Latest Advances in AI Agents: PaperBanana, SDPO, Lumine, Idea2Story, and Insight Agents

Research on artificial intelligence is transitioning from conversational large models to AI agents that can plan, execute, and collaborate, shifting the focus from single‑model performance to multi‑agent architectures that produce verifiable, reusable results in real‑world settings.

1. PaperBanana: Automating Academic Illustration for AI Scientists

Researchers from Peking University and Google Cloud AI propose PaperBanana, a VLM‑driven agent framework that automatically retrieves, plans, stylizes, and iteratively optimizes publication‑grade academic figures. In evaluations on a NeurIPS‑2025 benchmark of diverse, aesthetically complex charts, PaperBanana outperforms baseline methods in fidelity, simplicity, readability, and visual appeal.

2. Reinforcement Learning via Self‑Distillation (SDPO)

The paper introduces Self‑Distillation Policy Optimization, a strategy‑optimization method that converts token‑level feedback into dense learning signals without external teacher or explicit reward models. The current model’s output under given feedback serves as a self‑teacher; its next‑token predictions are distilled back into the policy, leveraging the model’s ability to recognize its own errors in context. On LiveCodeBench v6, SDPO achieves higher sample efficiency and final accuracy than strong RLVR baselines across scientific reasoning, tool use, and competitive programming tasks.

3. Lumine: An Open Recipe for Building Generalist Agents in 3D Open Worlds

Lumine is the first open‑source generalist‑agent solution capable of executing hour‑long complex tasks in rich 3D environments. It adopts a human‑like interaction paradigm, using a vision‑language model to unify perception, reasoning, and action end‑to‑end. The agent processes raw pixel input at 5 fps and generates precise keyboard‑mouse actions at 30 fps, invoking the reasoning module only when necessary. Experiments demonstrate robust adaptation across varied world settings, marking a significant step toward universal agents in open environments.

4. Idea2Story: An Automated Pipeline for Transforming Research Concepts into Complete Scientific Narratives

AgentAlpha presents Idea2Story, a pre‑computed framework that builds a methodological knowledge graph from peer‑reviewed papers, converting vague research ideas into structured, reusable patterns. This reduces large‑language‑model context limits and hallucinations, enabling efficient, novel scientific discovery without runtime reprocessing of literature. The accompanying dataset trains Idea2Story to learn expression and evaluation of research contributions, supporting retrieval and composition of reusable methodological patterns.

5. Insight Agents: An LLM‑Based Multi‑Agent System for Data Insights

Amazon researchers introduce Insight Agents, a multi‑agent system built on large language models with a planning‑execution architecture, hierarchical agents, and out‑of‑distribution (OOD) routing. The system delivers accurate business insights to Amazon sellers within 15 seconds, achieving 90 % human‑evaluated accuracy. Evaluation uses a curated dataset of 301 questions (178 in‑domain, 123 OOD) plus a benchmark of 100 popular queries with ground‑truth answers for end‑to‑end assessment.

The five papers collectively illustrate progress in agent framework design, cross‑modal coordination, self‑feedback learning, and closed‑loop task execution, offering a clear view of the evolution path toward next‑generation generalist AI agents. Additional cutting‑edge AI papers are updated daily on the HyperAI "Latest Papers" page.