EMNLP 2026 BoF on Embeddings, Rerankers and Small LMs for Better Search

In September, TypeSafe released Jev, a general-purpose discriminative model that returns typed decisions: a choice among options, a yes or no, or a score. TypeSafe positions it as a "System 1" that complements generative models. Han Xiao, who organizes this BoF, summarized his view:

Jev surprised me. A general-purpose discriminative model drew as much attention as a generative one, and TypeSafe's "System 1" framing fits it well. I have worked on discriminative models for years and never thought of taking them beyond search. My guess is that routing and tool calling move back to discriminative models.

Embedding models and rerankers are discriminative models too, which makes Jev a natural starting point for our 4th Birds of a Feather (BoF) session on embeddings, rerankers and small language models at EMNLP 2026 in Budapest. This year we especially welcome talks on Jev-like decision models and other general-purpose discriminative models, alongside retrieval work. The session runs as a mini-workshop of 10-minute lightning talks followed by an open discussion.

tagEvent Details

Date: Sunday, October 25, 2026
Time: 11:00 to 12:30, Budapest time (CET)
Location: Room G9-10 & G11, Gallery floor, HUNGEXPO Congress Centre, Budapest
Format: in-person, 10-minute lightning talks followed by open discussion

Walk-in attendance is super welcome. You do not need to register to attend. If you want to give a talk, you must register in advance with the form below and include a link to your slides or paper. We confirm selected talks by email before the session.

The slot is tentative until the EMNLP 2026 program on Underline is final. If it changes, we will update this page and email all registered speakers. Add to Google Calendar.

HUNGEXPO Congress Centre gallery floor with rooms G9, G10 and G11 marked — Rooms G9-10 and G11 are on the Gallery floor of the HUNGEXPO Congress Centre, next to Gallery Foyer II. Floor plan from HUNGEXPO.

tagRegister to Speak

If the form does not load, open it in a new tab.

tagTopics

We welcome any work on discriminative models and models that improve search. This year we are especially interested in:

Jev-like decision models with typed choice, yes/no and score outputs
General-purpose discriminative models beyond search
Routing, tool selection and classification for agents
Embedding and reranker architectures
Small language models for search and document understanding
Omni and multimodal retrieval over text, image, audio and video
Visual document retrieval and OCR-free pipelines
Reasoning-intensive retrieval and reasoning rerankers
Agentic search and deep research
Test-time compute for search
Synthetic data and distillation for retrieval models
On-device retrieval models
Benchmarks and evaluation

tagPrevious BoFs

We started this BoF at EMNLP 2023 in Singapore, where more than 80 people joined five lightning talks and a panel discussion.

Panel discussion at the EMNLP 2023 BoF in Singapore.

Lightning talk at the EMNLP 2023 BoF — Lightning talk on post-processing approaches for better embeddings, Singapore 2023.

Lightning talk comparing embedding models, rerankers and LLMs, Singapore 2023.

At EMNLP 2024 in Miami, we extended the scope to rerankers and small language models. Talks filled the whole 90 minutes, so the panel did not take place.

EMNLP 2024 BoF in the Miami Lecture Hall.

Lightning talk at the EMNLP 2024 BoF, Miami.

At EMNLP 2025 in Suzhou, the room was full again, with talks on jina-embeddings-v4, interpretable text embeddings and cross-lingual hard negatives.

Opening talk at the EMNLP 2025 BoF — Opening talk on trends in embeddings, rerankers and small language models, Suzhou 2025.

See you in Budapest on October 25.