Sakana AI, in collaboration with researchers at UC Berkeley, has introduced Multi-Agent Self-Supervision (MASS), a framework designed to let AI systems keep improving even when tasks become too complex for humans to reliably evaluate.

Sakana AI is a Tokyo-based research lab founded in 2023 by former Google researchers, among them Llion Jones, a co-author of the 2017 paper that introduced the Transformer architecture underlying much of today's AI. The lab takes its name from the Japanese word for fish — a nod to schooling behavior that matches the collective-intelligence framing of the new work: recursive self-improvement through teams of agents.

The work addresses a growing problem in AI development: as models move beyond humanity's accumulated knowledge, human experts may no longer be able to judge or guide their output. MASS tackles this by making the model its own optimizer and evaluator — a form of homogeneous recursive self-improvement, in which every agent in the loop is a copy of the same underlying model — while attempting to avoid the trap of a system reinforcing its own blind spots.

The framework alternates between two loops. In the inner loop, the model organizes copies of itself into a team of virtual subagents governed by a written workflow that specifies each subagent's role, its instructions, contracts for what information must be exchanged, and the order of calls. An orchestrator executes the workflow, an evaluator compares results against the current champion, and an optimizer proposes revisions. In the outer loop, the best-performing multi-agent trajectories are distilled back into the shared model's weights via fine-tuning, after which the updated model resumes all three roles and the cycle repeats.

After two full cycles starting from a 27B-parameter open-weights model, performance measured as score per output token — a compute-normalized view of capability — reached 1.2 to 1.6 times the base model's level across research benchmarks including MLR-Bench, ScienceAgentBench, DSBench, and AstaBench, with the gains confirmed by external judges withheld from the training process.

Introducing Multi-Agent Self-Supervision (MASS): Recursive self-improvement through collective intelligence Blog: As AI systems tackle increasingly open-ended research problems, human experts may struggle to assess results and guide further progress. The… pic.twitter.com/OWjbfIOEsy

— Sakana AI (@SakanaAILabs) October 11, 2026

Key Findings: Dense Signals, Role Generalization, and New Bottlenecks

The researchers report three principal findings. First, MASS enables role generalization: training only on task-solving trajectories improved not just the executor but also the workflow optimizer and the evaluator, with agreement between the model's judgments and strong external judges rising from 73% to 93% across cycles. This suggests the self-improvement loop can sustain itself, since capabilities for producing and selecting future training data improve alongside task performance.

Second, multi-agent trajectories proved substantially more efficient as training data than single-agent ones. The model reached a 68.3% win rate over the base while using 31% fewer supervised tokens than a comparable single-agent approach — evidence that jointly learning high-level orchestration from orchestrators and bounded execution from subagents provides a denser learning signal per token.

Third, and perhaps most counterintuitive, ablations showed that the bottleneck in recursive self-improvement can be the optimizer rather than the evaluator. Swapping in a stronger external model as the evaluator alone helped little, but upgrading both the evaluator and the optimizer allowed the system to find winning workflows for all 12 test tasks within five iterations, compared with 11 of 12 for the fully homogeneous loop, which simply took longer.

Analysis of the evolved workflows revealed that effective coordination shifts task-specific information into contracts and call order — how agents exchange and verify information — rather than into roles. The authors also emphasize the safety implications: in a homogeneous loop, shared blind spots could allow flawed results to be accepted as progress and reinforced through training, making independent verification mechanisms essential as self-improvement capabilities advance.