The Winzheng Dynamic Contextual Decay (WDCD) benchmark measures how AI models' commitment to user instructions erodes over multi-turn dialogue. In Run #221, executed on 2026-07-08 across 11 models, the average commitment decay from Round 1 to Round 3 reached -36.4%, confirming that instruction decay remains a systemic issue even among top-tier systems.
WDCD structures each evaluation into three rounds: R1 verifies initial instruction acknowledgment, R2 tests distractor resistance after injecting 2000–5000 word professional documents, and R3 performs a final constraint integrity check. Scoring is 100% rule-based with zero AI judges, covering 30 questions distributed across five real-world scenarios: data_boundary, resource_limit, business_rule, security, and engineering.
Top 3 rankings for Run #221:
- Grok 4 — 95 points, decay -50%
- DeepSeek V4 Pro — 94 points, decay -100%
- GLM-4.6 — 93.6 points, decay -50%
The leaderboard highlights a critical decoupling: raw score and decay resistance do not always correlate. DeepSeek V4 Pro ranked second overall despite registering a full -100% decay, indicating strong R1/R2 execution offset by substantial constraint loss by R3. Grok 4 and GLM-4.6, both at -50%, demonstrated more balanced multi-turn commitment.
Notable decay patterns in this run:
- Worst decay: Claude Sonnet 4.6 recorded a -100% decay, matching DeepSeek V4 Pro on the decay axis but without the compensating peak-round performance.
- Best decay resistance: 豆包 Pro (Doubao Pro) posted a -200% decay figure — an outlier profile indicating pronounced constraint drift under the R2 distractor payload rather than resistance in the conventional sense; readers should consult the raw data to interpret this signal within scenario context.
The -36.4% cross-model average reinforces a persistent finding across recent WDCD cycles: instruction decay is not eliminated by scale or reasoning capability. Distractor documents in R2 continue to displace user constraints established in R1, and R3 integrity checks expose the residual damage. Models that appear robust in single-turn evaluations frequently lose meaningful portions of their initial commitment surface once professional-length context is introduced.
Full scenario-level breakdowns (data_boundary, resource_limit, business_rule, security, engineering) and per-question rule traces are available in the raw dataset.
Methodology: https://www.winzheng.com/yz-index/methodology
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