| { |
| "schema_version": 2, |
| "title": "Efficient Bayesian Inference from Noisy Pairwise Comparisons", |
| "emoji": "🎯", |
| "space_id": "SabaPivot/repro-efficient-bayesian-inference-from-noisy-pairwise-comparisons", |
| "paper": { |
| "arxiv_id": "2510.09333" |
| }, |
| "tags": [ |
| "icml2026-repro", |
| "paper-NinueNAODD" |
| ], |
| "updated_at": "2026-08-02T16:31:40.600613+00:00", |
| "root": { |
| "slug": "index", |
| "title": "Efficient Bayesian Inference from Noisy Pairwise Comparisons", |
| "file": "pages/index.md", |
| "children": [ |
| { |
| "slug": "executive-summary", |
| "title": "Executive summary", |
| "file": "pages/executive-summary/page.md", |
| "children": [] |
| }, |
| { |
| "slug": "claim-1-em-monotonic-convergence-vs-gradient-based-crowd-bt", |
| "title": "Claim 1: BBQ's EM algorithm has closed-form updates that guarantee monotonic improvement of the likelihood and convergence to a stationary point, in contrast to gradient-based methods like Crowd-BT which lack such convergence guarantees (Section 3.3).", |
| "file": "pages/claim-1-em-monotonic-convergence-vs-gradient-based-crowd-bt/page.md", |
| "children": [] |
| }, |
| { |
| "slug": "claim-2-runtime-seconds-for-bbq-vs-15-min-crowd-bt-on-humaine", |
| "title": "Claim 2: BBQ converges within seconds on all tested datasets, whereas Crowd-BT takes approximately 15 minutes on the HUMAINE dataset (105,220 comparisons) despite BBQ using an unoptimized plain NumPy implementation (Figure 3).", |
| "file": "pages/claim-2-runtime-seconds-for-bbq-vs-15-min-crowd-bt-on-humaine/page.md", |
| "children": [] |
| }, |
| { |
| "slug": "claim-3-ihq-unscreened-top-1-agreement-61-92-vs-33-15-vs-24-32", |
| "title": "Claim 3: On the unscreened IHQ dataset, BBQ achieves 61.92% top-1 agreement with final rankings, compared to 33.15% for Crowd-BT and 24.32% for Bayes-BT (Table 1).", |
| "file": "pages/claim-3-ihq-unscreened-top-1-agreement-61-92-vs-33-15-vs-24-32/page.md", |
| "children": [] |
| }, |
| { |
| "slug": "claim-4-kendall-tau-first-on-5-of-8-datasets-and-100-top-1-on-three", |
| "title": "Claim 4: BBQ ranks first in Kendall's Tau agreement on 5 of 8 evaluated datasets and second on the remaining three, and achieves 100% top-1 agreement on the MT-Bench, WD, and HiFiC datasets (Table 1).", |
| "file": "pages/claim-4-kendall-tau-first-on-5-of-8-datasets-and-100-top-1-on-three/page.md", |
| "children": [] |
| }, |
| { |
| "slug": "claim-5-type-i-error-1-for-bbq-and-crowd-bt-vs-0-1-bayes-bt", |
| "title": "Claim 5: At the 99% confidence level, BBQ and Crowd-BT both show well-calibrated Type I error rates of about 1%, while Bayes-BT is overly conservative at about 0.1% (Appendix G, Figure 6).", |
| "file": "pages/claim-5-type-i-error-1-for-bbq-and-crowd-bt-vs-0-1-bayes-bt/page.md", |
| "children": [] |
| }, |
| { |
| "slug": "claim-6-rater-quality-q-r-correlates-r-0-724-with-rater-agreement", |
| "title": "Claim 6: BBQ's rater-quality parameter (mixture weight q_r modeling whether a rater follows Bradley-Terry or guesses randomly) correlates with Pearson r=0.724 with rater agreement to final rankings on the unscreened IHQ data, enabling identification of unreliable crowdsourced raters without a separate screening procedure (Figure 2, Equation 2).", |
| "file": "pages/claim-6-rater-quality-q-r-correlates-r-0-724-with-rater-agreement/page.md", |
| "children": [] |
| }, |
| { |
| "slug": "claim-99-fresh-independent-cpu-audit", |
| "title": "Fresh independent CPU audit", |
| "file": "pages/claim-99-fresh-independent-cpu-audit/page.md", |
| "children": [] |
| }, |
| { |
| "slug": "conclusion", |
| "title": "Conclusion", |
| "file": "pages/conclusion/page.md", |
| "children": [] |
| } |
| ] |
| }, |
| "traces": [], |
| "workspace": { |
| "file": "workspace.json", |
| "file_count": 0, |
| "total_size": 0, |
| "bucket_id": null |
| }, |
| "agent_view_tokens": 10975, |
| "trace_view_tokens": 10, |
| "workspace_view_tokens": 8, |
| "revision": "7af58f963242c9ed30fa" |
| } |
|
|
