ago