According to arxiv.org, researchers have introduced CANON (Consensus-ANchored self-distillatiON), a label-free training method that converts consensus signals from multiple model solutions into token-level supervision for large language models. The method samples multiple solutions for unlabeled prompts, extracts the majority answer, and uses a frozen model snapshot conditioned on a consensus-reaching solution to supervise training. Experiments on mathematical and scientific reasoning benchmarks showed CANON improved pass@1 accuracy by up to 12 points, outperforming label-free reinforcement learning by 6 points while using one-seventh of its compute, according to the paper.
Separately, arxiv.org published research on AgentHOI, a training-free framework for human-object interaction detection that uses multimodal large language models. The framework employs “Context-aware Multi-round Reasoning” and “Multifaceted Interaction Localization” mechanisms to progressively refine interaction hypotheses and enhance grounding precision. According to the researchers, AgentHOI achieved superior performance over state-of-the-art supervised methods despite requiring no HOID training data.
In medical AI, arxiv.org introduced ELF, a family of three encoder-free ECG-Language Models that simplify ECG interpretation. According to the paper, accepted to MLHC 2026, ELF models remain competitive with and often outperform prior state-of-the-art models across two datasets while using substantially simpler architectures and training pipelines.