Tag
The article highlights a counterintuitive finding: adding a weak model to a voting panel can degrade performance by adding noise, whereas a single independent uncorrelated model (e.g., a 32B) can outperform multiple same-vendor models. It emphasizes the value of uncorrelated voters over mere quantity.
DLLG (Dynamic Logit-Level Gating) is a novel framework that dynamically fuses multiple specialized LLMs at the token-level logit space using a lightweight learned gating module, outperforming routing, heuristic ensembling, and parameter-merging baselines across reasoning and code benchmarks. The approach requires only sparse response-level supervision and preserves expert modularity without retraining.
SwanNLP presents an LLM-based framework for plausibility scoring in narrative word sense disambiguation at SemEval-2026 Task 5, using structured reasoning and dynamic few-shot prompting to predict human-perceived plausibility of word senses in short stories. The work demonstrates that commercial large-parameter LLMs with few-shot prompting and model ensembling effectively replicate human judgment patterns in realistic narrative contexts.