Principle 01
Sandboxed Execution
- Every trial runs in an isolated container
- No internet access, no host filesystem leakage
- Reproducible environment via Docker images
§01 · Overview
MARS is a reinforcement learning environment for training autonomous agents to resolve complex CLI (command line interface) tasks. Each task includes a containerized environment with a human curated problem based on real-world operational scenarios, a natural language problem description, and a Docker container for sandboxed execution. Reward signals are generated using test suites that verify the final container state against expected outcomes. Two frontier models were tested on 10,000 instances covering 6 task categories across 3 difficulty tiers.
§02 · Key metrics
Three numbers that define the scope of MARS.
Instances
10,000
100% reward signal coverage
Models Used
2
GLM-5 & Kimi K2.5
Categories
6
Difficulty Tiers
3
Easy, Medium, Hard
§03 · Methodology
How the evaluation scores are computed.
Evaluation framework
Four principles govern the MARS evaluation system.
Principle 02
Automated Verification
- Test scripts execute automatically post-agent
- Binary reward: pass (1) or fail (0)
- No human evaluation in the loop
Principle 03
Single Attempt
- Each agent gets exactly one attempt per task
- Mirrors real-world single-submission workflow
- No retry, no cherry-picking best runs
Principle 04
Cost Tracking
- Full token usage (input + output) recorded
- USD cost computed per trial via LLM pricing
- Enables cost-efficiency comparisons
§04 · Evaluation pipeline
Three stages turn a raw task into a scored RL environment.
Phase 01
Environment Construction
- Identify target real-world programming problems
- Write human curated reward signals (deterministic test harnesses)
- Package sandboxed runtimes with graded difficulty metadata
Phase 02
Agent Execution
- Deploy frontier models into isolated environments
- Agents produce solutions under resource constraints
- Capture per-environment outputs
Phase 03
Reward Signal Evaluation
- Execute reward signal harnesses against agent outputs
- Score binary pass/fail per test case
- Aggregate to environment-level and category-level pass rates
§05 · Results
GLM-5 achieves overall pass rate 65%. Kimi K2.5 achieves overall pass rate 45%.
Both models show clear degradation as difficulty increases: from ~86% at Easy to ~33% (GLM-5) and 0% (Kimi K2.5) at Hard.
GLM-5 achieves higher pass rates at ~48× lower cost. ($0.72 vs $34.89)
Success rate by difficulty tier for two frontier models.
Difficulty breakdown
GLM-5 maintains a consistent lead across all tiers. Kimi K2.5 collapses entirely on Hard environments.
| Difficulty | GLM-5 | Kimi K2.5 |
|---|---|---|
| Easy | 85.7% | 71.4% |
| Medium | 71.4% | 57.1% |
| Hard | 33.3% | 0.0% |
Category breakdown
Performance varies sharply by category. File-operations and algorithms show the highest GLM-5 pass rates. Data-querying and optimization expose the largest model gap.
| Category | GLM-5 | Kimi K2.5 |
|---|---|---|
| algorithms | 100.0% | 100.0% |
| data-querying | 33.3% | 0.0% |
| debugging | 66.7% | 66.7% |
| file-operations | 100.0% | 50.0% |
| optimization | 50.0% | 0.0% |
| software-engineering | 66.7% | 50.0% |
§06 · Dataset Viewer
Sample instances evaluated on GLM-5 and Kimi K2.5.
| # | Instance | Category | Difficulty | GLM-5 | GLM-5 Time | Kimi K2.5 | Kimi K2.5 Time | Language |
|---|
§07 · Model comparison
GLM-5
- Overall Pass Rate
- 65.0%
- Easy
- 85.7%
- Medium
- 71.4%
- Hard
- 33.3%
- Total Cost
- $0.72
Kimi K2.5
- Overall Pass Rate
- 45.0%
- Easy
- 71.4%
- Medium
- 57.1%
- Hard
- 0.0%
- Total Cost
- $34.89
GLM-5 achieves higher pass rates at 48× lower cost.
§08 · Resources