blog
October 30, 2025
Wilmington

At ADI, AI means developing world-class models that overcome latency, power, and cost constraints to deliver insights. It’s October, and where ADI is based, we are all autumn people. That means apple picking, pumpkin spice Dunkin’ lattes, and going over papers members of our team are reading about all things physical intelligence:
LLM-JEPA: LARGE LANGUAGE MODELS MEET JOINT EMBEDDING PREDICTIVE ARCHITECTURES
Recommended by: Yakov Shkolnikov
Why it’s worth reading: This paper by AI researchers Yann LeCun, Hai Huang, and Randall Balestriero introduces a JEPA-style approach to representation learning for large language models. It combines next-token prediction with an embedding-space objective that maps one view (e.g., text) to another (e.g., code). The hybrid loss enables models to retain generative capabilities while learning more abstract and transferable features, yielding consistent performance gains on both text and code tasks. It also supports faster, more robust PEFT fine-tuning (e.g., LoRA) and better resistance to overfitting. However, while the method shows promise, it’s important to contextualize the results: the best-performing non-JEPA 1B-parameter model reaches 62.8% accuracy on GSM8k —above the ~1B parameter baselines used in this study. And performance of larger models exceed 97% accuracy also without using JEPA, albeit at higher inference compute. This suggests caution when interpreting the broader impact of the approach.
Structured Pruning for Deep Convolutional Neural Networks: A survey
Recommended by: Shanza Iftikar
Why it’s worth reading: This survey from the IEEE Transactions on Pattern Analysis and Machine Intelligence systematically reviews methods that remove entire filters, channels, or layers from CNNs to reduce computation while maintaining accuracy. It classifies approaches into importance-based, regularization-based, and search-based pruning, and analyzes their impact on efficiency, generalization, and deployment. It offers a clear, up-to-date roadmap of structured pruning techniques and their trade-offs, bridging algorithmic design with hardware efficiency. The survey is valuable for researchers or engineers aiming to design deployable lightweight models or extend pruning ideas to transformers and AutoML. It highlights open challenges and future research directions, making it an essential reference for anyone working on model compression or efficient deep learning.
Stability-Preserving Automatic Tuning of PID Control with Reinforcement Learning
Recommended by: Solomon Garber
Why it’s worth reading: This paper tackles automatic PID tuning with reinforcement learning while explicitly preserving closed-loop stability, offering a pragmatic route to bring RL into safety-conscious control applications. Published by a team from Texas Tech University’s Chemical Engineering department, it frames the problem as a stability-aware tuning strategy rather than end-to-end control, it aligns with how industrial systems actually deploy PID and clarifies where RL adds value. The article’s positioning and timing make it a helpful reference for researchers blending safe RL with classical control design in practice.
Paper2Agent: Reimagining Research Papers As Interactive and Reliable AI Agents
Recommended by: Parth Brahanpurkar
Why it’s worth reading: Paper2Agent from Stanford University proposes a concrete way to turn static papers into interactive Model Context Protocol (MCP) servers that expose methods as callable tools, resources, and prompts. A multi-agent pipeline automatically mines the repository, sets up environments, extracts tools, and assembles a reliable agent with minimal human effort. Reliability is enforced by validating against the paper’s own figures/results with traceable code links, curbing hallucinated tooling and boosting reproducibility. The authors present case studies (AlphaGenome, TISSUE, Scanpy) that reproduce results and enable new analyses in computational biology, pointing toward interactive, reproducible “paper-as-agent” publishing.
World Modeling with Probabilistic Structure Integration
Recommended by: Yash Kubavat
Why it’s worth reading: This paper—also from Stanford—introduces Probabilistic Structure Integration (PSI): a unified, LLM-style interface for promptable, controllable world models learned from raw video. A randomaccess autoregressive transformer builds a probabilistic graphical model, enabling flexible queries and sampling. PSI zeroshot extracts interpretable structures (flow, depth, segments) via causal, counterfactual probing and then re-injects them as tokenized signals, yielding compositional control surfaces without architectural changes. Trained at 1.4T tokens, the model attains state-of-the-art flow and strong self-supervised depth/segmentation.