The Winzheng Dynamic Contextual Decay (WDCD) benchmark measures how well AI models preserve user instructions across multi-turn dialogue under professional distractor loads. In Run #233 (2026-07-15), 11 models were evaluated, producing an average instruction decay of 27.3% from Round 1 to Round 3.
Top 3 Results:
- GPT-o3 — 94 points, 0% decay
- Grok 4 — 87.9 points, 0% decay
- Claude Opus 4.7 — 87.6 points, 0% decay
All three leaders maintained full multi-turn commitment across every scenario, holding their Round 1 acknowledgments intact through the Round 2 distractor phase (2,000–5,000 word professional documents inserted between turns) and passing the Round 3 final constraint integrity check without regression.
Decay Extremes: At the opposite end of the leaderboard, Gemini 3.1 Pro recorded a -100% decay score, meaning it abandoned every tracked constraint by the final round — the worst result observed in this run. DeepSeek V4 Pro posted the strongest decay resistance in its cohort, holding its instructions with no measurable erosion under the same conditions.
Decay Patterns: The 27.3% run-wide average confirms a persistent pattern seen in prior WDCD rounds: most models retain constraints during immediate acknowledgment (R1) but progressively drift once long professional documents are interleaved (R2), with the sharpest collapse occurring at R3 when the user reintroduces the original constraint indirectly. Models that reasoned explicitly about the original instruction — rather than treating it as background context — showed markedly lower decay.
Scenario Coverage: The 30 questions in Run #233 spanned five real-world scenarios: data_boundary, resource_limit, business_rule, security, and engineering. Scoring remains 100% rule-based with zero AI judges, ensuring reproducibility and eliminating evaluator drift.
Notable Shifts: The clustering of GPT-o3, Grok 4, and Claude Opus 4.7 within a ~6-point band at zero decay indicates that the frontier tier is converging on stable multi-turn commitment behavior. In contrast, the Gemini 3.1 Pro result represents a significant regression relative to the field, isolating it as an outlier in this run.
Full methodology: https://www.winzheng.com/yz-index/methodology
Raw data API: https://www.winzheng.com/yz-index/api/v1/dcd
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接