Jiaming (Jamin) Cheng

After completing my B.S. in Computer and Information Science (cum laude) at The Ohio State University in 2024, I have continued doing research with the group: day to day with Subhransu Das, on projects led by Profs. Rajiv Ramnath and Brijesh Soni.

I care about making capable AI usable without a data center behind it — running on the devices people already carry, and without shipping their data elsewhere to do it.

Doing that has taught me to distrust proxies for efficiency and go measure the thing itself. On the compression side the thing itself is the hardware: structured pruning, low-bit quantization, and knowledge distillation, with segmentation models cut by up to 87% in size and parameters and deployed on a Raspberry Pi inside a few watts, and phase-wise benchmarks that show where on-device LLM latency goes.

The same distrust applies away from the device. Multi-agent LLM systems that relay KV caches credit their gains to the “latent thoughts” they exchange; swapping in a mismatched cache shows a large cache effect need not be a pairing effect — across three released systems, one relay carries example-specific content, one carries it over a subset of layers, and for one the transfer is not detected.

I am applying to PhD programs for Fall 2027.

news

Oct 01, 2025	SPICE accepted to IEEE CCNC 2026 as an oral paper.

selected publications

Preprint
When Does Latent Communication Pay? A Causal Audit of Relayed KV Caches in Multi-Agent LLMs

Jiaming Cheng, Subhransu Das, and Rajiv Ramnath

2026

Preprint

Abs arXiv Bib

Multi-agent LLM systems relay key-value caches instead of text and credit their gains to exchanged "latent thoughts". That credit is a claim about which example’s cache is relayed, not merely that one is. We audit it causally in released systems, replacing the cache with deranged (mismatched-example), zeroed, and moment-matched random counterparts under two regimes defined by whether the receiver needs the sender’s private information. Where it does, the battery reads ceiling: 100% against 23-25% for answer-irrelevant relays on the primary backbone, replicated across three families, five checkpoints, and a prose document-QA surface. Where it does not, a pre-registered five-seed protocol establishes equivalence within 2.8 points under Holm-corrected TOST. Benchmark deltas do not by themselves establish latent-thought transmission; establishing it takes a mismatched-cache audit, which we release.
@misc{cheng2026latentaudit, title = {When Does Latent Communication Pay? A Causal Audit of Relayed {KV} Caches in Multi-Agent {LLMs}}, author = {Cheng, Jiaming and Das, Subhransu and Ramnath, Rajiv}, year = {2026}, archiveprefix = {arXiv}, primaryclass = {cs.CR}, note = {Preprint}, }

PEARC

Phase-Wise Analysis of LLM Inference Acceleration on GPU, CPU, and Edge Device

Subhransu Das, Jiaming Cheng, Swathi Vallabhajosyula, and 2 more authors

In Practice and Experience in Advanced Research Computing (PEARC ’26), 2026

To appear

Bib

@inproceedings{das2026phasewise,
  title = {Phase-Wise Analysis of {LLM} Inference Acceleration on {GPU}, {CPU}, and Edge Device},
  author = {Das, Subhransu and Cheng, Jiaming and Vallabhajosyula, Swathi and Soni, Brijesh and Ramnath, Rajiv},
  booktitle = {Practice and Experience in Advanced Research Computing (PEARC '26)},
  year = {2026},
  note = {To appear},
}

CCNC

SPICE: Structured Pruning for Inference on Constrained Edge Devices

Subhransu Das, Jiaming Cheng, Aniruddha Rakshit, and 3 more authors

In IEEE Consumer Communications & Networking Conference (CCNC), 2026

Bib

@inproceedings{das2026spice,
  title = {{SPICE}: Structured Pruning for Inference on Constrained Edge Devices},
  author = {Das, Subhransu and Cheng, Jiaming and Rakshit, Aniruddha and Soni, Brijesh and Boubin, Jayson and Ramnath, Rajiv},
  booktitle = {IEEE Consumer Communications \& Networking Conference (CCNC)},
  year = {2026},
}

PEARC

EPIC: Efficient Pruning for Inference on Constrained Devices

Subhransu Das, Jiaming Cheng, Aniruddha Rakshit, and 2 more authors

In Practice and Experience in Advanced Research Computing (PEARC ’25), 2025

Bib

@inproceedings{das2025epic,
  title = {{EPIC}: Efficient Pruning for Inference on Constrained Devices},
  author = {Das, Subhransu and Cheng, Jiaming and Rakshit, Aniruddha and Boubin, Jayson and Ramnath, Rajiv},
  booktitle = {Practice and Experience in Advanced Research Computing (PEARC '25)},
  year = {2025},
}