Four research papers published on arXiv introduce novel approaches to improving AI model capabilities and efficiency.
Brainstacks presents a modular architecture for continual multi-domain fine-tuning of large language models. According to arxiv.org, the system uses frozen adapter stacks with MoE-LoRA routing and achieves “2.5x faster convergence than parameter-matched single LoRA.” The research, validated on TinyLlama-1.1B and Gemma 3 12B IT, found that “domain stacks encode transferable cognitive primitives (instruction-following clarity, numerical reasoning, procedural logic, chain-of-thought structure) rather than domain-specific knowledge,” with medical prompts routing to chat and math stacks in 97% of cases despite those stacks containing zero medical data.
Science-T2I addresses scientific accuracy in image generation. According to arxiv.org, researchers introduced an expert-annotated dataset with over 20,000 adversarial image pairs across 16 scientific domains. The evaluation of 18 image generation models found that “none scores above 50 out of 100 under implicit scientific prompts,” though explicit prompts yielded scores “roughly 35 points higher.” The paper was accepted to CVPR 2025.
PixelPrune tackles visual token reduction in Vision-Language Models. According to arxiv.org, the training-free method exploits pixel-level redundancy, noting that “only 22-71% of image patches are pixel-unique” in document and GUI benchmarks. The approach delivers “up to 4.2× inference speedup and 1.9× training acceleration.”
Online Reasoning Calibration (ORCA) improves test-time efficiency. According to arxiv.org, the framework achieved “savings up to 47.5% with supervised labels” on in-distribution tasks using Qwen2.5-32B, and improved zero-shot out-of-domain MATH-500 savings “from 24.8% of the static calibration baseline to 67.0%.”