Sage‑A1.5 Shifts from Long‑Horizon Execution to Scientific Exploration Using Capability Composition
agents training
| Source: HF Papers | Original article
Sage‑A1.5, a 35B Mixture‑of‑Experts agentic model, extends A1 by using adaptive capability composition to tackle long‑horizon problems, enabling it to organize and revise its work across tasks.
Shanghai AI Lab has unveiled Sage‑A1.5, a 35‑billion‑parameter Mixture‑of‑Experts (MoE) foundation model purpose‑built for agentic tasks. The technical report, posted on arXiv (2610.28578), describes the model as the next step in the Agents‑A1 lineage, shifting the focus from sheer parameter count to extending the execution horizon of autonomous agents.
Sage‑A1.5 introduces “adaptive capability composition,” a second‑order development that lets the system not only solve individual tasks but also organise work, coordinate multiple problem‑solving abilities and revise its approach on the fly. The paper highlights three concrete upgrades: enhanced agentic orchestration, built‑in verification mechanisms and self‑check routines that together aim to make long‑horizon execution more reliable.
Early benchmarks underscore the impact. On a condensed‑matter physics suite, Sage‑A1.5 achieved a score of 80.0, markedly higher than the 50.0 recorded by the contemporary GPT‑5.6‑sol baseline. The model operates with roughly 3 billion active parameters at inference, leveraging the MoE architecture to keep compute demands modest while scaling capability breadth across domains.
The release matters because it pushes the frontier of AI‑driven scientific exploration. By embedding orchestration and verification directly into the model, researchers can delegate more complex, multi‑step investigations to an autonomous system, potentially accelerating discovery cycles in fields that require sustained, iterative reasoning.
Going forward, the community will watch for broader evaluations on diverse scientific benchmarks, integration into open‑source toolchains, and any follow‑up releases that expand the capability‑composition paradigm. As we have seen with recent work on long‑horizon environments and object‑oriented context, Sage‑A1.5 could become a cornerstone for the next generation of autonomous research agents.
Sources
Back to AIPULSEN