Sakana AI's RL Conductor: A 7B Model That Outperforms GPT-5 by Orchestrating AI Teams

A 7-billion-parameter language model, trained on just 960 examples with two H100 GPUs, outperforms GPT-5, Claude Sonnet 4, and Gemini 2.5 Pro on complex reasoning tasks.

It doesn’t solve problems directly. It orchestrates the models that do.

The RL Conductor, developed by Sakana AI in collaboration with the University of Michigan and the Institute of Science Tokyo, was accepted at ICLR 2026 [1]. It treats multi-agent coordination not as a routing problem (pick the best model) but as a meta-prompt engineering problem — generating natural-language workflows that divide, delegate, and verify across a pool of frontier models.

Here’s how it works, what it learned, and why it changes how we think about multi-agent systems.

The problem: Hardcoded pipelines don’t scale

Most multi-agent systems today use either:

Hardcoded pipelines (LangChain-style DAGs) — fixed topology per task type. They break when query distribution shifts.
Simple routers — pick a single model. They miss the value of complementary strengths across models.
Mixture-of-Agents — broadcast every subtask to every model. They work, but waste 6x tokens [2].

The RL Conductor solves all three: it learns to build custom workflows on the fly, using only the models it needs, with minimal token overhead.

How RL Conductor works

The Conductor is a fine-tuned Qwen2.5-7B that generates agentic workflows in natural language [1]. For each input, it outputs three Python lists:

model_id = [2, 0]                          # which workers to call
subtasks = ["Plan the approach",           # what to ask each
            "Implement in Python"]
access_list = [[], ["all"]]                # what context to pass

Each step specifies:

Which agent to call (by integer ID)
What specific subtask to assign (full natural language prompt)
What prior messages to include in context (using an access list)

The output is parsed, executed sequentially against the worker pool, and the final answer is extracted. Workflows can have up to 5 steps and produce chains, parallel forks, or recursive loops.

Training: GRPO on 960 problems

The Conductor was trained using GRPO (Group Relative Policy Optimization) — the same RL algorithm behind DeepSeek-R1 [1]. For each training problem, the model generates multiple workflow candidates. Rewards are:

Format reward: r = 1 if the output parses correctly, r = 0 if not
Correctness reward: r = 1 if final answer matches solution, r = 0.5 if not

Training dataset: 960 problems from MATH500, MMLU, RLPR, and LiveCodeBench V1 [1]. That’s it. 960 examples on two H100 GPUs.

Benchmarks: Surpassing every worker

Benchmark	GPT-5	Gemini 2.5 Pro	Claude Sonnet 4	RL Conductor (7B)
MATH500	99.0	96.0	96.0	99.4
MMLU	93.5	92.4	91.4	94.1
RLPR	42.20	40.55	36.70	44.75
LiveCodeBench	82.90	67.24	46.54	83.93
AIME25	90.8	78.3	74.3	93.3
BigCodeBench	32.75	37.51	37.16	37.86
GPQA-Diamond	82.3	84.8	77.7	87.5
Average	74.78	70.97	65.69	77.27

Source: Table 1 in the paper [1].

The Conductor beats every single worker on every benchmark. The gains on AIME25 (+2.5% over GPT-5) and GPQA-Diamond (+2.7% over Gemini 2.5 Pro) are roughly equivalent to a full generational model upgrade [2].

Cost efficiency

The Conductor uses an average of 1,820 tokens per question — compared to 11,203 for Mixture-of-Agents (MoA) and 9,909 for Smoothie [2]. At API pricing, that’s $0.02384 per question vs. $0.04855 for MoA [1].

Average workflow length: 3 steps (well below the 5-step max) [1]. The Conductor learned to be concise.

Emergent strategies discovered by RL

The most interesting results aren’t the benchmarks — they’re the behaviors the Conductor discovered on its own.

Task difficulty calibration

For simple factual questions, the Conductor uses a single call. For hard coding problems, it autonomously spins up a 4-agent pipeline with planner, implementer, verifier, and reviewer [2].

Model specialization

The Conductor learned which models to use for which roles:

Gemini 2.5 Pro and Claude Sonnet 4 get assigned as high-level planners and idea generators
GPT-5 is reserved for final optimized code generation [3]

It discovered the comparative advantages of these models by trial and error during RL training.

Role abdication

In some workflows, the Conductor hands the entire planning process to Gemini 2.5 Pro — effectively letting one worker dictate subtasks for the rest of the pool [2]. This emergent delegation pattern was not programmed.

Recursive test-time scaling

One of the paper’s most practical contributions: the Conductor can call itself as a worker [1]. When allowed to recurse, the Conductor reads its own earlier output, detects failures, and spins up corrective workflows.

Results on BigCodeBench improve from 37.8 to 40.0 with recursion enabled [1]. This opens a new axis for test-time compute scaling — instead of chain-of-thought length or beam search width, you scale by orchestration depth.

“This work proves that language models can become elite meta-prompt engineers, dynamically harnessing collective intelligence.” — Sakana AI research team [2]

Sakana Fugu: The commercial product

The RL Conductor research powers Sakana Fugu, Sakana AI’s commercial multi-agent orchestration service now in beta [3]. Fugu exposes the Conductor through a standard OpenAI-compatible API:

Fugu Mini — low-latency for production serving
Fugu Ultra — maximum performance for complex workloads

Target industries include finance and defense, where rigid hardcoded pipelines have failed to deliver productivity gains in multi-agent setups [3].

What this means for practitioners

Small orchestration models beat big monolithic models. A 7B Conductor with RL training outperforms GPT-5 at 1/50th the parameter count. The bottleneck isn’t model capacity — it’s coordination strategy.
RL is emerging as the dominant training paradigm for agent behavior. GRPO on 960 examples produced emergent strategies that no hand-crafted prompt engineering could match. Expect this pattern to spread.
Recursive orchestration is a new compute scaling axis. When you can’t make the model bigger or sample more chains, let the conductor replan in depth. This is the multi-agent equivalent of chain-of-thought.
Token budget matters. The Conductor achieves higher scores using 1,820 tokens vs. MoA’s 11,203. For production systems with cost constraints, orchestration efficiency is the lever to pull.

References

[1] Nielsen, S., Cetin, E., Schwendeman, P., Sun, Q., Xu, J., Tang, Y. “Learning to Orchestrate Agents in Natural Language with the Conductor.” ICLR 2026. arXiv:2512.04388. https://arxiv.org/abs/2512.04388

[2] Sakana AI. “Learning to Orchestrate Agents in Natural Language with the Conductor.” Official blog post. https://sakana.ai/learning-to-orchestrate/

[3] Dickson, B. “How Sakana trained a 7B model to orchestrate GPT, Claude and Gemini LLMs.” VentureBeat, May 7, 2026. https://venturebeat.com/orchestration/how-sakana-trained-a-7b-model-to-orchestrate-gpt-5-claude-sonnet-4-and-gemini-2-5-pro

ToolBrain — tool reviews, LLM comparisons, and AI workflow guides
CodeIntel Log — code quality, debugging, and software engineering benchmarks
Hermes Tutorials — Hermes Agent setup, configuration, and advanced workflows

Cross-links automatically generated from NiteAgent.

← Back to all posts