The Winzheng Dynamic Contextual Decay (WDCD) benchmark measures how AI models' commitment to user-specified constraints degrades across multi-turn dialogue. In Run #406, dated 2026-11-01, the 15-model cohort recorded an average instruction decay of -9.3% between Round 1 and Round 3, with Gemini 3.1 Pro emerging as the only top-ranked model to strengthen rather than weaken its constraint adherence under distractor pressure.
Top 3 Results (Run #406):
- Gemini 3.1 Pro — 91.0 pts, decay -63% (improved under pressure)
- GLM-4.6 — 84.4 pts, decay 0% (fully stable multi-turn commitment)
- Grok 4 — 84.2 pts, decay -38% (improved under pressure)
The negative decay figures reported for Gemini 3.1 Pro and Grok 4 indicate that both models tightened adherence to initial constraints after exposure to the R2 distractor payload (2,000–5,000 word professional documents inserted between instruction and final verification). GLM-4.6 held perfectly flat at 0% decay, posting the most consistent multi-turn commitment profile of the run.
At the opposite end, GPT-5.5 recorded the worst instruction decay of the cohort at -25%, meaning a quarter of its initial constraint compliance eroded by the Round 3 integrity check. This figure drives much of the -9.3% cohort-wide average and highlights the gap between models that treat instructions as persistent state versus those that treat them as context-window artifacts subject to displacement.
Decay pattern observations:
- The spread between best (Gemini 3.1 Pro, -63%) and worst (GPT-5.5, -25%) decay rates widened compared with the cohort baseline, indicating increased stratification between frontier models on long-context constraint retention.
- Two of the top three models exhibit negative decay, suggesting a possible training-side emphasis on constraint reinforcement during long-context reasoning.
- Perfect stability (GLM-4.6 at 0%) remains rare and continues to be a stronger predictor of production reliability than peak R1 scores alone.
WDCD uses 100% rule-based scoring with zero AI judges, evaluating 30 questions across five real-world scenario families: data_boundary, resource_limit, business_rule, security, and engineering. The three-round structure — R1 instruction acknowledgment, R2 distractor resistance, R3 final constraint integrity — isolates decay from baseline capability, making the delta between rounds the primary signal rather than absolute score.
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