Researchers have published multiple studies advancing AI reasoning and multimodal capabilities, according to papers posted on arXiv on June 3, 2026.
According to arxiv.org, CP-Agent introduces an “agentic multimodal large language model (MLLM)” designed for cellular morphological profiling under drug perturbations. The system uses a “context-aware alignment module, CP-CLIP, that jointly embeds high-content images and experimental metadata,” achieving a maximum F1-score of 0.896 for treatment and mechanism-of-action discrimination. The paper states CP-Agent can generate “mechanism-relevant, human-interpretable rationales for cell morphological changes under drug perturbations” to support drug discovery workflows.
In a separate study on arxiv.org, researchers propose Vision Inference Former (VIF), addressing limitations in multimodal language models where “visual information serves as the core evidential modality” but is “treated on par with textual tokens.” According to the paper, VIF “continuously injects visual semantics throughout the decoding phase” and was tested on 14 benchmark tasks covering general reasoning, OCR, and table understanding.
A third paper on arxiv.org reports a negative result on cross-model activation transfer between Pythia-160M and Pythia-410M models. According to the researchers, while a linear translation layer achieved “normalized cosine similarity near 0.97,” injecting translated activations at inference time “did not improve downstream answering.”
Finally, arxiv.org describes Ctx2Skill, a “self-evolving framework” for context learning that “autonomously discovers, refines, and selects context-specific skills without human supervision or external feedback,” using a multi-agent self-play approach.