Three research papers published on arXiv and accepted to major 2026 conferences examine different aspects of how large language models handle context and reasoning.
According to a paper accepted at AACL-IJCNLP 2026, researchers propose a knowledge graph-based evaluation framework to assess whether LLMs truly comprehend context or simply excel at pattern matching. The paper introduces Semantic Structural Similarity for KGs (S3KG), which combines structural and semantic signals to evaluate contextual understanding in question answering. According to the research, S3KG achieves F1 gains of up to +7.6 points over the strongest baseline and AUROC up to 0.973 across nine benchmarks.
A separate paper accepted at NeurIPS 2026 addresses “context pollution” in LLM chat systems by introducing “mutable transcripts,” which allow users to revise prior conversation turns through natural language edit requests rather than simply appending new information. According to the study with 17 participants, users significantly preferred mutable transcripts over standard chat across measures of clarity, confidence, and ease of use, with reduced intent to restart conversations.
Meanwhile, a third paper also accepted at NeurIPS 2026 presents PhysElite, a benchmark of 11,586 Olympiad-level physics problems to test multimodal LLM reasoning capabilities. According to the research, even the strongest model tested reached only 33.7% answer accuracy on these expert-level problems.