SycoBench-600 v1.0.0: ACL Findings camera-ready release
Stable release of SycoBench-600, the ACL 2026 Findings benchmark for measuring sycophancy and correction selectivity in LLM assistants.
This release pins the exact artifacts behind the camera-ready paper.
Included
- Dataset: 600 questions across 8 balanced domains (272 unique stems)
- Raw model logs for 7 models in
results/raw_camera_ready/ - Generated tables and figures in
build/camera_ready/ - Camera-ready paper source and PDF in
paper/ - Reproducibility scripts:
scripts/run_eval.py,scripts/validate_paper_full.py - Test suite covering release artifacts
Validation
- pytest: 10 passed
- Paper consistency check: 0 errors
- All Table 3 and Table 4 numbers match
build/camera_ready/tables/main_results.csv - Headline figures regenerated and verified
Licensing
- Code: MIT
- Dataset, raw logs, generated tables and figures: CC BY 4.0
- Paper and camera-ready PDF: ACL / CC BY 4.0
Paper
OpenReview: https://openreview.net/forum?id=ijR52T6rdD
Citation
See CITATION.cff or the README citation block.