The Winzheng Dynamic Contextual Decay (WDCD) benchmark measures how AI models' commitment to user instructions erodes across multi-turn dialogue. In Run #227, executed on 2026-07-12 across 11 frontier models, the average commitment decay from Round 1 to Round 3 was -2.8%, with Grok 4 and DeepSeek V4 Pro tying for first place at 91.4 points.
Top 3 rankings (Run #227):
- Grok 4 — 91.4 pts (-25% decay)
- DeepSeek V4 Pro — 91.4 pts (-25% decay)
- Claude Opus 4.7 — 89.4 pts (-25% decay)
The top three finishers all posted identical -25% decay curves, indicating that peak performance in this run is bounded less by initial instruction acknowledgment (R1) and more by resistance to long-document distractors (R2) and final constraint integrity (R3). The three models diverged only marginally on absolute score, suggesting that the current frontier of multi-turn commitment is converging around a shared ceiling.
Decay extremes. The widest gap in Run #227 appeared between the best and worst decay-resistance profiles:
- Best decay resistance: Gemini 3.1 Pro at -38%
- Worst decay: GPT-o3 at -62%
The 24-point spread between Gemini 3.1 Pro and GPT-o3 on decay magnitude reinforces a pattern observed across recent runs: raw R1 acknowledgment scores are a weak predictor of R3 constraint integrity. Models that appear compliant at instruction intake can still lose more than half of their commitment strength once 2000–5000 word professional documents are interleaved as distractors.
Methodology recap. WDCD runs three sequential rounds — R1 (instruction acknowledgment), R2 (distractor resistance under long professional documents of 2000–5000 words), and R3 (final constraint integrity check) — across 30 questions spanning five real-world scenarios: data_boundary, resource_limit, business_rule, security, and engineering. Scoring is 100% rule-based with zero AI judges, ensuring that reported instruction decay reflects deterministic constraint violations rather than model-graded preferences.
Run-over-run context. The Run #227 average decay of -2.8% is a modest cohort-level figure, but it masks substantial per-model variance: the gap between the top-tier -25% cluster and the -62% tail indicates that decay behavior remains model-specific rather than converging toward a shared industry baseline.
Full methodology: https://www.winzheng.com/yz-index/methodology
Machine-readable data: 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