news Release
July 30, 2026
At ADI, AI means developing world-class models that overcome latency, power, and cost constraints to deliver real-world insights. While we mixed summer vacations with office visits, we kept our neurons trained on the search for new ideas, concepts and inspiration. This month’s selections reflect that momentum, with audio performance research, device data reasoning, and a more efficient approach to LLM training and explainability. Whether you’re squeezing in a quick read or diving deeper, here’s our list of recommended papers to explore.
📄 SPEAR: A Unified SSL Framework for Learning Speech and Audio Representations
👍🏼 Recommended by: Kao Kitichotkul
💡 Why it’s worth reading: The paper shows how to build a single audio-speech backbone by combining self-supervised learning (SSL) with knowledge distillation from domain-specific speech and audio teachers. The framework uses masked modeling over discrete tokens generated by a multi-codebook vector quantizer. This preserves both semantic speech information and rich acoustic detail. Architecturally, both teachers and student are Transformer-based models with CNN frontends. While the more recent USAD 2.0 model achieves higher overall performance through broader distillation and significantly larger scale, SPEAR offers an alternative knowledge distillation scheme that is applicable to a broader range of architectures, because it does not require the student and teacher models to have similar architectures like in USAD 2.0.. For applications where latency, memory, and power are tighter constraints than absolute benchmark scores, SPEAR's design may represent a more practical starting point than the largest universal audio encoders.
📄 Midtraining Bridges Pretraining and Posttraining Distributions
👍🏼 Recommended by: Debanjan Ghosh
💡 Why it’s worth reading: The paper, authored by Carnegie-Mellon researchers, provides one of the first rigorous explanations for why the increasingly common “midtraining” stage works. Rather than treating midtraining as simply adding higher-quality data, the authors show it acts as a distributional bridge between pretraining and downstream fine-tuning. Through controlled experiments, they find the gains are largest for domains far from web-scale pretraining—especially code and math—where midtraining reduces the gap to the target task, improves adaptation, and mitigates forgetting. Perhaps most interestingly, they identify a model “plasticity window”: introducing specialized data earlier matters more than increasing its volume later. The paper offers practical guidance for designing efficient LLM training pipelines and suggests that carefully staged distribution shifts may be as important as scale itself.
📄 Dynamic Convolutional Neural Networks as Efficient Pre-trained Audio Models
👍🏼 Recommended by: Shanza Iftikhar
💡 Schmid et al. propose Dynamic MobileNets (DyMNs), an efficient CNN architecture for audio tagging that augments MobileNetV3 with three input-adaptive components: Dynamic Convolution (input-conditioned kernel mixing), Dynamic ReLU (adaptive channel-wise activations), and Coordinate Attention (separate temporal and frequency attention). All three are driven by a shared Context Generation Module, keeping the additional compute overhead minimal. Trained on AudioSet using Transformer-to-CNN distillation from a PaSST teacher, DyMNs achieve Transformer-level accuracy while retaining CNN efficiency, outperforming both static MobileNets and several Transformer baselines on AudioSet and downstream benchmarks. The key insight is that lightweight dynamic operations improve CNN scaling more effectively than simply increasing model size.
📄 The Imperfective Paradox in Large Language Models (ACL 2026)
👍🏼 Recommended by: Debanjan Ghosh
💡 Why it’s worth reading: The paper probes a surprisingly fundamental weakness in LLM reasoning: distinguishing between an event being in progress and actually completed. Using a new benchmark, ImperfectiveNLI, the authors show that state-of-the-art models exhibit a strong “teleological bias,” often inferring that goal-oriented actions were completed even when the text explicitly says otherwise (e.g., “was building” → “built”). Perhaps the most interesting finding is that models internally represent the distinction, yet their final predictions are dominated by narrative priors. The paper argues that current LLMs behave more like predictive story-completion engines than faithful logical reasoners, highlighting a limitation that prompting alone is unlikely to fix.
📄 Circuit Tracing: Revealing Computational Graphs in Language Models
👍🏼 Recommended by: Muhammad Aneeq uz Zaman
💡 Why it’s worth reading: This publication from Anthropic is one of the most ambitious attempts yet to make LLM reasoning inspectable rather than merely observable. Instead of analyzing opaque neurons, the authors replace MLP layers with interpretable Cross-Layer Transcoders and construct attribution graphs that trace how individual features interact to produce specific outputs. The key contribution is methodological: a scalable way to uncover circuits for behaviors such as factual recall, acronym generation, and arithmetic while validating that the recovered pathways reflect the original model. For anyone interested in mechanistic interpretability, this paper moves the field from identifying isolated features toward mapping end-to-end computations inside language models.
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!