PDB-Results: model outputs and scores

📄 Paper (coming soon)  ·  💻 Code  ·  🌐 Project page  ·  🏆 Leaderboard

Raw debugging outputs and evaluator scores for every model evaluated on the PDB (Precise Debugging Benchmarking) suite, so that every reported number can be inspected and recomputed. The benchmarks themselves are anon-pdb/PDB-Single and anon-pdb/PDB-Wild.

Anonymous release for NeurIPS 2026 Datasets & Benchmarks review.

Evaluation sets

Filename tag	Set	Tasks	Models
pdb_single_hard	PDB-Single	5,751 (BigCodeBench 2,525 + LiveCodeBench 3,226)	4
pdb_single	PDB-Single-Full, the unfiltered single-line pool that PDB-Single is drawn from	7,589 (3,697 + 3,892)	9
pdb_multi	PDB-Multi, the multi-line BigCodeBench / LiveCodeBench part of PDB-Wild	256 (37 + 219)	14

The nine models with pdb_single files were run on PDB-Single-Full; their PDB-Single numbers below restrict those files to the task IDs in task_ids/pdb_single_<benchmark>.json.

Files

<benchmark>/debug_results/<model>_on_<benchmark>_<tag>_round_1.json          # model outputs
<benchmark>/eval_results/<model>_on_<benchmark>_<tag>_round_1_scores.json    # evaluator scores
task_ids/<set>_<benchmark>.json                                               # task IDs of each set
results_summary.csv                                                            # union metrics below

• Debug results are lists of benchmark entries (task_id, buggy_code, gt_solution, gt_diff, bug_count, task_prompt, ...) extended with debug_results = {model, solution, pred_diff}: the model's revised program and its line-level edit script relative to buggy_code.
• Scores map each task_id to its unit-test outcome (Unit score, 0 or 1) and to edit-level precision, bug-level recall, F1, and the matched / unmatched edit blocks (Symbolic block scores). Precision is ε-relaxed with ε = 2 on the single-line sets and ε = 1 on PDB-Multi.

Results

Means over BigCodeBench + LiveCodeBench tasks; ± is the 95% interval (1.96 × standard error).

PDB-Single (5,751 tasks)

Model	Precision	Recall	Unit (%)	F1	n
Claude-Opus-4.7	83.3 ± 0.7	89.3 ± 0.6	85.7 ± 0.9	84.1	5,751
Gemini-3.1-Pro	82.5 ± 0.8	89.4 ± 0.7	86.4 ± 0.9	84.3	5,751
Claude-Sonnet-4.5	71.2 ± 0.9	81.3 ± 0.8	76.4 ± 1.1	73.3	5,751
Gemini-2.5-Pro	71.2 ± 0.9	84.0 ± 0.8	79.4 ± 1.0	74.0	5,751
Qwen3-Coder-480B	65.5 ± 1.0	76.9 ± 0.9	70.5 ± 1.2	67.7	5,751
Qwen3.6-Plus	62.9 ± 1.0	77.4 ± 0.9	78.7 ± 1.1	66.4	5,751
Kimi-K2.6	59.2 ± 0.8	83.8 ± 0.8	51.7 ± 1.3	66.4	5,751
Kimi-K2-Instruct	56.2 ± 1.0	72.4 ± 1.0	65.0 ± 1.2	60.0	5,751
Grok-Code-Fast	54.0 ± 1.0	65.9 ± 1.0	58.5 ± 1.3	55.5	5,751
Kimi-K2-Thinking	50.7 ± 0.9	75.4 ± 0.9	75.3 ± 1.1	56.9	5,751
DeepSeek-V3.2	46.9 ± 1.0	69.0 ± 1.0	71.7 ± 1.2	52.2	5,751
DeepSeek-V3.2-Thinking	44.2 ± 0.9	70.8 ± 1.0	80.5 ± 1.0	50.8	5,751
GPT-5.1-Codex	39.4 ± 0.8	72.0 ± 1.0	77.7 ± 1.1	46.9	5,751

PDB-Multi (256 tasks)

Model	Precision	Recall	Unit (%)	F1	n
Gemini-3.1-Pro	83.2 ± 3.7	93.2 ± 2.7	96.5 ± 2.3	85.8	256
Claude-Opus-4.7	70.1 ± 4.6	80.7 ± 4.0	70.7 ± 5.6	71.0	256
Claude-Sonnet-4.5	65.9 ± 4.8	73.9 ± 4.7	64.8 ± 5.9	67.5	256
GPT-5.5	64.7 ± 4.4	86.8 ± 3.6	92.2 ± 3.3	71.0	256
Qwen3-Coder-480B	58.2 ± 4.9	67.3 ± 4.9	56.6 ± 6.1	60.1	256
Gemini-2.5-Pro	57.8 ± 4.9	73.2 ± 4.8	72.7 ± 5.5	61.6	256
Kimi-K2.6	49.1 ± 5.0	65.5 ± 5.6	52.7 ± 6.1	53.9	256
Kimi-K2-Instruct	43.4 ± 4.8	57.9 ± 5.1	44.1 ± 6.1	47.0	256
Grok-Code-Fast	41.6 ± 5.1	48.4 ± 5.2	41.8 ± 6.1	41.9	256
Qwen3.6-Plus	41.5 ± 5.0	59.9 ± 5.3	71.5 ± 5.5	45.2	256
Kimi-K2-Thinking	30.3 ± 4.6	49.0 ± 5.5	71.1 ± 5.6	34.3	256
DeepSeek-V3.2-Thinking	30.0 ± 4.6	47.9 ± 5.5	77.3 ± 5.1	33.9	256
GPT-5.1-Codex	27.9 ± 4.0	59.4 ± 5.4	77.0 ± 5.2	33.9	256
DeepSeek-V3.2	25.4 ± 4.5	38.9 ± 5.5	50.0 ± 6.1	28.3	256

PDB-Single-Full (7,589 tasks)

Model	Precision	Recall	Unit (%)	F1	n
Claude-Sonnet-4.5	77.9 ± 0.7	85.7 ± 0.7	81.9 ± 0.9	79.6	7,589
Gemini-2.5-Pro	77.8 ± 0.7	87.6 ± 0.6	84.0 ± 0.8	79.9	7,589
Qwen3-Coder-480B	73.4 ± 0.8	82.3 ± 0.7	77.4 ± 0.9	75.1	7,589
Kimi-K2-Instruct	65.7 ± 0.8	78.7 ± 0.8	72.9 ± 1.0	68.8	7,589
Grok-Code-Fast	63.6 ± 0.9	73.0 ± 0.8	67.1 ± 1.1	64.7	7,589
Kimi-K2-Thinking	61.0 ± 0.8	81.1 ± 0.7	81.0 ± 0.9	66.1	7,589
DeepSeek-V3.2	58.3 ± 0.9	76.0 ± 0.8	78.3 ± 0.9	62.5	7,589
DeepSeek-V3.2-Thinking	55.8 ± 0.9	77.4 ± 0.8	84.9 ± 0.8	61.2	7,589
GPT-5.1-Codex	50.2 ± 0.8	77.8 ± 0.8	82.3 ± 0.9	56.7	7,589

Model names

File prefix	Model
claude-opus-4.7	Claude-Opus-4.7
gemini-3.1-pro-preview	Gemini-3.1-Pro
claude-sonnet-4-5-20250929	Claude-Sonnet-4.5
gemini-2.5-pro	Gemini-2.5-Pro
Qwen3-Coder-480B-A35B-Instruct-FP8	Qwen3-Coder-480B
qwen3.6-plus	Qwen3.6-Plus
Kimi-K2-Instruct	Kimi-K2-Instruct
kimi-k2	Kimi-K2-Instruct
Kimi-K2-Thinking	Kimi-K2-Thinking
kimi-k2-thinking	Kimi-K2-Thinking
kimi-k2.6	Kimi-K2.6
grok-code-fast-1	Grok-Code-Fast
deepseek-chat	DeepSeek-V3.2
deepseek-reasoner	DeepSeek-V3.2-Thinking
gpt-5.1-codex	GPT-5.1-Codex
gpt-5.5	GPT-5.5