All evidence cases
Terminal-Bench 2.1terminalPython

Recover and tune a PyTorch model architecture

Reconstruct the network represented by a state dictionary, freeze all but output_layer, improve dataset MSE, and export the result as TorchScript.

pytorch-model-recovery

Normalized profile

What this task asks for

Reconstruct the network represented by a state dictionary, freeze all but output_layer, improve dataset MSE, and export the result as TorchScript. Define RecoveredModel to match /app/weights.pt exactly. Leave /app/weights.pt unchanged. Modify only output_layer parameters. Reduce MSE on /app/dataset.pt. Save a TorchScript model compatible with the original state dictionary.

  • feature implementation
  • model training
  • test and validation
  • model architecture
  • state dictionary
  • training loop
  • PyTorch
  • TorchScript
  • transformers
  • /app/model.pt
  • reverse engineering from incomplete or indirect evidence
  • strict behavioral compatibility or exact-output validation
  • changes span multiple components or interfaces

Published evidence

Model × harness outcomes

Counts are tied to these exact configurations. Harnesses are not treated as equivalent across benchmarks.

Models
13
Configurations
20
Trials
100

Showing 20 of 20 configurations

ModelHarnessEffortResultPasses
claude-fable-5anthropic · 2026-06-07claude-code2.1.167xhighMixed1 errors4/5
claude-fable-5anthropic · 2026-06-05terminus-22.0.0highMixed1 errors4/5
claude-opus-4-7anthropic · 2026-05-01claude-code2.1.123maxMixed3 errors2/5
claude-opus-4-7anthropic · 2026-05-01terminus-22.0.0maxMixed1 errors4/5
claude-opus-4-8anthropic · 2026-07-09claude-code2.1.205highAll passed5/5
claude-sonnet-5anthropic · 2026-07-09claude-code2.1.205highAll passed5/5
gemini-3-pro-previewgemini · 2026-05-01gemini-cli0.40.0highAll passed5/5
gemini-3-pro-previewgemini · 2026-05-01terminus-22.0.0highAll passed5/5
gemini-3.1-pro-previewgemini · 2026-05-05gemini-cli0.40.0highMixed3 errors2/5
gemini-3.1-pro-previewgemini · 2026-05-05terminus-22.0.0highAll passed5/5
glm-5.1zai · 2026-05-01claude-code2.1.123maxMixed3 errors2/5
gpt-5.5openai · 2026-05-01codex0.125.0xhighMixed3/5
gpt-5.5openai · 2026-05-01terminus-22.0.0xhighAll passed5/5
gpt-5.6-lunaopenai · 2026-07-10codex0.144.0maxAll passed5/5
gpt-5.6-lunaopenai · 2026-07-11codex0.144.1maxAll passed5/5
gpt-5.6-solopenai · 2026-07-10codex0.144.0maxMixed1/5
gpt-5.6-terraopenai · 2026-07-10codex0.144.0maxAll passed5/5
gpt-5.6-terraopenai · 2026-07-11codex0.144.1maxAll passed5/5
grok-4.5cursor · 2026-07-09cursor-cli2026.07.08-0c04a8adefaultAll passed5/5
muse-spark-1.1openai · 2026-07-09mini-swe-agent2.4.5xhighAll passed5/5