Research
73 updates on Research.
FunctionGemma 270M fine-tune reaches 76.5% on 15 mobile actions, with an Android demo
Fine-tuning FunctionGemma 270M on 9,500 synthetic mobile-action conversations raises end-to-end accuracy from 29.3% to 76.5%, with a Pixel 8 Pro demo app.
PalmClaw is an open-source agent framework that runs natively on Android phones
PalmClaw runs sessions, memory, tools and the agent loop on the phone and reports 97.1% task success on 70 mobile tasks.
Graphcore runs Llama 3.2 11B Vision on a Pixel 8a with 2.7-bit quantisation
Graphcore Research fits Llama-3.2-11B-Vision-Instruct into 3.6 GB with a 2.7-bit format and runs it on a Pixel 8a at 3.8 tok/s.
S2-MoE speeds up MoE decoding by up to 5.3x in llama.cpp
S2-MoE is a self-speculative decoding method for mixture-of-experts models that reaches up to 5.3x speedup and about 2.0x on average.
Online-SDFT reports 70.28% routing accuracy and updates a LoRA adapter on the phone
I-Ju Lin and Zhang-Wei Hong measure 70.28% accuracy against 52.78% for the best baseline, and run the LoRA update itself on an Android phone.
MobiBench paper benchmarks small LLMs in llama.cpp on two laptops and an RTX 3060
MobiBench is a llama.cpp benchmark suite for speed, memory and accuracy of small LLMs, but the reported tests ran on two laptops and not on phones.
FBLayout fine-tunes transformers on phone GPUs 2.2 to 5.7 times faster
A MobiSys 2026 paper fine-tunes seven transformer models on phone GPUs 2.2 to 5.7 times faster than MNN, TFLite and TVM, with 4.2 times fewer cache misses.
NPUs lead LLM prefill, CPUs lead decoding on Snapdragon phones
A measurement study of five mobile LLM frameworks finds NPUs excel at prefill, CPUs at decoding, and framework gaps of up to 10x on NPUs.
Google adds multi-token prediction to Gemini Nano v3 on Pixel 9 and 10
Google added a prediction head to frozen Gemini Nano v3 on Pixel 9 and 10, saving 130 MB per instance over a standalone drafter.
llada.cpp runs a diffusion LLM on a Snapdragon NPU up to 42 times faster
Tsinghua and Beihang researchers report LLaDA-8B generating 128 tokens 17 to 42 times faster on a Hexagon NPU than on the phone CPU.
CAPED redacts phone screenshots before a cloud GUI agent sees them
Researchers at CUHK, Southeast University and Shandong University report a phone-side screenshot filter that cuts seeded leakage from 0.766 to 0.268.
Quant.npu runs 4-bit LLMs up to 15.1 percent faster on a Qualcomm NPU
Eight authors report a fully static integer quantisation method that runs 4-bit models up to 15.1 percent faster on a Qualcomm SM8650 NPU.