Logits-to-Logic strengthens and filters last-layer logits to raise LLM logic consistency and reach state-of-the-art on multiple KGQA benchmarks

The paper states that LLMs pre-trained on vast unstructured text can understand logic in natural language and generate logic-consistent responses, yet in structured knowledge reasoning such as Knowledge Graph Question Answering (KGQA), the representational differences between unstructured and structured knowledge make LLMs struggle to maintain logic consistency, producing Logic Drift. It explicitly frames logic inconsistency in structured knowledge reasoning as Logic Drift and attributes it to representational differences between unstructured and structured knowledge rather than to prompt design alone. This claim comes from the paper's abstract-level problem statement and is an argument at the problem-definition level; the abstract provides no quantitative evidence.

Existing methods design complex workflows embedded in prompts to guide LLM reasoning, but the paper argues that these approaches only provide input-level guidance, fail to fundamentally address Logic Drift in LLM outputs, and rely on inflexible reasoning workflows that cannot adapt to different tasks and knowledge graphs. It shifts the locus of improvement from the input prompt level to the model output level, pointing out the limits of input-level guidance and the inflexibility of fixed workflows. This is a comparative positioning statement about prior methods; the abstract names no specific compared methods and reports no experimental data.

The paper proposes the Logits-to-Logic framework, which specifically targets the logits output from the autoregressive generation process and incorporates logits strengthening and logits filtering as core modules to correct logical defects in LLM outputs. Unlike input-level prompt guidance, the framework intervenes directly at the logits level during generation to correct logical defects. The abstract clearly gives the framework name and the functional role of its two core modules, but does not disclose their algorithmic details or implementation.

Experiments show that the approach significantly improves LLMs' logic consistency in structured knowledge reasoning and achieves state-of-the-art performance on multiple KGQA benchmarks. It reports both improved logic consistency and state-of-the-art performance across multiple KGQA benchmarks, supporting the effectiveness of logits-level intervention. The abstract states that extensive experiments were conducted and state-of-the-art performance was achieved, but gives no benchmark names, dataset sizes, comparison baselines, or numerical results.

The framework targets structured knowledge reasoning, especially Knowledge Graph Question Answering (KGQA), in settings that require maintaining logic consistency during autoregressive generation; its intended audience is researchers and practitioners applying LLMs to structured knowledge reasoning. Because the method centers on logits strengthening and logits filtering at the model output layer, a precondition is the ability to access or intervene on the logits of the generation process. The reported multi-benchmark state-of-the-art performance indicates the idea works under KGQA evaluation settings, but the abstract does not specify task types, knowledge graph scale, or deployment conditions, so the practical scope should be judged against the original experimental setup.

The abstract does not disclose how logits strengthening and logits filtering are implemented, which base LLM is used, the names and sizes of the KGQA benchmarks, the comparison baselines, or how logic consistency is measured, so the magnitude and statistical reliability of the gains cannot be judged from the abstract. It mentions extensive experiments and state-of-the-art results but gives no ablation information, leaving the individual contribution of the two core modules unclear. Whether the reliance on logits remains feasible across different models and knowledge graphs is also unaddressed, an open question that requires the full text.