Toroidal Logit Bias for Hallucination Reduction in Large Language Models

Superseded (notice added 2026-07-06): the effects reported here did not survive controlled replication — no statistically significant hallucination reduction (McNemar p=0.22); a factorial decomposition attributed the apparent gains to the system prompt, not the toroidal bias. See the replication and null result: DOI 10.5281/zenodo.21206458. Toroidal Logit Bias for Hallucination Reduction in Large Language Models v1.1 - Added TruthfulQA evaluation (817 samples) Key Results: - Custom benchmark (100 prompts): +40% error reduction (Qwen), +15.4% (OLMo) - TruthfulQA (817 prompts): +6.8% error reduction (Qwen) - Paired analysis: 46 improvements vs 32 regressions (McNemar p=0.14) - Consistent directional improvement (b > c) Method: Inference-time toroidal logit bias. No fine-tuning required, ~5% latency overhead. Scope: This work focuses narrowly on an inference-time intervention for hallucination reduction. It makes no claims about ontology, training dynamics, or universal representations. The contribution is operational and empirical. Changelog v1.1: - Added TruthfulQA evaluation (817 samples) with generation-based matching - Added paired McNemar's test analysis - Confirmed directional improvement across both benchmarks