Frontis / OpenRSI Integration for MLEvolve

An experimental integration of Frontis/OpenRSI-released language models into the open-source MLEvolve machine-learning engineering framework.

This repository is based on InternScience/MLEvolve. The work here focuses on one concrete question:

Can a Frontis/OpenRSI model serve as an MLEvolve backend through an OpenAI-compatible API while preserving the structured interactions the agent pipeline expects?

Contribution boundary

The underlying MLEvolve search engine, agents, memory, execution, and evaluation framework are upstream work.

My integration changes are concentrated in:

• llm/model_profiles.py — Frontis model profile/capability handling;
• llm/openai.py — structured-output compatibility and tolerant JSON extraction;
• integrations/frontis/ — endpoint probes, smoke tests, native mini-loops, backend comparison, and real-task runners;
• cross-platform guards for runtime assumptions such as CPU affinity.

This repository does not claim ownership of MLEvolve itself.

Integration architecture

Frontis / OpenRSI model
        │
        │ OpenAI-compatible API
        ▼
llm/model_profiles.py
        │
        ▼
llm/openai.py
   ├── native structured output when supported
   └── prompt-only JSON fallback + defensive parsing
        │
        ▼
upstream MLEvolve agents / search / execution
        │
        ▼
integration validation harness
integrations/frontis/

Validation workflow

Start with the interface boundary:

python integrations/frontis/endpoint_probe.py \
  --model Frontis-MA1-30B \
  --base-url http://127.0.0.1:8000/v1

Then run the structured smoke test:

python integrations/frontis/smoke_test.py \
  --model Frontis-MA1-30B \
  --base-url http://127.0.0.1:8000/v1

Then test a small native MLEvolve loop:

python integrations/frontis/native_mini_loop.py \
  --model Frontis-MA1-30B \
  --base-url http://127.0.0.1:8000/v1

The real-task runner is available for an already prepared MLE-Bench dataset:

python integrations/frontis/real_task.py \
  --task <competition-id> \
  --dataset-dir <mle-bench-root> \
  --model Frontis-MA1-30B \
  --base-url http://127.0.0.1:8000/v1 \
  --allow-prompt-tool-fallback

See integrations/frontis/README.md for the full validation harness.

Engineering issues addressed

1. Explicit model profiling

Frontis is handled explicitly at the model boundary instead of being silently treated as another model family based only on implementation ancestry.

2. Structured-output fallback

MLEvolve frequently expects schema-shaped outputs. When a backend does not reliably support the exact structured-output mechanism used by another provider, the integration can enforce JSON through prompting and recover a schema-shaped object defensively.

3. OpenAI-compatible endpoints

The integration preserves the existing MLEvolve call path while allowing a separately served compatible endpoint.

4. Cross-platform validation

Small tests isolate endpoint/model behavior before expensive native search runs, and platform-specific runtime assumptions are guarded where needed.

Upstream projects

• MLEvolve: https://github.com/InternScience/MLEvolve
• OpenRSI / Frontis: https://github.com/FrontisAI/OpenRSI

Please refer to those projects for the original frameworks, papers, model releases, setup, and licenses.

Status

This is an integration experiment rather than a new ML-engineering framework. The current focus is backend compatibility and discriminative testing, not reproducing or claiming upstream leaderboard results.