The AI @ ADI Reading Digest: Recommended papers and research about AI at the edge.

The May 2026 Edition

At ADI, AI means developing world‑class models that overcome latency, power, and cost constraints to deliver real‑world insights. For many of us, May kicks off a stretch of travel, field work, and (hopefully) a few outdoor breaks, perfect moments to queue up a shortlist of papers that are both practical and forward-looking. Whether you’re reading on a plane, sitting in a patio, or between meetings, here’s a set of research recommendations from our team, fresh ideas for getting more capability out of smaller, faster, edge-ready models.

📄 Sample-efficient Integration of New Modalities into Large Language Models

👍🏼 Recommended by: Kao Kitichotkul

💡 Why it’s worth reading: This work introduces SEMI, a practical framework for extending multimodal LLMs to entirely new, low‑resource modalities with minimal paired data. By training a hypernetwork on high‑resource modalities to generate lightweight adapters for a shared projector, SEMI enables few‑shot integration, on the order of tens of samples, without retraining the core model. The approach demonstrates large sample‑efficiency gains across diverse unseen modalities (e.g., satellite imagery, inertial sensors, molecules), making incremental modality expansion technically and economically viable.

📄 Reinforcement Learning for Machine Learning Engineering Agents

👍🏼 Recommended by: Debanjan Ghosh

💡 Why it’s worth reading: This paper challenges the prevailing assumption that strong ML engineering agents must rely on ever‑larger, static frontier models. Instead, it shows that smaller language models trained with reinforcement learning can outperform much larger prompted models by improving through experience. The authors introduce duration‑aware policy gradients to handle variable‑cost actions by penalizing agents that are solving tasks faster by discovering easier solutions and environment instrumentation by adding simple logging/print statements to provide dense, partial‑credit rewards in complex engineering workflows. On MLEBench, the RL‑trained agent exceeds Claude‑3.5‑Sonnet–based agents by ~22%, making a compelling case for learning‑based, self‑improving ML agents. The only caveat is that small models often do not follow complex instructions very well initially, so you need a strongly aligned small model to begin with.

📄 TOM: A Ternary Read-only Memory Accelerator for LLM-powered Edge Intelligence

👍🏼 Recommended by:

Comments

Join the conversation at EngineerZone

Opinions expressed in the content posted here are the personal opinions of the original authors, and do not necessarily reflect those of Analog Devices, Inc (ADI) or its subsidiaries. The content is provided for informational purposes only and is not meant to be an endorsement or representation by ADI or any other party. This site may also provide links or references to non-ADI sites and resources. ADI makes no representations, warranties, or other commitments whatsoever about any non-ADI sites or third-party resources that may be referenced, accessible from, or linked to this site.

💡 This paper presents a chip architecture that stores heavily quantized ternary LLM weights on chip and can run LLM inference at 1000 tokens per second in a power and area efficient manner. Their approach combines implementing ROM using logic gates rather than standard memory cells, and power gating the layers to prevent leakage loss. Their ROM implementation takes advantage of an imbalance between the number of 0s and 1s in the stored representations of ternary quantized weights, however the weight density improvements they show cannot be explained by just this imbalance, nevertheless they do not provide any further explanation. They demonstrate that their approach achieves better latency, throughput, power consumption and chip size than A100 GPU and is designed for edge deployment.

👍🏼 Recommended by: Yuval Zukerman

💡 Why it’s worth reading: This paper introduces NutriEar, an in‑ear acoustic sensing system that reframes food recognition around nutrition rather than brittle food IDs. Instead of classifying specific items, it maps chewing sounds to an eight‑class, food‑science‑grounded taxonomy that jointly encodes dominant macronutrient and texture (e.g., Crunchy‑Fat, Soft‑Protein). The key insight is a hybrid pipeline: acoustic features plus contrastively fine‑tuned CLAP embeddings that generalizes across users and conditions, achieving ~80% LOSO accuracy. It’s a compelling example of physiology‑aware representation learning on earables with real implications for continuous, privacy‑preserving dietary monitoring.

👍🏼 Recommended by: Giulia Vilone

💡 Why it’s worth reading: Being‑H0.5 makes a strong case for treating human interaction data as a universal prior for robotic control. The paper introduces a large‑scale Vision‑Language‑Action model trained on 35k+ hours of multimodal human and robot data, paired with a unified action space and a mixture‑of‑experts Transformer architecture. This design enables robust cross‑embodiment generalization, allowing a single policy to transfer across diverse robot morphologies with minimal adaptation, and achieves state‑of‑the‑art results on standard manipulation benchmarks and real‑world platforms.

Like what you read here? Join us to help build the next generation of ultra-efficient AI for devices and robotics! Visit our careers page now!

The AI @ ADI Reading Digest: Recommended papers and research about AI at the edge.

The May 2026 Edition

Tags

Comments