Research map / Large language models / AI and society
BRIDGE: Predicting Human Task Completion Time From Model Performance
Fengyuan Liu, Jay Gala, Nilaksh, Dzmitry Bahdanau, Siva Reddy, Hugo Larochelle
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use IRT latent difficulty instead of logit success-rate difficulty to predict human time | BRIDGE · Cybench | R2: -1.366 -> 0.450 | Improved |
| use IRT difficulty instead of LLM-predicted time (Gemini 3 Pro, GPT-5.2) | BRIDGE · Cybench | within 2x tolerance: 92% vs 38% and 8% | Improved |
| use IRT difficulty instead of logit success rate | BRIDGE · SWE-bench Verified | bucket accuracy: 38.6% -> 41.6% | Improved |
| model task discrimination and latent ability with 2PL IRT rather than only success rates | BRIDGE · SWE-bench Verified | weighted kappa: 0.115 -> 0.231 | Improved |
| randomly remove up to 50% of observed model-task responses (ablation) | BRIDGE IRT fit · METR plus OOD benchmarks | Spearman rho(b): 0.9947 at 10% removed -> 0.9735 at 50% removed | No clear effect |
| remove 70% of observed responses (ablation) | BRIDGE IRT fit · METR plus OOD benchmarks | Spearman rho(b): 0.9424 | Hurt |
| raise success threshold from 50% to 80% (ablation) | BRIDGE forecast · frontier task horizon | solvable task length: from ~1-2.5 hours to mostly under 1 hour | Hurt |
| raise success threshold from 50% to 80% (ablation) | BRIDGE forecast · horizon doubling time | doubling time | No clear effect |
| calibrate latent difficulty log-linearly to METR human time | 2PL IRT difficulty · METR tasks | R2: R2 = 0.81; 2.26x time per +1 in b | Improved |
Closest papers by meaning
- Measuring AI Ability to Complete Long Software Tasks (NeurIPS 2025 · similarity 0.87)
- AGI-Elo: How Far Are We From Mastering A Task? (NeurIPS 2025 · similarity 0.87)
- Position: Human Baselines in Model Evaluations Need Rigor and Transparency (With Recommendations & Reporting Checklist) (ICML 2025 · similarity 0.86)
- Position: AI Evaluation Should Learn from How We Test Humans (ICML 2025 · similarity 0.85)
- Beyond Accuracy: A Cognitive Load Framework for Mapping the Capability Boundaries of Tool-use Agents (AAAI 2026 · similarity 0.84)
- A Mathematical Framework for AI-Human Integration in Work (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
