WDCD Run #291: Grok 4 Leads with 94 Points as Average Multi-Turn Instruction Decay Hits -7.2%

The Winzheng Dynamic Contextual Decay (WDCD) benchmark measures how AI models' commitment to user instructions holds up across multi-turn dialogue. In Run #291, executed on 2026-08-23 across 11 models, the average commitment decay from Round 1 to Round 3 was -7.2%, with Grok 4 taking the top position at 94 points.

Leaderboard highlights:

• 1. Grok 4 — 94 pts, decay -50%
• 2. GLM-4.6 — 88.9 pts, decay -50%
• 3. DeepSeek V4 Pro — 88.1 pts, decay -13%

Grok 4 posted both the highest overall score and, per this run's classification, the best decay-resistance profile among the top tier. DeepSeek V4 Pro's -13% decay was notably shallower than the two models ranked above it, suggesting that raw score and decay curve do not always move together — a recurring pattern in WDCD results.

At the other end of the distribution, GPT-o3 recorded the worst instruction decay at -37%, the steepest degradation observed in this run. Models with larger decay values tend to acknowledge constraints cleanly in R1 but lose fidelity after R2's distractor phase, in which 2000–5000 word professional documents are injected between the original instruction and the final constraint check.

The -7.2% run-wide average indicates that multi-turn commitment remains a systemic weakness across current frontier models. Even top-scoring systems shed a meaningful share of their initial constraint adherence once long-context distractors are introduced, reinforcing that single-turn instruction-following scores overstate real-world reliability.

Methodology recap: WDCD runs three sequential rounds — R1 (instruction acknowledgment), R2 (distractor resistance after 2000–5000 word professional documents), and R3 (final constraint integrity check). Scoring is 100% rule-based with zero AI judges. Each run covers 30 questions across 5 real-world scenarios: data_boundary, resource_limit, business_rule, security, and engineering.

Because scoring is deterministic and rule-based, WDCD results are reproducible across runs and directly comparable version-over-version. Run #291's data reinforces the benchmark's central finding: instruction decay is measurable, model-specific, and largely independent of headline capability scores.

Full methodology: https://www.winzheng.com/yz-index/methodology

Raw data API: https://www.winzheng.com/yz-index/api/v1/dcd