Skip to content

Inference and Verification

Uni-Agent reuses the same Task, Agent, Sandbox, and reward definitions for parallel inference and verification. Each task runs in its own stateful sandbox, and scores are produced by the task's verifier.

SWE-Bench Verified

Model Score Rollouts Setting
Qwen3-Coder-30B-A3B-Instruct 49.2 Avg@4 100 turns, 128K context
Qwen3-Coder-480B-A35B-Instruct 64.2 Avg@4 500 turns, 256K context
Qwen3-Coder-Next 67.6 Avg@4 300 turns, 128K context
Qwen3.5-4B 45.2 Avg@1 100 turns, 64K context
Qwen3.5-9B 53.8 Avg@1 100 turns, 64K context
Qwen3.5-9B 65.6 Avg@1 200 turns, 128K context
Qwen3.5-35B-A3B 68.4 Avg@1 200 turns, 128K context

The Qwen3-Coder runs use temperature 0.8 and top-p 0.9. The Qwen3.5 runs use task-specific sampling configurations; consult the associated recipe before comparing rows.

SWE-Bench Multilingual

Model Score Rollouts Setting
Qwen3-Coder-30B-A3B-Instruct 32.3 Avg@1 200 turns, 128K context

Terminal-Bench v2

Model Score Rollouts Setting
Qwen3.6-35B-A3B 42.53 Avg@1 200K context

Run the Evaluation

Prepare the dataset and run either inference path described in Run Agent Inference:

  • External API mode for direct endpoint evaluation.
  • verl-managed rollout mode for training-path parity and token-level trajectory collection.

SWE-Bench verification is implemented by the Task reward function. The task executes benchmark tests inside the sandbox and reports resolved, evaluation status, runtime, and a detailed report.

Reporting Checklist

When adding a new row, include:

  • Exact model identifier.
  • Agent implementation or harness.
  • Sandbox backend.
  • Temperature, top-p, and top-k.
  • Max turns and token budget.
  • Number of rollouts.
  • Benchmark revision and verifier.
  • Result JSON or evaluation log.