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

👍🏼 Recommended by: Muhammad Aneeq uz Zaman

💡 Why it’s worth reading:  ChipMind proposes a method to enable LLMs to reason over long, complex chip specifications documents. The method converts documents into a structured knowledge graph (ChipKG) that captures circuit semantics (signals, conditions, dependencies). It then uses an iterative retrieval + reasoning loop with adaptive evidence selection to answer multi-hop questions more accurately than standard RAG. The key insight is that structure-aware retrieval and step-by-step reasoning outperform naive long-context or vector-based approaches for technical documents. 

📄 Constrained Decoding with Speculative Lookaheads

 👍🏼 Recommended by: Debanjan Ghosh

💡 Why it’s worth reading:  This paper tackles a practical bottleneck in aligning large language models: constrained decoding with lookahead heuristics delivers strong constraint satisfaction but is far too slow for real-world use. The authors introduce Constrained Decoding with Speculative Lookaheads (CDSL), borrowing ideas from speculative decoding to offload expensive lookahead generation to a smaller draft model while verifying with the target model and task-specific rewards. The result is a compelling 2×–12× speedup with minimal quality loss, making high-quality constrained decoding far more deployable in production settings.

📄 Steering Your Diffusion Policy with Latent Space Reinforcement Learning

👍🏼 Recommended by: Marlon Ares Milian

💡 Why it’s worth reading:  

This 2025 preprint presents a modular, practical, and sample-efficient methodology to improve the performance of diffusion-based policies, particularly for the application of robotic behavioural cloning (BC). The authors introduce Diffusion Steering via Reinforcement Learning (DSRL), which develops a reinforcement learning (RL) policy to modify the latent noise space of a pretrained diffusion policy. Instead of finetuning weights, DSRL learns how to bias the initial noise distribution of the diffusion policy to indirectly modify the action space; this results in strong sample efficiency and black-box compatibility. The authors show performance improvements from 20% (diffusion policy before steering) to around 90% (RL-steered diffusion policy) success rate for a variety of tasks including: offline, online, and offline-to-online single task learning, as well as steering of generalist (multi-task) diffusion policies, both in simulation and real-world benchmarks. This work provides precise practical guidelines to adjust the methodology to different scenarios, as well as a modular, readily available, codebase. Overall, this paper bridges BC and RL elegantly, and provides a practical, readily available, methodology for developing diffusion-based policies for robot manipulation.

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!