news Release
September 28, 2026

At ADI, AI means developing world-class models that overcome latency, power, and cost constraints to deliver real-world insights. Summer may be winding down, but the pace of AI research certainly is not. After taking August off for a well-earned vacation break, the AI @ ADI Reading Digest returns with a fresh selection of papers spanning time-series reasoning, reinforcement learning, knowledge distillation, and efficient speech models. Whether you are easing back into your reading list or ready for a deeper technical dive, September offers plenty of new ideas to explore.
📄 Adaptive Time Series Reasoning via Segment Selection
👍🏼 Recommended by: Erkan Bayram
💡 Why it’s worth reading: Adaptive Time Series Reasoning via Segment Selection reframes time-series reasoning as active evidence acquisition rather than static full-sequence processing. ARTIST alternates between a controller that selects temporal windows and decides when to stop, and a reasoner that updates its answer from accumulated evidence. Training combines LoRA-based supervised fine-tuning with collaborative self-play reinforcement learning and hierarchical policy optimization. Across six benchmarks, it reports roughly six percentage points of average improvement, with larger gains on tasks where the critical portion of the signal is brief and concentrated, though some datasets still require most of the series.
📄 Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
👍🏼 Recommended by: Solomon Garber
💡 Why it’s worth reading: This paper challenges the idea that reinforcement learning with verifiable rewards (RLVR_ teaches language models to reason in fundamentally new ways. The researchers used the pass@k metric to assess performance. The metric tracks the percentage of problems a model solves correctly with k generated attempts. Although RLVR improves performance by making known solution paths easier to sample, large-k evaluation reveals that base models can solve some problems their RL-trained counterparts cannot. The results suggest that current RLVR primarily reshapes the base model’s existing output distribution rather than expanding its reasoning frontier, sometimes trading broader solution coverage for better average-case performance.
📄 Agree to Disagree: Adaptive Ensemble Knowledge Distillation in Gradient Space
👍🏼 Recommended by: Alessandro Ragano
💡 Why it’s worth reading: This paper improves ensemble knowledge distillation by recognizing that multiple teacher models may give conflicting guidance to a student model. Instead of simply averaging all teachers equally, it estimates which teachers provide compatible and useful learning signals during training and adaptively balances their influence. This allows the student to benefit from diverse teachers while reducing the impact of noisy or contradictory ones. The main takeaway is that ensemble distillation works better when teacher disagreement is handled explicitly rather than hidden inside a simple average.
👍🏼 Recommended by: Shanza Iftikar
💡 Why it’s worth reading: LongCat-Audio-Codec is a neural audio tokenizer/detokenizer designed specifically for end-to-end Speech LLMs. Its main idea is to separate semantic information from acoustic information. The model consists of:
The model uses a multi-stage training strategy: first training the semantic encoder for strong speech understanding, then training the acoustic encoder for robustness and intelligibility, followed by decoder training for high-quality reconstruction and optional speaker-specific fine-tuning.
📄 Reinforcement-aware Knowledge Distillation for LLM Reasoning
👍🏼 Recommended by: Erkan Bayram
💡 Why it’s worth reading: The paper introduces RL-Aware Distillation, a method for transferring reasoning capabilities from a large teacher model to a smaller student during reinforcement learning. Instead of continuously forcing the student to imitate the teacher through a Kullback-Leibler (KL) penalty, the method uses reward signals to determine when imitation is beneficial. The authors propose a Trust Region Ratio Distillation (TRRD) objective. TRRD combines two techniques:
The benefit: improved training stability and performance on mathematical and logical reasoning benchmarks.
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!