Several new research papers published to arXiv examine different aspects of large language model reasoning capabilities, revealing both strengths and limitations.
According to a study on planning problems (arXiv:2604.02910), reasoning-enhanced LLMs “significantly outperform traditional satisficing planners” on complex, multi-goal configurations in classic AI planning domains like Blocksworld. The research found that LLMs “track theoretical optimality limits with near-perfect precision,” even when semantic hints are removed. The researchers propose two potential explanations: “active Algorithmic Simulation executed via reasoning tokens” and a “Geometric Memory” that allows models to represent topologies as navigable geometries.
In mathematical proof verification (arXiv:2604.02450), researchers found that smaller open-source models are “only up to ~10% behind frontier models in accuracy but they are up to ~25% more inconsistent.” However, the study demonstrated that smaller models “do possess the mathematical capabilities to verify proofs at the level of frontier models” but struggle to reliably elicit these capabilities. Through specialized prompts, researchers boosted performance by up to 9.1% in accuracy and 15.9% in self-consistency, allowing models like Qwen3.5-35B to perform “on par with frontier models such as Gemini 3.1 Pro.”
A third study (arXiv:2604.02512) examining social meaning found that LLMs “reliably reproduce the qualitative structure of human social inferences but differ substantially in magnitude calibration,” while research on constraint optimization (arXiv:2509.12643) introduced AutoCO, a method coupling constraint relaxation with LLM reasoning for solving complex optimization problems.