milo-bench-samples
Long-horizon software evolution

Measuring long-horizon software-engineering competence at the granularity of milestones.

Milo-Bench measures long-horizon software-engineering capability, not just isolated coding ability. It evaluates whether an agent can complete milestone-scale engineering tasks that span multiple files, languages, and architectural components while keeping an existing production codebase correct — not the individual functions or short bug fixes that narrower benchmarks target. Where SWE-bench and Multi-SWE-bench score single-issue fixes, Milo-Bench targets whole milestones resolved across a codebase over a long agent trajectory.

  • Milestone-scale tasks spanning multiple files and architectural components
  • Multi-language coverage across 8 programming languages, 19 codebases
  • Continuous partial-credit scoring so long-task progress stays visible
  • Network-sealed verifier — reference solutions cannot influence a run
  • Three frontier models (Claude Opus 4.8, Gemini 3.1 Pro, GPT-5.5) × 3 runs each
Per-model pass rate vs reference-patch complexity ( src_hunks).

Five outcome-based tiers

Tiers are computed from observed model performance across runs. Trivial tasks are solved by nearly every run; Expert tasks by none. This is what the models actually experienced, not any property fixed in advance.

Pass rate by difficulty tier.
Mean agent cost per run by difficulty tier.

Head-to-head, three frontier models

Pass rate captures how hard the tasks are, not how far apart the models are. Mean score keeps crediting partial progress where pass rate bottoms out — this can even flip the model ranking (GPT-5.5 leads by pass@3, Claude Opus 4.8 by mean score).

Model Provider Runs Pass@3 Mean score Total cost

Where the models differ

Tier
Tier

Network-sealed, continuous, partial credit

The verifier computes a continuous score per run: score = recall × regression_factor. A run solves a task only when score is exactly 1.0. The verifier runs network-sealed, so the reference solution cannot influence a run.

1

Task

instruction.md + Docker environment

2

Agent

long tool-use trajectory

3

Patch

edits across files & languages

4

Verifier

network-sealed target tests

5

Score

per-run graded result

8 languages across 19 codebases

Milo-Bench spans eight programming languages drawn from nineteen production codebases; each language card summarises how many of the 30 tasks land in that language.

Browse the 30 tasks

Filter and search the full delivery. Expand a row to see per-model, per-run results and links to the raw trajectory and instruction file.

—
Task Codebase Language Tier Src hunks Pass rate Mean score

Explore the sources