TeamBench: Evaluating Agent Coordination under Enforced Role Separation
Paper • 2605.07073 • Published
Datasets:
task_id stringlengths 10 33 | title stringlengths 1 80 | category stringclasses 19
values | difficulty stringclasses 4
values | has_generator bool 2
classes | ablation_scores stringclasses 129
values | tni float32 -2 2 ⌀ | classification stringclasses 4
values |
|---|---|---|---|---|---|---|---|
API1_version_compat | Version Compat | Other | hard | true | null | null | null |
CR1_review_respond | Review Respond | Code Review | medium | true | {"oracle": 0.5, "restricted": 0.5, "team": 0.0, "team_no_plan": 0.62, "team_no_verify": 0.62} | null | TEAM-HURTS |
CR2_style_enforce | Style Enforce | Code Review | easy | true | {"oracle": 0.92, "restricted": 0.69, "team": 0.77, "team_no_plan": 0.77, "team_no_verify": 0.69} | 0.3478 | HIGH-TNI |
CR3_perf_review | Perf Review | Code Review | medium | true | {"oracle": 0.9, "restricted": 0.9, "team": 1.0, "team_no_plan": 0.9, "team_no_verify": 1.0} | null | TEAM-HELPS |
CR4_api_review | Api Review | Code Review | hard | true | {"oracle": 0.56, "restricted": 0.56, "team": 0.62, "team_no_plan": 0.56, "team_no_verify": 0.25} | null | TEAM-HELPS |
CR5_test_coverage | Test Coverage | Testing | medium | true | {"oracle": 0.0, "restricted": 0.12666666666666668, "team": 0.25666666666666665, "team_no_plan": 0.12666666666666668, "team_no_verify": 0.15333333333333335} | -1.0263 | TEAM-HELPS |
CR6_review_checklist | Review Checklist | Code Review | hard | true | {"oracle": 0.06, "restricted": 0.06, "team": 0.30333333333333334, "team_no_plan": 0.2733333333333333, "team_no_verify": 0.30333333333333334} | null | TEAM-HELPS |
CROSS1_api_contract | Go server + Python client with 3 field name/format mismatches to reconcile | Cross-System Integration | hard | true | {"oracle": 1.0, "restricted": 1.0, "team": 1.0, "team_no_plan": 0.9, "team_no_verify": 0.4} | null | NEUTRAL |
CROSS2_schema_evolution | Schema Evolution | Cross-System Integration | hard | true | {"oracle": 1.0, "restricted": 1.0, "team": 1.0, "team_no_plan": 1.0, "team_no_verify": 0.8889} | null | NEUTRAL |
CROSS3_protocol_bridge | JSON->proto3 bridge with 6 type bugs: int64, bytes, oneof, enum, 2 error codes | Cross-System Integration | hard | true | {"oracle": 1.0, "restricted": 1.0, "team": 1.0, "team_no_plan": 1.0, "team_no_verify": 1.0} | null | NEUTRAL |
CROSS4_auth_federation | Auth Federation | Cross-System Integration | hard | true | {"oracle": 1.0, "restricted": 1.0, "team": 1.0, "team_no_plan": 1.0, "team_no_verify": 1.0} | null | NEUTRAL |
CROSS5_event_schema | Event Schema | Cross-System Integration | hard | true | {"oracle": 0.7, "restricted": 1.0, "team": 0.7, "team_no_plan": 0.7, "team_no_verify": 1.0} | 1 | TEAM-HURTS |
CROSS6_grpc_rest_bridge | gRPC-style service definition + REST gateway with 4 type conversion errors | Multi-language | hard | true | {"oracle": 1.0, "restricted": 1.0, "team": 0.0, "team_no_plan": 1.0, "team_no_verify": 1.0} | null | TEAM-HURTS |
CROSS7_config_drift | Config Drift | Operations | hard | true | {"oracle": 0.6, "restricted": 0.6, "team": 0.6, "team_no_plan": 0.6, "team_no_verify": 0.6} | null | NEUTRAL |
CRYPTO1_nonce_reuse | AES-GCM counter nonce, PBKDF2 1000 iters, 8-byte tag: 3 bugs, salt is correct | Security | expert | true | {"oracle": 0.5, "restricted": 1.0, "team": 1.0, "team_no_plan": 0.7, "team_no_verify": 1.0} | 0 | NEUTRAL |
CRYPTO3_cert_chain | X | Security | expert | true | {"oracle": 0.06666666666666667, "restricted": 0.06666666666666667, "team": 0.06666666666666667, "team_no_plan": 0.0, "team_no_verify": 0.13333333333333333} | null | NEUTRAL |
CRYPTO4_kdf_migration | MD5->Argon2id migration with format marker bug and hex-salt-as-base64 error | Security | expert | true | {"oracle": 0.20000000000000004, "restricted": 0.06666666666666667, "team": 0.20000000000000004, "team_no_plan": 0.20000000000000004, "team_no_verify": 0.20000000000000004} | 1 | HIGH-TNI |
CRYPTO5_tls_config | Tls Config | Security | hard | true | {"oracle": 0.3, "restricted": 0.19999999999999998, "team": 0.3, "team_no_plan": 0.3, "team_no_verify": 0.19999999999999998} | 1 | TEAM-HELPS |
CRYPTO6_token_validation | Token Validation | Security | hard | true | {"oracle": 0.9, "restricted": 0.9, "team": 0.9, "team_no_plan": 0.9, "team_no_verify": 0.9} | null | NEUTRAL |
D1_schema_drift | Schema Drift | Data Engineering | medium | true | {"oracle": 0.44, "restricted": 0.5, "team": 0.5, "team_no_plan": 0.38, "team_no_verify": 0.38} | 0 | NEUTRAL |
D2_data_quality | Data Quality | Data Engineering | medium | true | {"oracle": 0.27, "restricted": 0.82, "team": 0.36, "team_no_plan": 0.82, "team_no_verify": 0.36} | 0.8364 | TEAM-HURTS |
D3_schema_migration | Schema Migration | Data Engineering | hard | true | {"oracle": 0.73, "restricted": 0.82, "team": 0.82, "team_no_plan": 0.73, "team_no_verify": 0.82} | 0 | NEUTRAL |
D4_data_pipeline | Data Pipeline | Data Engineering | hard | true | {"oracle": 0.59, "restricted": 0.59, "team": 0.0, "team_no_plan": 0.59, "team_no_verify": 0.65} | null | TEAM-HURTS |
D5_query_optimize | Query Optimize | Data Engineering | hard | true | {"oracle": 0.8, "restricted": 0.87, "team": 0.8, "team_no_plan": 0.93, "team_no_verify": 0.87} | 1 | TEAM-HURTS |
D6_data_reconcile | Data Reconcile | Data Engineering | expert | true | {"oracle": 0.83, "restricted": 0.83, "team": 0.89, "team_no_plan": 0.83, "team_no_verify": 0.72} | null | TEAM-HELPS |
D7_etl_reconciliation | Etl Reconciliation | Data Engineering | hard | true | {"oracle": 1.0, "restricted": 1.0, "team": 1.0, "team_no_plan": 1.0, "team_no_verify": 0.9} | null | NEUTRAL |
D8_csv_cleanup | Csv Cleanup | Data Engineering | easy | true | {"oracle": 0.2966666666666667, "restricted": 0.26, "team": 1.0, "team_no_plan": 0.5933333333333334, "team_no_verify": 1.0} | null | TEAM-HELPS |
D9_schema_validation | Schema Validation | Data Engineering | medium | true | {"oracle": 0.8, "restricted": 0.8, "team": 0.9, "team_no_plan": 0.8, "team_no_verify": 0.8} | null | TEAM-HELPS |
DB1_migration_conflict | Migration Conflict | Data Engineering | hard | true | null | null | null |
DB2_query_regression | Query Regression | Data Engineering | hard | true | null | null | null |
DEVOPS1_dockerfile_fix | Dockerfile Fix | Operations | medium | true | null | null | null |
DEVOPS2_cicd_pipeline | Cicd Pipeline | Data Engineering | hard | true | null | null | null |
DIST1_queue_race | Queue Race | Distributed Systems | expert | true | {"oracle": 0.3333, "restricted": 0.3333, "team": 0.4167, "team_no_plan": 0.25, "team_no_verify": 0.4167} | null | TEAM-HELPS |
DIST2_consensus_partition | Consensus Partition | Distributed Systems | expert | true | {"oracle": 0.4167, "restricted": 0.4167, "team": 0.44446666666666673, "team_no_plan": 0.4167, "team_no_verify": 0.4167} | null | NEUTRAL |
DIST3_idempotency | Idempotency | Distributed Systems | hard | true | {"oracle": 0.5277666666666666, "restricted": 0.6667, "team": 0.4166666666666667, "team_no_plan": 0.7222, "team_no_verify": 0.8055333333333333} | 1.7997 | TEAM-HURTS |
DIST4_clock_skew | Distributed clock skew handling bugs in event ordering and lease expiry | Distributed Systems | hard | true | {"oracle": 0.9167, "restricted": 0.9444666666666667, "team": 0.9722333333333334, "team_no_plan": 0.8333666666666666, "team_no_verify": 0.7222333333333334} | null | NEUTRAL |
DIST5_leader_election | Bully leader election with 3 bugs + 2 network partition edge cases | Distributed Systems | hard | true | {"oracle": 0.2, "restricted": 0.2, "team": 0.2, "team_no_plan": 0.2, "team_no_verify": 0.2} | null | NEUTRAL |
DIST6_distributed_lock | Distributed Lock | Distributed Systems | hard | true | {"oracle": 1.0, "restricted": 0.7000000000000001, "team": 0.6333333333333333, "team_no_plan": 0.5333333333333333, "team_no_verify": 1.0} | -0.2222 | TEAM-HURTS |
EA1_security_scan | Security Scan | Security | hard | true | null | null | null |
EA2_coverage_gap | Add tests to reach 90% branch coverage using coverage analysis findings | Testing | hard | true | null | null | null |
EA4_code_quality | Code Quality | Code Review | hard | true | null | null | null |
EA5_dependency_audit | Fix 5 CVEs: 3 pin-only upgrades and 2 requiring API migration | Security | hard | true | null | null | null |
GH1000_numpy_19869 | BUG: ensure np | Other | medium | true | null | null | null |
GH1001_numpy_30855 | BUG: fix infinite recursion in np | Other | medium | true | null | null | null |
GH1002_scipy_24753 | BUG: Rotation | Other | medium | true | null | null | null |
GH1003_scipy_24496 | BUG:sparse: make `sum` apply `dtype` before accumulation | Other | medium | true | null | null | null |
GH1004_dask_11665 | Fix filtering on parquet file containing a struct column | Other | medium | true | null | null | null |
GH1005_ray_60236 | [Data] - Only return selected data columns in hive partitioned parquet files | Other | medium | true | null | null | null |
GH1006_spaCy_13068 | Fix displacy span stacking | Other | medium | true | null | null | null |
GH1007_spaCy_13040 | Revert 'Load the cli module lazily for spacy | Other | medium | true | null | null | null |
GH1008_spaCy_12749 | Add spancat_singlelabel to debug data CLI | Other | medium | true | null | null | null |
GH1009_spaCy_12486 | Fix pickle for ngram suggester | Other | medium | true | null | null | null |
GH1010_great_expectati_10406 | [BUGFIX] Fix Databricks SQL Regex and Like based Expectations | Other | medium | true | null | null | null |
GH1011_great_expectati_8062 | Great Expectati 8062 | Other | medium | true | null | null | null |
GH1012_autogluon_3480 | Tabular: Fix train_test split edge-case | Other | medium | true | null | null | null |
GH1013_autogluon_4290 | [tabular] Fix Stacker max_models logic | Other | medium | true | null | null | null |
GH1014_pymc_7809 | Only add `dim_length` shared variable when adding a new coordinate | Other | medium | true | null | null | null |
GH1015_scikit_learn_28095 | FIX divide by zero in line search of GradientBoostingClassifier | Other | medium | true | null | null | null |
GH1016_scikit_learn_17575 | FIX escape double quotes when exporting tree with Graphviz | Other | medium | true | null | null | null |
GH1017_scikit_learn_21654 | [MRG] FIX segmentation fault on memory mapped contiguous memoryview | Other | medium | true | null | null | null |
GH1018_scikit_learn_25432 | FIX Support readonly sparse datasets for `manhattan_distances` | Other | medium | true | null | null | null |
GH1019_transformers_36455 | Fix loading zero3 weights | Other | medium | true | null | null | null |
GH1020_transformers_20318 | Fix flakey no_trainer test with seed | Other | medium | true | null | null | null |
GH1021_transformers_17331 | Transformers 17331 | Other | medium | true | null | null | null |
GH1022_keras_20768 | fix(ops): Fix issue with map_coordinates for uint8 dtype | Other | medium | true | null | null | null |
GH1023_pytorch_lightni_21187 | Fix edgecase in batch size scaling tuner alg | Other | medium | true | null | null | null |
GH1024_pytorch_lightni_6942 | [fix] Add a cluster environment teardown to clean up environment state | Other | medium | true | null | null | null |
GH1025_numpy_30818 | BUG: Add test to reproduce problem described in #30816 | Other | medium | true | null | null | null |
GH1026_numpy_30521 | BUG: validate contraction axes in tensordot | Other | medium | true | null | null | null |
GH1027_scipy_24761 | BUG: `from_rotvec` read-only support | Other | medium | true | null | null | null |
GH1028_statsmodels_8633 | ENH/BUG: archimedean k_dim > 2, deriv inverse in generator transform | Other | medium | true | null | null | null |
GH1029_dask_12081 | Fix mixed HLG/Expr handling in ``_ExprSequence | Other | medium | true | null | null | null |
GH1030_mlflow_21824 | Fix tar path traversal vulnerability in `extract_archive_to_dir` | Other | medium | true | null | null | null |
GH1031_mlflow_21922 | Fix Anthropic structured outputs compatibility in gateway adapter | Other | medium | true | null | null | null |
GH1032_ray_29438 | [Tune] Catch SyncerCallback failure with dead node | Other | medium | true | null | null | null |
GH1033_ray_23187 | `map` and `map_unordered` cancel previous tasks before submitting new ones | Other | medium | true | null | null | null |
GH1034_airflow_64145 | Fix FAB DB manager discovery in migration-only contexts | Other | medium | true | null | null | null |
GH1035_autogluon_5475 | [timeseries] fixes to Toto and add to smoke tests | Other | medium | true | null | null | null |
GH1036_FLAML_1405 | fix: Fixed bug where group folds and sample weights couldn't be used together | Other | medium | true | null | null | null |
GH1037_gpytorch_1685 | Kernel batch size recurses through module lists | Other | medium | true | null | null | null |
GH1038_gpytorch_1416 | Make NGD is compatible with batch-mode variational GPs | Other | medium | true | null | null | null |
GH1039_darts_2776 | Fix/categorical validation features | Other | medium | true | null | null | null |
GH103_redis-py_3998 | Redis-Py 3998 | Other | medium | false | null | null | null |
GH1040_darts_2957 | Fix/tft explainer index error | Other | medium | true | null | null | null |
GH1041_matplotlib_31061 | BUG: Fix text appearing far outside valid axis scale range | Other | medium | true | null | null | null |
GH1042_matplotlib_31091 | BUG: Fix IndexLocator | Other | medium | true | null | null | null |
GH1043_plotly.py_5415 | Fix bug where numpy datetime contained in Python list gets converted to integer | Other | medium | true | null | null | null |
GH1044_plotly.py_5193 | Fix issue breaking `fig | Other | medium | true | null | null | null |
GH1045_pytorch_170555 | [cherry-pick] Fix vllm issue for flex (#170499) | Other | medium | true | null | null | null |
GH1046_transformers_43675 | make sure hub errors are surfaced | Other | medium | true | null | null | null |
GH1047_transformers_16669 | Fix example logs repeating themselves | Other | medium | true | null | null | null |
GH1048_keras_22294 | fix(progbar): handle target=0 to prevent crash with empty dataset | Other | medium | true | null | null | null |
GH1049_numpy_31053 | BUG: avoid warning on ufunc with where=True and no output | Other | medium | true | null | null | null |
GH1050_numpy_31045 | BUG: avoid warning on ufunc with where=True and no output | Other | medium | true | null | null | null |
GH1051_scipy_24742 | BUG: stats | Other | medium | true | null | null | null |
GH1052_statsmodels_9524 | BUG: Fix bug in Runs | Other | medium | true | null | null | null |
GH1053_dask_10590 | Use the inferred filesystem/region by default when ``filesystem='arrow'`` | Other | medium | true | null | null | null |
GH1054_dask_9852 | Satisfy `broadcast` argument in `DataFrame | Other | medium | true | null | null | null |
GH1055_mlflow_1227 | Fix download from URI | Other | medium | true | null | null | null |
GH1056_ray_46321 | Ray 46321 | Other | medium | true | null | null | null |
End of preview. Expand in Data Studio
TeamBench: Evaluating Agent Coordination under Enforced Role Separation
Overview
TeamBench is a benchmark of 851 task templates that expand to 931 seeded evaluation instances across 19 base categories (the leaderboard uses 21 refined categories; see paper §3.1). It evaluates whether LLM-based agent teams outperform a single oracle agent under OS-enforced role separation (Planner / Executor / Verifier in isolated sandboxes with distinct tool allow-lists), and reports the Teamwork Necessity Index (TNI): a paired metric that quantifies how much a task requires coordinated multi-agent effort beyond what a capable single agent achieves alone.
Category-distribution caveat. Of the 851 templates, 733 are GitHub-derived (GH*) and are aggregated into a single Other category in the category column; the remaining 118 originally-authored templates split across 18 fine-grained categories (Security, Software Engineering, Distributed Systems, Adversarial, and so on; see the Domain Distribution table below). The "19 base categories" number counts Other + the 18 fine-grained labels; readers analysing category-level distributions should keep this in mind.
The release includes mirrored task artifacts and deterministic shell-script graders, structured ablation results, role-mixing studies, and human-study metadata. The complete seeded workspace generators and evaluation harness are maintained in the accompanying GitHub repository.
Dataset Files
| File | Rows | Description |
|---|---|---|
teambench_dataset.json |
931 | Canonical full release: every seeded instance with task_id, title, category, difficulty, has_generator, ablation_scores (when available; columns oracle/restricted/team/team_no_plan/team_no_verify), tni (= (S_team − S_restricted) / max(0.05, S_solo − S_restricted); null when the necessity gap is below 0.05), and classification (HIGH-TNI/TEAM-HELPS/NEUTRAL/TEAM-HURTS). |
tasks/ |
851 templates | Mirrored task artifacts, including task metadata and deterministic grade.sh graders. Available files vary by task. |
croissant.json |
-- | Dataset metadata (Croissant 1.0 + RAI). |
Quickstart
import json, urllib.request
URL = "https://huggingface.co/datasets/ybkim95/teambench/resolve/main/teambench_dataset.json"
with urllib.request.urlopen(URL) as r:
tasks = json.load(r)
print(f"{len(tasks)} instances")
print(tasks[0])
Or via the datasets library:
from datasets import load_dataset
ds = load_dataset("ybkim95/teambench")
print(ds)
Task artifacts and deterministic graders are mirrored under
tasks/<TASK_ID>/. The complete evaluation harness, seeded workspace
generators, sandbox configuration, and analysis scripts remain available
in the GitHub repository. The structured dataset files in this repository
support programmatic indexing through the Hugging Face datasets library.
Task Origin Mix
| Origin | Templates | Description |
|---|---|---|
| Originally authored | 161 | Critical constraints placed exclusively in the full specification (absent from the brief and the workspace), so a single agent cannot solve the task without the Planner. |
| GitHub bug reports | 650 | Adapted from active open-source repositories (Flask, Click, httpx, Requests, Pydantic, Django, pytest, FastAPI, SQLAlchemy, Celery, Werkzeug, NumPy, SciPy, Keras, spaCy, etc.). Issue text and user-facing symptom go into the brief; the upstream fix patch becomes the deterministic grader. |
| UCI data-science | 30 | Canonical UCI public datasets (cited in the paper). |
| Public post-mortems | 10 | Adapted from public incident-response post-mortems. |
| Total | 851 | Each template has a parameterized generator that emits byte-identical workspaces from a fixed integer seed; held-out seeds are reserved for periodic leaderboard refresh. |
Domain Distribution (Top Categories)
| Category | Tasks | Coordination signal |
|---|---|---|
GitHub-derived (Other) |
733 | Library maintenance bug fixes |
| Security | 32 | Vulnerability patching, audit triage |
| Software Engineering | 31 | Hidden specs, backward compatibility |
| Incident Response | 26 | Cascade failure, memory leak, rollback |
| Operations | 17 | Container debugging, monitoring |
| Data Engineering | 15 | Schema drift, ETL repair |
| Testing | 12 | Spec-to-tests, mutation resistance |
| Policy | 9 | Access control, license compliance |
| Information Retrieval | 8 | Evidence QA, misinformation traps |
| Distributed Systems | 7 | Race conditions, Raft consensus |
| Adversarial | 7 | Spec conflicts, false bug reports, security theater |
| Code Review | 6 | API review, style enforcement |
| Multi-language | 6 | Go concurrency, JavaScript XSS |
| Long-Horizon | 6 | Multi-step migrations, staged deployments |
| Pipeline | 6 | API gateway, message queues |
| Cross-System Integration | 5 | API contract mismatches, schema evolution, auth federation |
| Specification | 3 | Feature implementation from RFC |
| Integration / Negotiation | 2 | Pipeline repair, trade-off configuration |
Five-Condition Ablation
Each core task is evaluated under five conditions. The paper and website use the human-readable names (Solo, Restricted, Full Team, etc.); this dataset and the harness CLI store the machine names (oracle, restricted, etc.). The mapping is:
| Paper / Website name | Dataset / CLI key | Roles | Purpose |
|---|---|---|---|
| Solo | oracle |
Single agent, full tool access | Capability ceiling |
| Restricted | restricted |
Single agent, executor-only tools | Capability floor |
| Full Team | team |
Planner + Executor + Verifier | Full team |
| Team, No Plan | team_no_plan |
Executor + Verifier | Isolates planner contribution |
| Team, No Evaluate | team_no_verify |
Planner + Executor | Isolates verifier contribution |
Scores are in [0, 1] (fraction of grader checks passed).
Ablation columns are populated for the 153 originally-authored core templates and the 120-task hard leaderboard subset. GitHub-derived rows ship with ablation_scores: null because they are scored against upstream patches rather than the 5-condition pipeline.
TNI Metric
The Teamwork Necessity Index measures how much of the Solo-versus-Restricted gap the team recovers:
TNI = (S_team − S_restricted) / max(ε, S_solo − S_restricted), ε = 0.05
S_solo is the oracle column in this dataset. The denominator is the "necessity gap": when Solo barely beats Restricted, TNI is unstable, so we only report it when the gap exceeds ε. TNI = 1 recovers the full single-agent advantage; TNI > 1 exceeds it. This matches Equation 1 of the paper.
Classification bands (from harness/compute_tni.py):
| Band | Rule | Interpretation |
|---|---|---|
HIGH-TNI |
necessity gap > 0.1 and TNI > 0.2 | Coordination is necessary and the team recovers a substantial fraction of the necessity gap |
TEAM-HELPS |
team uplift > +0.05 (vs Restricted) | Team helps |
NEUTRAL |
absolute(team uplift) <= 0.05 | No clear team effect |
TEAM-HURTS |
team uplift < −0.05 (vs Restricted) | Coordination overhead hurts |
where team uplift = S_team − S_restricted. Tasks with ablation_scores: null are unrated.
Headline Findings
- Verifier false-accept rate of 49% on grader-failing runs in the role-mixing pool, with removing the Verifier improving mean partial score in the reference ablation.
- Prompt-only and sandbox-enforced teams reach statistically indistinguishable pass rates, but prompt-only runs produce 3.6 times more cases where the Verifier rewrites the Executor's code.
- Conditional team value: teams help most when single agents struggle (lowest Solo-score quintile, +15.7 points) but hurt on tasks where Solo already performs well.
- Human pilot (40 sessions): Solo participants work through the task directly, Hybrid sessions often collapse into quick approval, and human teams spend more effort coordinating missing information across roles.
Responsible-Use Note
Adversarial-trap (TRAP*) and security-vulnerability (CRYPTO*, SEC*) tasks contain plausible-but-incorrect security patterns by design (intentional nonce reuse, low PBKDF2 iterations, truncated authentication tags, etc.). These are synthetic evaluation cases and must not be deployed. Recommended use is in network-isolated containers, per the responsible-use note in the accompanying paper.
Citation
@article{kim2026teambench,
title={TeamBench: Evaluating Agent Coordination under Enforced Role Separation},
author={Kim, Yubin and Park, Chanwoo and Kim, Taehan and Park, Eugene and Schmidgall, Samuel and Rahman, Salman and Park, Chunjong and Breazeal, Cynthia and Liu, Xin and Palangi, Hamid and others},
journal={arXiv preprint arXiv:2605.07073},
year={2026}
}
License
Released under the MIT License. Tasks adapted from public GitHub issue trackers and UCI datasets retain their respective upstream licenses; only the issue text and patch reference are used to construct deterministic graders.
Links
- Paper: https://arxiv.org/abs/2605.07073
- Code: https://github.com/ybkim95/TeamBench
- Croissant metadata:
croissant.json - Issue tracker: https://github.com/ybkim95/TeamBench/issues
- Downloads last month
- 75