Yunkai Zhang
Publications
When 2D Tasks Meet 1D Serialization: On Serialization Friction in Structured Tasks
Large language models (LLMs) conventionally process structured inputs as 1D token sequences. While natural for prose, such linearization may introduce additional representational burden for tasks whose computation depends directly on explicit 2D structure, because row--column alignment and local neighborhoods are no longer directly expressed in the input. We study this setting, which we refer to as serialization friction, on a small diagnostic testbed of synthetic tasks with explicit 2D structure: matrix transpose, Conway's Game of Life, and LU decomposition. To examine this question, we compare a text-only language pathway over serialized inputs with a vision-augmented pathway, built on the same language backbone, that receives the same underlying content rendered in task-faithful 2D layout, yielding a system-level comparison between two end-to-end input pathways. Across the tasks and settings we study, the visual pathway consistently outperforms the textual pathway; the gap often widens at larger dimensions, and error patterns under serialization become increasingly spatially structured. These findings indicate that the relationship between input representation and model performance on such tasks warrants further investigation, and suggest that preserving task-relevant 2D layout is a promising direction for structured 2D tasks.
TARo: Token-level Adaptive Routing for LLM Test-time Alignment
Large language models (LLMs) exhibit strong reasoning capabilities but typically require expensive post-training to reach high performance. Recent test-time alignment methods offer a lightweight alternative, but have been explored mainly for preference alignment rather than reasoning. To bridge this gap, we propose, Token-level Adaptive Routing (TARo), which steers frozen LLMs toward structured reasoning entirely at inference time. Specifically, we first train reward models on step-wise mathematical traces to capture fine-grained logical consistency signals, then introduce a learnable token-level router that automatically controls the guidance of the reward model to the base model. Extensive experiments show that TARo significantly improves reasoning performance by up to +22.4% over base model and +8.4% over existing token-level test-time alignment methods, while also boosting out-of-distribution clinical reasoning (MedXpertQA) and instruction following (AlpacaEval). Furthermore, TARo also generalizes from small to large backbones without retraining, extending test-time alignment from preference optimization to robust, cross-domain reasoning.