π New preprint: Ripple
On unified-memory edge devices like DGX Spark, LLM/VLM inference can starve time-critical robot CPU tasks of bandwidth.
Ripple turns CPU deadline slack into GPU memory-byte budgets, no bandwidth counters needed.
https://t.co/9SvdZxo1Wl
@mbzuai 3οΈβ£ On DGX Spark (GB10), memory search adds only 7β87 ms, which is small next to model inference.
Takeaway: on-device memory needs better evidence selection and real access control, not just bigger models.
Thanks to my advisor Prof. Xiaosong Ma. π Code and data coming soon.
Our paper MemArena got accepted as a poster at #NeurIPS2026 ! π I did this work in my first year of undergrad at @MBZUAI.
It's an ego-centric benchmark for on-device personal memory assistants: 50 simulated agents, 15 days, 10.3M dialogue tokens.
π https://t.co/dwfij8tz4P
π§΅π
@mbzuai Across 5 models Γ 5 memory backends:
1οΈβ£ The backend matters more than model size: switching backends gives +32.5pp recall, while scaling Qwen3 from 0.6B to 32B gives +10.6pp.
2οΈβ£ Permission-aware access fails everywhere. Systems either leak the private fact or never retrieve it.
@mbzuai A personal assistant should only know what its user actually saw, and it should know when not to share it.
Existing memory benchmarks give models an omniscient view of sparse, disconnected sessions. MemArena uses one coherent simulated world with dense, human-scale daily dialog.
Nice to witness the conferral of the inaugural honorary doctorate by @mbzuai to @sama Sam Altman! Cool to be an undergraduate student studying AI in Abu Dhabi!
If you are an alumni from NFLS εδΊ¬ε€ε½θ―ε¦ζ ‘ and currently work/study in the Middle East(especially Dubai and Abu Dhabi), I will be more than happy to know you. Currently I am an undergraduate in MBZUAI located in Masdar City, Abu Dhabi. I'm hoping for more connections!