Three papers published on arxiv.org in July 2026 present advances in understanding and improving AI model capabilities across different dimensions.
According to arxiv.org, researchers introduced M-JudgeBench, a ten-dimensional capability-oriented benchmark designed to assess Multimodal Large Language Models (MLLMs) used as judges for evaluation tasks. The benchmark decomposes evaluation into pairwise Chain-of-Thought comparison, length bias avoidance, and process error detection tasks. The researchers also proposed Judge-MCTS, a data construction framework that generates pairwise reasoning trajectories, which they used to train M-Judger, a series of judge models that demonstrated superiority on existing benchmarks.
In a separate study, arxiv.org reported on research auditing open-weight language models using persona vectors—behavioral directions in activation space. The work compiled a 53-trait inventory across four behavioral domains, labeling traits as “natural” (expressed at baseline), “steerable” (latent but amplifiable), or “intractable” (resistant to extraction). According to the source, both models studied defaulted to helpful behavior, with all nine agentic traits being natural.
A third paper on arxiv.org, accepted at ICLR 2026, investigated representation-based exploration for language models. According to the research, exploration with a representation-based diversity bonus derived from pre-trained model hidden states “significantly improves diversity and pass@k rates” for both post-training and inference-time scaling. For example, the approach achieved over 50% improvement in verifier efficiency for Qwen-2.5-14b-Instruct across various tasks.