Retrieval that actually works
Zero-shot, few-shot, and instruction-following retrieval systems that can handle diverse video collections and complex user intent.
ACL 2026 · La Jolla · In-person only
Video
Retrieval
A casual, high-signal ACL 2026 session for researchers thinking about video retrieval, multimodal retrieval, and efficient video understanding — from zero-shot systems to scalable search over millions of videos without melting GPUs.
Session Chair
Sourajit Saha
Location
La Jolla
Mode
In-person only
Why this session?
Video is everywhere, but retrieving the right video moment remains expensive, messy, and surprisingly unsolved. This session is for people who want systems that are not only accurate, but also scalable, interactive, and deployable.
Efficiency without sadness
Memory-efficient search, compressed representations, smarter candidate filtering, and practical ways to avoid brute-force everything.
Friendly academic chaos
Informal discussion, idea sharing, open problems, and probably too many opinions about embeddings, rerankers, and compute budgets.
Topics we want to discuss
Bring your papers, failed experiments, weird benchmark observations, and half-formed ideas. Especially the half-formed ideas.
Zero-shot video retrieval
Few-shot retrieval
CoT-guided retrieval
Multimodal reranking
Memory-efficient search
Compressed representations
Test-time adaptation
Interactive retrieval
Scalable video recognition
Annotation-efficient learning
Retrieving from millions of videos
Deployable retrieval systems
Suggested format
The session is designed to be lightweight and discussion-heavy. Replace the placeholder times below once the ACL schedule is finalized.
00–10 min
Opening: what are we even retrieving?
A quick framing of current video retrieval problems: queries, videos, moments, events, captions, embeddings, and where current systems break.
10–30 min
Lightning idea sharing
Short informal contributions from attendees: recent work, promising directions, painful bottlenecks, and open questions.
30–50 min
Open discussion
How do we retrieve from huge video databases? How do we reduce annotation burden? When should we use rerankers? What should future benchmarks measure?
50–60 min
Takeaways + possible collaborations
Wrap up with promising problems, shared resources, and people who should probably talk to each other after the session.
Come for the retrieval. Stay for the embeddings. Leave with at least one dangerously good research idea.
Where
La Jolla. This session is available in-person only.
Who should come?
Researchers working on video retrieval, multimodal retrieval, efficient video understanding, representation learning, reranking, or scalable search.
What to bring
Questions, opinions, open problems, negative results, benchmark frustrations, and ideas that are not fully baked yet.