AI / ML

OpenLanguageModel (OLM) is an open-source PyTorch library designed for transparent pretraining of small language models. It structures model code to mirror architecture diagrams using explicit wiring modules like Block and Parallel, enabling seamless transitions from educational notebooks to full-scale research ablations. The library integrates tokenizers, streaming datasets, optimization, mixed precision, and hardware-aware execution across CPU and single-node GPU setups.

• Code readability mirrors architecture diagrams, easing debugging and teaching.
• Reusable components allow moving from notebooks to production pretraining unchanged.
• Built-in support for FineWeb-Edu and other streaming datasets simplifies data pipelines.
• Hardware-aware execution covers CPU, single-GPU, and single-node multi-GPU modes.

ColGraphRAG addresses retrieval accuracy issues in graph-grounded multimodal QA by swapping single-vector visual ranking for late-interaction MaxSim scoring. This ColBERT/ColPali-inspired approach preserves patch and token-level structure often lost in bi-encoder similarity, ensuring fine-grained alignment of graph-linked images. The method keeps offline graph construction and text retrieval unchanged while improving retrieval-stage performance on MultimodalQA.

• Late-interaction scoring retains token-level detail for better image-text alignment in GraphRAG
• Visual ranking is upgraded to MaxSim-style multi-vector scoring without changing graph construction
• Retrieval accuracy improves on MultimodalQA compared to standard bi-encoder approaches

Shapley Context Pruning introduces a framework that models RAG context retrieval as a cooperative game to determine token importance. It replaces heuristic loss functions with Shapley values for interpretable attribution. The method utilizes a Deep Sets architecture to balance fine-grained and coarse-grained context representations effectively.

• Replaces heuristic losses with cooperative game theory for context importance
• Uses Shapley values to provide interpretable attribution for retrieved tokens
• Employs Deep Sets architecture to balance representation granularity
• Offers a unified framework for context reranking and pruning

Researchers propose Generative Ontology Induction (GOI), a domain-agnostic framework that extracts structured ontologies from document corpora using large language models. The system produces a generative blueprint of entities, relationships, and constraints, exporting them as typed graphs in YAML or JSON. To evaluate structural completeness, the authors introduce the Node Coverage Score, which measures how many ontology classes and properties appear in the generated outputs.

• GOI automates schema discovery without requiring predefined schemas or narrow domain constraints.
• Outputs are structured as typed graphs (6 node types, 7 edge types) in YAML/JSON for easy integration.
• New Node Coverage Score metric evaluates the fraction of structural ontology nodes in generated outputs.
• Addresses the bottleneck of manual ontology engineering for knowledge-intensive AI systems.

PPO-HSC is a reinforcement learning framework that addresses mode collapse in LLM fine-tuning by optimizing wide-area policy coverage. It introduces a High-order Sampling Coverage reward that encourages the discovery of unique, high-validity reasoning patterns rather than over-optimizing known solutions. The method maintains a dynamic library of verified unique trajectories to sustain exploration during training.

• Prevents mode collapse by rewarding low-similarity, high-validity reasoning paths
• Uses dynamic trajectory libraries to track and incentivize unique solutions
• Shifts RLVR focus from pure reward maximization to broader exploration
• Aims to preserve model curiosity and solution manifold diversity
• Targets LLM fine-tuning stability in complex reasoning tasks