Closed-Loop Evaluation of LLM Agents for Embedded Software Development
Agents write firmware from engineering briefs and test it on simulated equipment. Hidden checks score device behavior across five control tasks.
The authors publish 420 runs for seven models under four feedback conditions. In the primary condition, agents test their own work before one graded submission.
- Made by
- Jorge García-Carrasco and colleagues
- Kind
- Electronics
- Level
- University
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Agentic
- Environment
- Terminal, A simulator
- Items from
- Written by experts
- Refresh
- One fixed set
- Run by
- Anyone
- Strongest model tested
- GPT-5.4 #28 on the capability index
- Last model added
- March 2026
- Scores refreshed
- November 20th, 2026
GPT-5.4 leads with 100% among models released by March 2026
Best score per model from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.
The newest model with a score here came out in March 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 1 GPT-5.4 ci_red_green 100%
- 2
Qwen3.5-27B oracle_full 80%
- 3 GPT-5.4 Mini ci_red_green 80%
- 4
Qwen3.5-35B-A3B oracle_full 66.7%
- 5
Qwen3.5-9B oracle_full 46.7%
- 6
Qwen3.5-4B oracle_full 13.3%
- 7
Qwen3.5-2B ci_red_green 0%
Source: Official Closed-Loop Evaluation of LLM Agents for Embedded Software Development leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-5.4 | 100% | ci_red_green | Mar 2026 | Official leaderboard |
| 2 | | 80% | oracle_full | Feb 2026 | Official leaderboard |
| 3 | GPT-5.4 Mini | 80% | ci_red_green | Mar 2026 | Official leaderboard |
| 4 | | 66.7% | oracle_full | Feb 2026 | Official leaderboard |
| 5 | | 46.7% | oracle_full | Feb 2026 | Official leaderboard |
| 6 | | 13.3% | oracle_full | Feb 2026 | Official leaderboard |
| 7 | | 0% | ci_red_green | Feb 2026 | Official leaderboard |
Scores at each effort level
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: Official Closed-Loop Evaluation of LLM Agents for Embedded Software Development leaderboard
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| GPT-5.4 | ci_red_green | 100% | |||
| GPT-5.4 | oracle_full | 93.3% | |||
| GPT-5.4 | realistic_self_verify | 93.3% | |||
| GPT-5.4 | oneshot_blind | 86.7% | |||
| oracle_full | 80% | ||||
| GPT-5.4 Mini | ci_red_green | 80% | |||
| GPT-5.4 Mini | oracle_full | 73.3% | |||
| oracle_full | 66.7% | ||||
| GPT-5.4 Mini | oneshot_blind | 66.7% | |||
| ci_red_green | 60% | ||||
| realistic_self_verify | 60% | ||||
| oneshot_blind | 53.3% | ||||
| ci_red_green | 53.3% | ||||
| GPT-5.4 Mini | realistic_self_verify | 53.3% | |||
| oracle_full | 46.7% | ||||
| oneshot_blind | 26.7% | ||||
| realistic_self_verify | 26.7% | ||||
| ci_red_green | 20% | ||||
| oracle_full | 13.3% | ||||
| realistic_self_verify | 13.3% | ||||
| oneshot_blind | 13.3% | ||||
| realistic_self_verify | 13.3% | ||||
| ci_red_green | 6.70% | ||||
| ci_red_green | 0% | ||||
| oneshot_blind | 0% | ||||
| oracle_full | 0% | ||||
| realistic_self_verify | 0% | ||||
| oneshot_blind | 0% |
2 models have held number one on the official leaderboard since 2023
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official Closed-Loop Evaluation of LLM Agents for Embedded Software Development leaderboard
GPT-5.4 on the official leaderboard passed 85% in March 2026
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official Closed-Loop Evaluation of LLM Agents for Embedded Software Development leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Engineering and CAD
- EEBench In use 13 electrical-engineering circuit and PCB design tasks graded by deterministic build checks, simulation and BOM constraints. No LLM judge.
Gemini 4 Carbon 83.9%
- 3DCodeBench In use Agentic procedural 3D modeling via code: author engine-ready 3D assets through software APIs. Graded for executability and shape fidelity.
LongCat-2.5 55.4%
- CADGenBench Only in system cards Agents construct CAD models from design prompts. The generation split is graded for executability and geometric correctness. Claude Fable 5 45.1%
- PartBench Only in system cards Parametric mechanical part design and CAD-style generation, graded for geometric and manufacturing fidelity. Grok 4.6 58.5%
- Razavi-Bench Not in model cards Models explain what happens in analog circuits and why. A language model awards up to four points per answer against a reference and a rubric.
Gemini 4 Carbon 98.3%
- CircuitRubric Not in model cards Models write analog circuit netlists from design requests. A program compares transistor connections and relative device sizes with accepted reference circuits. Claude Opus 4.8 96%
Data audit
All 8 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 2 pass 3 notes
- Pass Independent runs Scores from the official leaderboard (7 entries).
- Pass Models scored 7 models we track.
- Note Newest score The newest came 8 months ago, in March 2026, for GPT-5.4 Mini. That’s expected for a saturated test.
- Note Frontier models None of the top 10 has a score, which is fine for a retired test. The strongest tested is GPT-5.4 (#28).
- Note Headroom Saturated. The best is 100%, past 85% since March 2026.
Leaderboard import 3 pass
- Pass Importer We read the board on github.com: 28 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.