Sage-A1.5: From Long-Horizon Execution to Scientific Exploration with Capability Composition

AI summary

Overview

Research area: Agentic AI / large language model post-training — specifically long-horizon agentic execution, multi-agent orchestration, self-verification, and AI-assisted scientific research.

Technical level: Advanced. The report assumes familiarity with Mixture-of-Experts (MoE) architectures, supervised fine-tuning, PPO-style reinforcement learning, agent harnesses, and benchmark suites.

Scope (one sentence): This is a technical report describing Sage-A1.5, a 35B MoE agentic model from the Shanghai Artificial Intelligence Laboratory that adds learned task delegation across parallel subagents and learned self-verification to a long-horizon agent, and applies it to coding, machine-learning engineering, long-context reasoning, web research, real-world professional workflows, and scientific discovery.

What This Paper Is About

Earlier agent models such as Agents-A1 improved long-horizon performance by scaling the agent horizon — chaining tool calls, observations, and verification feedback over trajectories averaging 45K tokens — rather than scaling parameters. The authors argue that stronger domain skills alone are insufficient: complex work also requires organizing subtasks, running investigations in parallel, comparing approaches, and revising decisions when evidence changes, and a single agent working step by step suffers from growing context, slow exploration, and compounding errors. Sage-A1.5's goal is to make how work is organized an explicit training target, so that delegation and verification become learned model capabilities rather than fixed rules in an external harness.

Key Contributions

1. Adaptive Agentic Execution. The authors integrate Agentic Orchestration (decomposing work, delegating subtasks to parallel subagents with isolated contexts, and integrating their findings) and Agentic Verification (recognizing stalled progress or unverified conclusions and revising the plan) into the model's execution policy for long-horizon problem solving.
2. A unified training pipeline. Multi-domain supervised fine-tuning is combined with mixed-reward reinforcement learning, fusing capabilities across domains, reasoning modes, and agent harnesses into a single agentic model, via a three-stage pipeline that first builds multi-domain teachers and then applies unified outcome–process post-training.
3. Domain-specific data and environment construction. The report details pipelines for Kaggle-style ML engineering (Kaggle-MLE Factory), repository-level software engineering (SWE-rebench-V2 plus in-house tasks, with leakage controls), scientific research (a three-stage science teacher pipeline with search/visit/code/scholar tools), long-context reasoning (recursive task evolution on LongBlocks documents from 16K–128K tokens across 24 task categories), open-ended web research, and economically valuable real-world workflows synthesized from Agents' Last Exam seeds.
4. Demonstration of agentic scientific discovery. Benchmarks and multidisciplinary case studies are used to argue for a shift from isolated task solving toward long-horizon, collaborative, verifiable, and inspectable scientific exploration.

Main Findings

• Small model, stronger benchmark claims: Despite having only 35B parameters, Sage-A1.5 is reported to outperform the substantially larger Kimi-K3, GLM-5.3, and GPT-5.6-sol on FrontierScience-Olympiad, CMT, PinchBench, and IFBench, and to remain competitive with them on SWE-Bench Pro and BrowseComp.
• Consistent gains among similarly sized models: Sage-A1.5 matches or surpasses the equally sized Agents-A1 and Ornith-1.5 on all twelve benchmarks the authors evaluate.
• Two distinct composition patterns: Parallel agent execution is characterized as exploratory composition (multiple agents test alternative hypotheses for the same question, analogous to A ∨ B ∨ C) and constructive composition (complementary subproblems assembled into one solution, analogous to A ∧ B ∧ C); complex tasks interleave both.
• Verification can be a learned, on-demand behavior: A conditioning marker, Verification protocol: enabled, added to the system prompt activates Agentic Verification, after which the model performs verification inside its own reasoning and tool-use process — with no separate verification tool, no online supervisor, and no harness-implemented procedure.
• Structured subagent feedback aids the main agent: Subagents return their result together with a self-assessment, supporting evidence, and remaining issues, which helps the main agent detect gaps and conflicts and also supplies signals for trajectory selection.
• Long-context task difficulty can be raised deliberately: Recursive evolution over ten difficulty axes produced 3,512 hard QA pairs, on which Qwen3.5-35B-A3B achieves a mean per-question success rate of 31.57%; after answer filtering and joint LLM-judge selection over three sampled trajectories per question from DeepSeek-V4-Pro, 1949 LongBlocks trajectories were retained for SFT.
• Real-world task supply is the bottleneck: For Agents' Last Exam-style tasks (55 subdomains, 13 industry clusters, deterministic normalized [0,1] grading with no LLM judge), a manual cold start of a few dozen tasks was evolved through controlled variation into a pool of more than one thousand professional workflow environments.
• Quantitative per-benchmark scores are not reported in the text: The paper's numeric results appear in Figure 1 and in the case studies, which are not included in the provided content; only the relative comparisons above are stated.

Methodology in Plain English

The work starts from an existing 35B Mixture-of-Experts agent model, Agents-A1, and builds two new behaviors into it.

Delegation. The model is trained to decide when to investigate a problem itself and when to hand pieces of it to helper agents that work in separate contexts at the same time. To create training data, the authors build hard search questions by picking a target answer, tracing several different chains of related clues outward from it (forming a star-shaped graph), then merging those chains into a single question with identifying names masked. Because different clues constrain the same hidden answer, helper agents can pursue different clues in parallel and the main agent can combine their findings. A second task family, multi-source research synthesis, trains division of a broad research goal into complementary assignments plus integration into a report. The main agent's trajectory and each helper's trajectory are treated as separate training samples with role-specific prompts, tools, and budgets.

Verification. A supervisor model watches a researching model work through live search and page reading and inserts a verification step whenever the search becomes repetitive, progress stalls, or a key conclusion needs confirmation. Each verification uses only the evidence available at that moment, separates confirmed facts from unverified hypotheses, and proposes a next direction. The supervisor prompts are then stripped out and the reflection text is folded into the model's own reasoning, so the behavior is internalized; the supervisor is only a data-construction device.

Training. The pipeline is three-stage: build domain teachers on coding, long-context reasoning, deep research, scientific reasoning, and co-working data using environment construction, trajectory selection, staged SFT, and RL; then apply unified outcome–process post-training to produce one model. For ML engineering, a two-stage SFT is followed by Windowed Interaction Reinforcement Learning, which groups interactions between successive experiment results into windows with shared outcome-based rewards, trained asynchronously with token-probability-ratio correction weights and PPO clipping, KL regularization, and entropy terms. For software engineering, tasks are built from SWE-rebench-V2 and in-house sources, validated with reference-patch and no-op checks, and rolled out in isolated containers with leakage controls (Git objects removed, proxy redirected to a closed port, audits of trajectories). For science, a teacher is distilled using four tools (search, visit, code, scholar) inside a sandbox with tool-result replay, then filtered on solution correctness, reasoning depth, and tool-call success rate before token-level SFT on Qwen3.5-35B-A3B with an extended multi-epoch schedule.

Why This Matters

Impact on research. The report reframes agent progress along two axes — capability coverage (which domains a model can handle) and capability composition (how those capabilities are organized and revised during execution) — and shows that the second can be trained rather than engineered into a scaffold. Making verification an intrinsic, on-demand behavior rather than an external harness procedure is a concrete alternative to fixed agent frameworks.

Real-world applications:

• Scientific research assistance: literature retrieval, computational analysis, and evidence-grounded synthesis across mathematics, physics, chemistry, and biology, producing inspectable research artifacts.
• Machine-learning engineering: automated competition-style experimentation with branch comparison, backtracking, and submission management.
• Software maintenance: resolving repository-level issues from natural-language descriptions into validated patches.
• Professional workflow automation: multi-step file and data deliverables in specialized software across 13 industry clusters, graded deterministically.
• Private, local deployment: the report notes local deployment for data privacy, access control, and offline support.

Industry relevance. A 35B MoE model that is competitive with or better than much larger systems on several agentic benchmarks is directly relevant to cost-sensitive deployment, and the report explicitly frames the work as an accessible foundation for open scientific discovery. The engineering detail on leakage control, containerized rollouts, and grading separation is also practically useful for teams building agent training infrastructure.

Future Directions

• Verification-coupled context management. The authors plan to use intermediate verification points as boundaries for compressing exploration history into stage records, replacing only the segment between the two most recent verification points while preserving key evidence, rejected hypotheses, unresolved questions, and current task state.
• Longer research horizons. Whether the composition-and-verification approach holds up on research tasks substantially longer than those evaluated remains open.
• Broader scientific validation. The scientific case studies described are demonstrations; systematic evaluation of agentic discovery quality beyond benchmark suites is not established in the provided content.
• Scaling the environment supply. Since real-world task supply, not model capacity, is identified as the binding constraint, further work on automated task synthesis and verification without corrupting reward signals (the stated reason for separating builder, reviewer, and judger agents) is a natural next step.

Target Audience

Researchers and engineers working on agentic LLMs, post-training (SFT and RL) for tool-using models, and multi-agent system design; practitioners building coding, ML-engineering, or deep-research agents who want concrete recipes for data construction, harness design, and leakage-safe evaluation; and scientific-computing groups interested in deploying locally hosted agent models for exploratory research. Readers looking for a controlled ablation study or full numeric tables will need the complete paper, since the benchmark figures and case-study details are not present in the text summarized here.

Authors’ abstract

In this report, we introduce Sage-A1.5, a 35B Mixture-of-Experts agentic model that builds on A1 to advance long-horizon problem solving through adaptive capability composition. Compared with A1, Sage-A1.5 learns not only to solve individual tasks, but also to organize work and revise its approach during execution. Sage-A1.5 highlights three improvements: agentic orchestration, verification, and open scientific discovery. First, agentic orchestration extends single-agent execution through task delegation and parallel collaboration among subagents with isolated contexts, enabling complex tasks to be decomposed and addressed concurrently and supporting efficient execution while reducing interference across concurrent subagents. Second, agentic verification strengthens long-horizon reliability through progress checking, evidence validation, and reflective strategy revision, allowing the agent to recognize unproductive search paths and pursue alternative approaches. Third, open scientific discovery extends agentic capabilities beyond answering predefined questions to conducting exploratory research, connecting scientific inquiry with interactive analysis and inspectable research artifacts. We showcase these capabilities through complicated agentic tasks and multidisciplinary scientific applications. Besides, we hope Sage-A1.5 provides the community with an accessible foundation for open scientific discovery, offering a practical path from long-horizon task execution to collaborative, verifiable, and inspectable scientific exploration.