Researchers have released four papers addressing different aspects of improving reasoning capabilities in large language models.
According to arxiv.org, one study introduces Group Relative Policy Optimization (GRPO) for heart-focused medical question answering. The researchers developed a Variance-Aware Reward Framework using rubric-based supervision from RaR-Medicine. On a heart-related subset of HealthBench, their best GRPO variant improved the Qwen3-14B base model’s accuracy from 0.362 to 0.502 and F1 score from 0.532 to 0.668, performing competitively with GPT-OSS-120B (0.508 accuracy, 0.674 F1).
A separate arxiv.org paper examined reducing hallucinations in question answering by using lightweight graph structures to support retrieval-augmented generation (RAG) systems. Testing on the MoNaCo Wikipedia QA benchmark, the researchers report that graph-based tools “can significantly increase the precision and recall of factual correctness, can halve the number of hallucinated answers, and achieves the highest fine-grained truthfulness score.”
In video understanding, arxiv.org describes “pause-and-think-T,” a reasoning-centric training dataset that encourages models to reason over visual evidence before responding. A fine-tuned 4B-parameter model achieved 58.0% accuracy—matching GPT-5.2 on scene understanding while using 59x fewer parameters than Qwen3-VL-235B (58.9%).
Finally, arxiv.org introduces ReLAT (Reconstruction-Guided Latent Reasoning At Test Time), which uses query reconstruction to verify latent reasoning states. On Qwen3-8B, ReLAT raised AIME 2024 accuracy from 56.7% to 73.3%, according to the paper.