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

What it claims

Sage-A1.5 is a 35B Mixture-of-Experts agent model from Shanghai Artificial Intelligence Laboratory, released October 2026. It builds on Agents-A1, which the report calls "A1" in its figure. The report's Figure 1 compares it on twelve benchmarks against two 35B models (Agents-A1, Ornith-1.5) and three larger ones (Kimi-K3, GLM-5.3, GPT-5.6-sol). The authors say it beats the larger models on FrontierScience-Olympiad, CMT, PinchBench and IFBench. They say it is competitive with them on SWE-Bench Pro and BrowseComp, and that it matches or beats both 35B models on all twelve benchmarks. The figure also shows it trailing the larger models on SciCode, MLE-Lite, HLE with tools, ALE-CLI and SWE-Multilingual.

How it works

Orchestration runs through a spawn_subagent(tasks) call. Each task carries three fields: task (a self-contained subproblem), context (background, evidence, constraints) and expected_output. Subagents search and read pages in separate contexts within a concurrent batch. They report only after the whole batch completes.

Each subagent returns JSON with five fields: answer, evidence (each item has url, snippet, supports), unresolved_issues, and status. The status is one of ok, unknown, insufficient_evidence, conflict or ambiguous.

The report names two patterns. In exploratory composition, agents pursue alternative hypotheses (A∨B∨C). In constructive composition, they solve complementary subproblems (A∧B∧C).

flowchart LR
  M[Main agent] -->|answer directly| R[Answer / report]
  M -->|spawn_subagent| S1[Subagent 1]
  M --> S2[Subagent n]
  S1 --> F[Structured JSON feedback]
  S2 --> F
  F -->|gaps or conflicts| M
  F -->|sufficient| R

Orchestration training data uses "starlike tree" questions. Several relation chains lead to one answer entity, and names are masked, so each branch can be investigated in parallel. A teacher model runs in the orchestration harness. Its main-agent trajectory and each subagent trajectory become separate training samples.

For verification training, a supervisor model inserts reflection points during data generation. It does this at repeated searches, stalls, or key conclusions. The supervisor prompts are then removed, and the reflections are folded into the model's own reasoning text. At inference no external verifier or online supervisor runs. A conditioning marker switches the capability on demand; the fetched text cuts off before explaining it.

The overall pipeline has two stages. First, multi-domain teachers are trained with SFT and RL across coding, long-context reasoning, deep research, science and co-working. Second, these teachers go through unified outcome-process post-training with mixed rewards.

Evidence

The scores are read from Figure 1's bar chart, in the order Sage-A1.5 / A1 / Ornith-1.5 / Kimi-K3 / GLM-5.3 / GPT-5.6-sol:

On SciCode, Sage-A1.5 ties A1 rather than beating it. The largest gains over A1 are on SWE-Bench Pro (+17.6), MLE-Lite (+21.3) and CritPt (+15.7).

Limits

The fetched text stops partway through Section 2.1.2. The evaluation settings (Section 3.1), the ablations, the case studies and the authors' own limitations section (Section 5) are not included. Figure 1 gives no metrics, harnesses or sampling settings for its scores. The page links a "Model" and a "Token-plan", but the text states no weights location, licence, active-parameter count, context length or serving requirements. The report mentions local deployment for privacy and offline use but gives no detail.