Reasoning benchmarks often collapse heterogeneous problem types into aggregate scores, obscuring which structural computations large language models (LLMs) can reliably perform. Existing graph-reasoning benchmarks cover graph traversal, broad graph computation, and graph-language interaction, but do not directly test whether graph-structured reasoning transfers across topological mechanisms under matched conditions. We introduce GraphReason-Bench (GR-Bench), comprising 1,800 programmatically generated and oracle-verified problems across three prompt-length regimes and six task families: local motif enumeration, exact-depth breadth-first search, bounded path enumeration, articulation-point identification, non-induced subgraph matching, and non-monotonic rule execution. All tasks share a text-only executable interface, exact oracles, canonical output normalization, and set-based scoring, enabling controlled cross-mechanism comparison. On the 600-instance main evaluation split across nine proprietary and open-source LLMs, we find mechanism-level dissociation: models that succeed on one structural operation often fail on others, and cross-task performance does not reduce to a stable graph-reasoning ability. Trace-level cases reveal mechanism-specific failures such as premature search termination and loss of intermediate state. An initial sparse-autoencoder probe, supported by cross-model robustness checks, shows that topology-sensitive representations become less separable as structural difficulty increases. Together, these results establish graph-structured reasoning as a measurable domain of LLM evaluation.

Lossless compression techniques are crucial in an era of rapidly growing data. Traditional universal compressors like gzip offer low computational overhead, high speed, and broad applicability across data distributions. However, they often lead to worse compression rates than modern neural compressors, which leverage large-scale training data to model data distributions more effectively.Despite their advantages, neural compressors struggle to generalize to unseen data. To address this limitation, we propose a novel framework that performs Test-Time Steering via a Weighted Product of Experts (wPoE).At inference, our method adaptively combines a universal compression model with a pretrained neural language model, ensuring the compression rate is at least as good as the best individual model.Extensive experiments demonstrate that our approach improves the performance of text compression without requiring fine-tuning. Furthermore, it seamlessly integrates with any autoregressive language model, providing a practical solution for enhancing text compression across diverse data distributions.