ARIS: Autonomous Research via Adversarial Multi-Agent Collaboration

Summary

ARIS is an open-source research harness for autonomous ML research, built around the assumption that "any long-term task performed by a single agent is unreliable." It coordinates research workflows through cross-family adversarial collaboration: an executor model drives forward progress while a reviewer drawn from a different model family critiques intermediate artifacts and requests revisions. The system is organised into three architectural layers — execution (65+ reusable Markdown-defined skills, MCP integrations, persistent research wiki, deterministic figure generation), orchestration (five end-to-end workflows with adjustable effort settings and reviewer routing), and assurance (claim auditing, mathematical-proof checks, scientific editing pipeline, visual PDF inspection).

Contribution

A concrete instantiation of harness engineering for autonomous ML research that treats assurance as a first-class workflow layer rather than a single review pass. The three named bottlenecks — persistent research state, modular execution, independent assurance — are presented as system-level consequences of the "single-agent long-horizon research is unreliable" assumption, not as separate desiderata bolted on after the fact. A prototype self-improvement loop records research traces and proposes harness changes, gated by reviewer approval.

Method

System-design technical report with early deployment experience across three executor platforms. No controlled benchmark study; the paper documents architecture, assurance mechanisms, and qualitative deployment observations rather than measured win-rates against baselines.

Relevance to RISE

ARIS is one of the most architecturally explicit recent agentic research harnesses and shares the RISE perspective that workflow harness (skills, state, review) matters as much as model weights. The cross-family executor/reviewer pairing is a clean operational answer to single-agent failure modes; the three-stage claim audit is a useful template for evaluation rigor in RISE-style pipelines. ARIS is also the canonical reference behind the aris skills pack bundled in this KB.

Critique / open questions

The "any long-term task by a single agent is unreliable" framing is stated as an assumption rather than an empirical finding; the paper acknowledges it may understate current single-agent capabilities and prefers the strict version for high-rigor settings. The "more than 65 skills" and three-layer architecture are not yet evaluated against ablations (e.g., what if assurance is removed? what if same-family review is used?). Cross-family review introduces real coordination cost; the trade-off vs. single-family pipelines is asserted but not quantified.

Key quotes

"Any long-term task performed by a single agent is unreliable. We need to divide the total workflow into sub-workflows and cross-family models to review the output at each step independently."

"The central failure mode is not visible breakdown but plausible unsupported success: a long-running agent can produce claims whose evidential support is incomplete, misreported, or silently inherited from the executor's framing."

"We default to cross-family pairings because prior work suggests that mixed-model agent configurations can produce less correlated and more varied critiques … we adopt this as a recommended configuration rather than a hard system constraint."