Several research teams have introduced novel techniques to make AI reasoning more efficient across different domains, according to recent arXiv papers accepted for presentation at major conferences in 2026.
According to arxiv.org, researchers developed a training-free approach for Table Question-Answering (TQA) featuring two frameworks: TableGrid Navigation (TGN) and Progressive Inference Prompting (PIP). Testing 17 LLMs against 6 baselines, TGN improved over the strongest baseline by 3.8 points on TableBench, while PIP achieved state-of-the-art performance over ReAct and Chain-of-Thought on the FeTaQa dataset. The paper notes this work was “Accepted for Presentation in ICDAR 2026, Vienna, Austria.”
For video reasoning, another arxiv.org paper introduces Internalized Visual Thinking (IVT), a post-training framework that addresses overhead in visual chain-of-thought reasoning. IVT “improves over direct-answer fine-tuning on all six evaluation settings” and “reduces average end-to-end latency by more than 5x” compared to explicit Visual CoT, according to the source.
In the domain of adaptive reasoning, arxiv.org reports on QA-Merging (Query-adaptive Layer Selective Merging), which merges Long-CoT and Short-CoT models without retraining. The method was “Accepted to CIKM 2026,” according to the paper.
Additionally, arxiv.org describes PertMind, which uses cellular perturbation data as reinforcement-learning environments to improve biological reasoning in language models.