Researchers introduced and released ScienceBuddy, an interactive scientific research workspace whose agents improve continually from researcher requests, feedback, and execution evidence. Its core method, recursive-in-recursive self-improvement, couples harness evolution with model reinforcement learning. The release makes the paradigm available to the scientific community as a step toward discovery intelligence.
A new arXiv paper introduces ScienceIDE, infrastructure that converts scientific code repositories into executable environments for training scientific agents. Agents transform repos into environments supporting task generation, execution, and verification, guided by expert-defined cases and acceptance criteria. The authors train a PhAI-IDE model family (72B, 9B, 4B) on verified trajectories, reporting gains in scientific-code repair and positive transfer to general code, reasoning, and knowledge benchmarks.
A new arXiv paper introduces ScienceIDE, infrastructure that converts scientific code repositories into executable environments for training scientific agents. Agents transform repos into environments supporting task generation, execution, and verification, guided by expert-defined cases and acceptance criteria. The authors train a PhAI-IDE model family (72B, 9B, 4B) on verified trajectories, reporting gains in scientific-code repair and positive transfer to general code, reasoning, and knowledge benchmarks.
KADATH is an open-source evolutionary multi-agent runtime that breeds, evaluates, and improves autonomous agents over reproducible epochs to optimize a goal. It spends large token budgets evolving populations, grading them against benchmarks, and selecting the strongest performers.
KADATH is an open-source evolutionary multi-agent runtime that breeds, evaluates, and improves autonomous agents over reproducible epochs to optimize a goal. It spends large token budgets evolving populations, grading them against benchmarks, and selecting the strongest performers.
A startup called Poetiq, founded by former DeepMind researchers, published a paper on harness-in-the-loop learning, where user-constructed harnesses are optimized for both immediate agent performance and the quality of execution traces used to train future models. A separate open-source project demonstrates that a harness trained on a small Qwen 3.6 35B model improves Terminal-Bench scores for four other models.
A startup called Poetiq, founded by former DeepMind researchers, published a paper on harness-in-the-loop learning, where user-constructed harnesses are optimized for both immediate agent performance and the quality of execution traces used to train future models. A separate open-source project demonstrates that a harness trained on a small Qwen 3.6 35B model improves Terminal-Bench scores for four other models.
An arXiv paper introduces Atria Dawn Preview, a foundation agentic language model for scientific research and engineering workflows. It is trained via a Verifiable Experience Pipeline tying tool-mediated interactions to executable environments and externally verified outcomes. The model is competitive with frontier agents across 16 benchmarks, and the paper analyzes the human-AI collaboration behind its development as a case study.
An arXiv paper introduces Atria Dawn Preview, a foundation agentic language model for scientific research and engineering workflows. It is trained via a Verifiable Experience Pipeline tying tool-mediated interactions to executable environments and externally verified outcomes. The model is competitive with frontier agents across 16 benchmarks, and the paper analyzes the human-AI collaboration behind its development as a case study.
A new arXiv paper introduces AutoResearch, a two-stage autonomous research system that separates Idea Generation from Idea Execution to keep automated research scientifically grounded. It aims to reduce hallucination by connecting emerging research signals with domain knowledge and multi-model generation.
A new arXiv paper introduces AutoResearch, a two-stage autonomous research system that separates Idea Generation from Idea Execution to keep automated research scientifically grounded. It aims to reduce hallucination by connecting emerging research signals with domain knowledge and multi-model generation.