A talk introduces a computational framework for learning a compact structured representation for real-world datasets, that is both discriminative and generative. It proposes to learn a closed-loop transcription between the distribution of a high-dimensional multi-class dataset and an arrangement of multiple independent subspaces, known as a linear discriminative representation (LDR). The optimality of the closed-loop transcription can be characterized in closed-form by an information-theoretic measure known as the rate reduction. Why it matters: The framework unifies concepts and benefits of auto-encoding and GAN and generalizes them to the settings of learning a both discriminative and generative representation for multi-class visual data.
This article discusses retrieval augmentation in text generation, where information retrieved from an external source is used to condition predictions. It references recent work on retrieval-augmented image captioning, showing that model size can be greatly reduced when training data is available through retrieval. The author intends to continue this work focusing on the intersection of retrieval augmentation and in-context learning, and controllable image captioning for language learning materials. Why it matters: This research direction has the potential to improve transfer learning in vision-language models, which could be especially relevant for downstream applications in Arabic NLP and multimodal tasks.
The paper introduces Sparse-Quantized Representation (SpQR), a new compression format and quantization technique for large language models (LLMs). SpQR identifies outlier weights and stores them in higher precision while compressing the remaining weights to 3-4 bits. The method achieves less than 1% accuracy loss in perplexity for LLaMA and Falcon LLMs and enables a 33B parameter LLM to run on a single 24GB consumer GPU. Why it matters: This enables near-lossless compression of LLMs, making powerful models accessible on resource-constrained devices and accelerating inference without significant accuracy degradation.
A new method is proposed to reduce the verbosity of LLMs in step-by-step reasoning by retaining moderately easy problems during Reinforcement Learning with Verifiable Rewards (RLVR) training. This approach acts as an implicit length regularizer, preventing the model from excessively increasing output length on harder problems. Experiments using Qwen3-4B-Thinking-2507 show the model achieves baseline accuracy with nearly twice shorter solutions.
The paper introduces Yet another Policy Optimization (YaPO), a reference-free method for learning sparse steering vectors in the latent space of a Sparse Autoencoder (SAE) to steer LLMs. By optimizing sparse codes, YaPO produces disentangled, interpretable, and efficient steering directions. Experiments show YaPO converges faster, achieves stronger performance, exhibits improved training stability and preserves general knowledge compared to dense steering baselines.