Conditional Program Shrinks LLM Lookup Tables 75% While Improving Fidelity
dnotitia · hf · 2026-10-20
A paper from Dnotitia Inc. and KAIST introduces Conditional Program (CP), a compressed representation for conditional-memory embedding tables in LLMs.
Instead of keeping the full table resident, CP reconstructs only the rows an input requests, representing each row as a short code plus a local scale, with the reconstruction program shared across the table and rows independently decodable.
Result: 75% storage reduction while improving fidelity over the uncompressed baseline.
More from Research
- Causal inference tutorial series covers DAG engines for separating correlation from causation — Nasereliver · 2026-10-20
- Google's FireSat constellation uses AI to spot wildfires as small as 5x5 meters from space — GoogleAI · 2026-10-20
- CyberArena benchmarks AI agents in live attack-and-defense combat — 2077AIDataFoundation · 2026-10-20
- Training MOSS-Image LoRAs locally on 8GB VRAM: 500 steps beats 8000, full settings inside — AcademiaSD · 2026-10-20
- PNAS study: brief AI chats drive 17,950 people to sign petitions and donate, attitudes don't predict behavior — _FelixSimon_ · 2026-10-20
- Citrini Used Anthropic's Internal Model to Design Scorpion Antivenom Proteins Overnight — citrini · 2026-10-20