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

This 2020 Interspeech paper addresses the challenge of detecting and counting overlapping speakers in distant‑microphone audio. It reframes related tasks—Voice Activity Detection (VAD), Overlapped Speech Detection (OSD), joint VAD+OSD, and speaker counting under a unified problem definition called Overlapped Speech Detection and Counting (OSDC). The paper’s central contribution is a Temporal Convolutional Network (TCN)–based approach designed to handle the temporal structure and complexity of multi‑speaker acoustic scenes. TCNs are well‑suited for this domain because they process time‑ordered data using causal convolutions, ensuring predictions rely only on past inputs, thereby avoiding leakage of future information. The use of dilated convolutions efficiently captures long‑range temporal dependencies without excessive depth or parameter count. Residual connections stabilize training and enable deeper architectures.

The system operates on 80‑dimensional log‑mel filterbank features extracted from short‑time windows (25 ms, 10 ms stride). The features are normalized and passed through an initial convolutional layer followed by a stack of residual TCN blocks (three groups of five). Forced‑alignment‑based labeling is implemented using Hidden Markov Models or deep learning models. This provides precise and consistent time‑aligned supervisory targets. Together, these components form a coherent architecture capable of identifying when speech is present, determining whether multiple talkers overlap, and estimating the number of active speakers. This unified TCN‑driven formulation yields an efficient, scalable solution for overlapping‑speech scenarios common in far‑field and real‑world audio environments.

📄 Less is More: Recursive Reasoning with Tiny Networks

👍🏼 Recommended by: Kao Kitichotkul

💡 Why it’s worth reading:  Less is More: Recursive Reasoning with Tiny Networks introduces the Tiny Recursive Model (TRM), a remarkably small yet highly capable architecture designed to perform recursive reasoning with far fewer parameters than prior hierarchical models. TRM simplifies the more complex HRM by collapsing reasoning into a single tiny 2‑layer network that is repeatedly applied through a set of nested refinement loops, aligning with the triply‑nested structure highlighted in your presentation. This recursive structure allows the model to iteratively refine a latent state z and an output y, effectively deepening the computation without adding depth to the network itself. The authors demonstrate that this lightweight recursive process significantly improves generalization, outperforming HRM across a diverse set of reasoning tasks despite having dramatically fewer parameters. The paper positions TRM as an alternative to Deep Equilibrium Models,—both approaches emulate deep computation without expanding model size, but TRM does so with simpler training dynamics and without requiring fixed‑point solvers. Overall, TRM shows that intelligently structured recursion can substitute for scale, offering a high‑capacity reasoning mechanism that is compute‑efficient and attractive for edge‑constrained environments—an insight particularly aligned with the Edge AI perspective.

 📄 Knowledge Graphs are Implicit Reward Models: Path-Derived Signals Enable Compositional Reasoning

👍🏼 Recommended by: Debanjan Ghosh

💡 Why it’s worth reading:  This Princeton paper preprint introduces a post‑training framework in which knowledge graphs (KGs) function as implicit reward models. The framework enables reinforcement‑learning (RL) pipelines to supervise large language models using grounded, verifiable, and scalable signals derived directly from KG paths rather than final‑answer correctness alone. Prior RL‑for‑reasoning approaches optimize only end outcomes, whereas the paper’s method encourages compositional reasoning by training models on shorter multi‑hop KG paths that generalize to harder long‑hop queries. This results in a particularly compelling result in the Unified Medical Language System (UMLS) domain, where structured disease–symptom–drug relations lend themselves to path‑aligned reasoning. The dataset construction includes 24,660 KG‑grounded QA tasks and a non‑overlapping evaluation benchmark (ICD‑Bench). Evaluations revealed that smaller models fine‑tuned with path‑derived signals can outperform larger frontier models on higher‑hop reasoning tasks.
The paper’s reward modeling integrates four signals: binary correctness, expert‑trace similarity, thinking‑quality scoring, and the key innovation—path alignment, which scores a model’s reasoning based on how closely its intermediate steps cover ground‑truth KG triples. KGs also provide process‑level supervision that mitigates hallucinations, enforces factual grounding, and improves out‑of‑distribution generalization in multi‑step reasoning tasks. By using structured domain knowledge as a foundation layer for rewards, the framework offers a scalable alternative to costly human‑generated reasoning traces and positions KG‑guided RL as a powerful path toward more trustworthy and compositional LLM reasoning.

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!