benjaminhan - Network

5/nREFERENCES[1] Orion Weller, Michael Boratko, Iftekhar Naim, and Jinhyuk Lee. 2025. On the theoretical limitations of embedding-based...

5/n

REFERENCES

[1] Orion Weller, Michael Boratko, Iftekhar Naim, and Jinhyuk Lee. 2025. On the theoretical limitations of embedding-based retrieval. https://arxiv.org/abs/2508.21038

[2] Yifu Qiu, Varun Embar, Yizhe Zhang, Navdeep Jaitly, Shay B. Cohen, and Benjamin Han. 2025. Eliciting in-context Retrieval and reasoning for long-context large language models. https://machinelearning.apple.com/research/eliciting-in-context repo: https://github.com/apple/ml-icr2

#AI #LLM #RAG #Embeddings #Retrieval #RecSys #Search #AgenticAI #paper

4/4. Three ways out: (1) Cross-encoders—placing docs in the prompt of a Long-Context LM; accurate but costly. (2) Multi-vector retrieval....

4/

4. Three ways out: (1) Cross-encoders—placing docs in the prompt of a Long-Context LM; accurate but costly. (2) Multi-vector retrieval. (3) Sparse retrieval.

For cross-encoders, this links directly to our earlier work on ICR² [2], where combining training data design with model re-architecting improved retrieval performance (see picture 4). Many other paths remain open!

#AI #LLM #RAG #Embeddings #Retrieval #RecSys #Search #AgenticAI #paper

3/3. LIMIT dataset: To stress-test real models, they construct the LIMIT dataset consisting of 50k docs and 1k queries, each with 2 relevant...

3/

3. LIMIT dataset: To stress-test real models, they construct the LIMIT dataset consisting of 50k docs and 1k queries, each with 2 relevant docs (picture1). All single-vector models fail badly, while BM25 and multi-vector methods perform much better (see picture 2).

#AI #LLM #RAG #Embeddings #Retrieval #RecSys #Search #AgenticAI #paper

2/2. Oracle experiment: They empirically confirm this bound by directly optimizing embeddings (“free embeddings”) against the relevance...

2/

2. Oracle experiment: They empirically confirm this bound by directly optimizing embeddings (“free embeddings”) against the relevance matrix. The critical corpus size grows only cubically with dimension. For example, with embedding dimension d = 1024, you can only represent all possible 2-doc query combinations up to about 4 million documents — far below typical web-scale retrieval needs.

#AI #LLM #RAG #Embeddings #Retrieval #RecSys #Search #AgenticAI #paper

1/Embeddings are the beating heart of modern AI—powering RAG and serving as memory for agentic AI. But a new paper [1] shows a ceiling:1....

1/

Embeddings are the beating heart of modern AI—powering RAG and serving as memory for agentic AI. But a new paper [1] shows a ceiling:

1. Dot-product retrieval is bounded by embedding dimension d; if the relevance matrix has sign-rank r, then d >= r is required—and no amount of training can avoid it.

#AI #LLM #RAG #Embeddings #Retrieval #RecSys #Search #AgenticAI #paper

Skipped today’s #Parkrun, and shadow-ran with my daughter in the community 5K race. It's her first race coming back to the school...

Skipped today’s #Parkrun, and shadow-ran with my daughter in the community 5K race. It's her first race coming back to the school cross-country team after *two* years in recovery from her running injury. She did quite well in the first mile, but ran out of steam after finishing the first 2 miles. But this was only the first week, and i know she will outrun me in no time!

I did manage to sprint a bit at the end -- 4'28"/mi -- my fastest stride ever! 🙂

(half marathon race next Monday!)

#running

Coachable human — achievement unlocked!#coachGPT #running

Coachable human — achievement unlocked!

#coachGPT #running

So much wonderful tech and engineering goes into this teeny tiny daily driver. (you also get to learn why dogs tilt their heads when...

So much wonderful tech and engineering goes into this teeny tiny daily driver.

(you also get to learn why dogs tilt their heads when confused)

https://youtu.be/PB_8dGKh9JI?feature=shared

#airpods #apple

Maier, #Franck, #Schumann: Sonatas for #Violin & #Piano Duo...

Maier, #Franck, #Schumann: Sonatas for #Violin & #Piano Duo Concertante

https://classical.music.apple.com/us/album/1817421838?l=en-US

#classicalMusic #appleMusicClassical #commute #chamber

Beware my #British friends. https://mstdn.social/@Free_Press/115099995291652362

Beware my #British friends. https://mstdn.social/@Free_Press/115099995291652362

Oliver Davis & Antonio Vivaldi: SeasonsTrafalgar Sinfonia, Grace Davidson, Kerenza Peacock(looking forward to...

Oliver Davis & Antonio Vivaldi: Seasons
Trafalgar Sinfonia, Grace Davidson, Kerenza Peacock

(looking forward to fall)

https://classical.music.apple.com/us/album/1024672141?l=en-US

#classicalMusic #appleMusicClassical #commute #vocal

Cause and effect #running #health https://mas.to/@SmudgeTheInsultCat/115092405352945410

Cause and effect

#running #health https://mas.to/@SmudgeTheInsultCat/115092405352945410

“Exposure to heat waves over just two years could add up to 12 extra days of age-related health damage.”Humans produce greenhouse gases...

“Exposure to heat waves over just two years could add up to 12 extra days of age-related health damage.”

Humans produce greenhouse gases -> global warming -> people age faster -> population decreases faster -> reduced greenhouse gases

https://www.nytimes.com/2025/08/25/climate/heat-waves-aging.html

#climateCrisis #Heat #health #earth

#Scarlatti: #piano sonatasJavier Perianeshttps://classical.music.apple.com/us/album/1810545013?l=en-US#classicalMusic #appleMusicClassical...

#Scarlatti: #piano sonatas
Javier Perianes

https://classical.music.apple.com/us/album/1810545013?l=en-US

#classicalMusic #appleMusicClassical #commute

“Check all constants and make sure they are immutable.”Life is WONDERFUL!#genAI #coding #RooCode

“Check all constants and make sure they are immutable.”

Life is WONDERFUL!

#genAI #coding #RooCode

Happy Monday! Let’s watch #Pika fight first! https://youtu.be/Q2ua2luJW00#animals #video #bbc

Happy Monday! Let’s watch #Pika fight first! https://youtu.be/Q2ua2luJW00

#animals #video #bbc

Great reading while training for #marathon races, I guess??? 🤔Are Marathon Runners More Likely to Get Cancer? - VICE...

Great reading while training for #marathon races, I guess??? 🤔

Are Marathon Runners More Likely to Get Cancer? - VICE https://apple.news/A9uZaF_VXSxqNdeM5Tt6cjQ

#running #health

Adding this rule to project documentation guidelines to prevent #RooCode from going into the infinite-loop hell.#genAI #coding

Adding this rule to project documentation guidelines to prevent #RooCode from going into the infinite-loop hell.

#genAI #coding

Today’s run went according to #ChatGPT coach’s plan. The coach even gave a remark on the little drama I encountered! 😆#Running

Today’s run went according to #ChatGPT coach’s plan. The coach even gave a remark on the little drama I encountered! 😆

#Running

J. S. #Bach & Sons: #Flute SonatasToshiyuki Shibata, Anthony...

J. S. #Bach & Sons: #Flute Sonatas
Toshiyuki Shibata, Anthony Romaniuk

https://classical.music.apple.com/us/album/1816634319?l=en-US

#classicalMusic #appleMusicClassical #commute