In AQUALAB gameplay logs, the estimated RLMM person parameter is positively associated with cumulative trajectory reward (Pearson 0.5842, Spearman 0.5760)¶

Claim · All claims
Evidence · 1 study · design r2 · q2 · i3 large
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i3 Episodes with larger estimated β achieved higher cumulative reward, with a printed Pearson correlation of 0.5842 and Spearman correlation of 0.5760. → Wenqian Xu 2026

Evidence¶

Wenqian Xu 2026¶

Wenqian Xu, Feng Ji. (2026). Reinforcement Learning Measurement Model. arXiv. https://arxiv.org/abs/2605.09305 doi:10.48550/arxiv.2605.09305

q2 · i3 · design · r2

Empirical analysis of the February 2026 AQUALAB log export; the filtered analytic sample contained 1,133 episodes and 8,474 transitions, reduced from 106,315 raw transitions. The printed correlation r = 0.5842 (Pearson) is the effect size. The article frames this as "a validity check rather than as evidence thatˆβrepresents a pure ability construct".

"After20block-coordinateMAPiterationsonthefiltereddataset,thecorrelationbetween ˆβandcumulative trajectory reward was 0.5842 by Pearson correlation and 0.5760 by Spearman correlation."

Discussion¶

Related Claims¶

• In AQUALAB, estimated β is ordered across episode completion status (means 1.181, 1.506, 2.343; Kruskal–Wallis p = 8.62×10−29) and higher-β quintiles complete more content more efficiently — related
• In peg-solitaire simulations with known true abilities, the RLMM attains lower RMSE of log βj than the MDP-MM on all four original boards
• The RLMM recovers ability parameters on peg-solitaire boards with up to more than 23 million states, where tabular MDP-MM estimation is impractical
• RLMM runtime stays within approximately 15–20 seconds across the four peg-solitaire boards while MDP-MM runtime grows with task complexity, reaching a 19.1× speedup on the diamond board
• Influence-function diagnostics identify steps 3–8 of the Big-L peg-solitaire board as the most consequential region, with the lowest-β participants showing substantially larger absolute influence values there