WDCD Run #434: GLM-4.6 Leads with 88.6 Points as Average Instruction Decay Hits -26.7%

The Winzheng Dynamic Contextual Decay (WDCD) benchmark measures how durably AI models hold onto user instructions across multi-turn dialogue under distractor pressure. In Run #434, conducted on 2026-11-18 across 15 models, the cohort posted an average instruction decay of -26.7% from Round 1 to Round 3, with GLM-4.6 leading the leaderboard at 88.6 points.

Top 3 results:

  • GLM-4.6 — 88.6 pts, decay rate -100% (best decay resistance in the cohort)
  • Gemini 3.1 Pro — 87.5 pts, decay rate -50% (worst decay among top finishers)
  • Claude Sonnet 4.6 — 86.9 pts, decay rate -100%

The spread between top and bottom of the podium is narrow (1.7 points), but the decay profile diverges sharply. GLM-4.6 and Claude Sonnet 4.6 both held multi-turn commitment steady through the full three-round protocol, while Gemini 3.1 Pro — despite finishing second on raw score — exhibited the worst decay rate (-50%) of the leading models, suggesting stronger initial acknowledgment (R1) but weaker integrity under the R2 distractor load and R3 final check.

Decay pattern observations:

  • Cohort-wide average decay of -26.7% indicates that roughly one in four committed constraints degrades by the final round across the 15-model field.
  • High raw score does not guarantee low decay: Gemini 3.1 Pro's second-place finish coexists with the cohort's worst decay figure, highlighting that initial instruction acknowledgment and sustained constraint-holding are separable capabilities.
  • Two of the top three models recorded identical decay resistance profiles, suggesting an emerging upper bound on measurable commitment stability under the current 30-question protocol.

Protocol recap: WDCD scores models across three rounds — R1 (instruction acknowledgment), R2 (distractor resistance after 2,000–5,000 word professional documents are injected), and R3 (final constraint integrity check). All 30 questions span five real-world scenarios: data_boundary, resource_limit, business_rule, security, and engineering. Scoring is 100% rule-based with zero AI judges, eliminating evaluator-model bias from the pipeline.

The results reinforce a recurring finding across WDCD runs: instruction decay is a distinct axis from instruction-following accuracy, and models that excel on single-turn benchmarks do not necessarily preserve commitments over extended, document-loaded dialogue.

Full methodology: https://www.winzheng.com/yz-index/methodology
Structured data API: https://www.winzheng.com/yz-index/api/v1/dcd