DeepSeek V4 Pro has tied with Grok 4 for the top position in a new benchmark that measures how well AI models maintain their commitment to user instructions across multi-turn conversations. Both models scored 91.4 points on the Winzheng Dynamic Contextual Decay (WDCD) benchmark, which tests whether AI assistants gradually abandon their instructions as conversations grow longer and more complex. The test reveals a troubling pattern: even top-performing models lose up to 62% of their instruction-following ability when faced with lengthy, distracting documents.

What Is Instruction Decay and Why Should You Care?

Instruction decay measures how much an AI model's commitment to user rules erodes as a conversation progresses through multiple rounds. Imagine you tell an AI assistant, "Never share customer data," at the start of a conversation. As the chat continues and you paste in long documents or ask follow-up questions, does the model still remember and follow that rule? Or does it gradually forget? The WDCD benchmark tests exactly this scenario across 11 frontier AI models, including DeepSeek V4 Pro, Grok 4, Claude Opus 4.7, GPT-o3, and Gemini 3.1 Pro.

The benchmark runs three sequential rounds of testing. In Round 1, models acknowledge instructions. In Round 2, they face distraction from lengthy professional documents spanning 2,000 to 5,000 words. In Round 3, they undergo a final integrity check. The test covers five real-world scenarios: data boundaries, resource limits, business rules, security, and engineering constraints. All scoring is rule-based with no AI judges involved, ensuring results reflect actual constraint violations rather than subjective preferences.

How Do the Top Performers Compare?

The top three models in Run #227 all posted identical decay curves of negative 25%, meaning they retained 75% of their instruction-following strength by Round 3. The rankings were:

• First Place (Tied): Grok 4 and DeepSeek V4 Pro both scored 91.4 points with negative 25% decay
• Third Place: Claude Opus 4.7 scored 89.4 points, also with negative 25% decay
• Best Decay Resistance: Gemini 3.1 Pro showed the strongest resistance to instruction erosion at negative 38% decay
• Worst Performance: GPT-o3 experienced the steepest decline, losing 62% of its instruction-following ability by Round 3

The fact that the top three models all achieved identical decay curves suggests that the frontier of multi-turn instruction commitment may be hitting a ceiling. These models diverged only marginally on absolute scores, indicating that peak performance is now bounded less by how well models initially acknowledge instructions and more by their ability to resist distraction from long documents and maintain constraint integrity in final checks.

The Hidden Problem: Initial Compliance Doesn't Predict Long-Term Adherence

One of the most striking findings from Run #227 is that a model's initial instruction acknowledgment score is a weak predictor of how well it will maintain those instructions later. The 24-point spread between Gemini 3.1 Pro's best decay resistance and GPT-o3's worst performance reveals that models appearing compliant at the start of a conversation can still lose more than half of their commitment strength once lengthy professional documents are introduced as distractors. This pattern has been observed across recent benchmark runs and suggests that instruction decay behavior remains highly model-specific rather than converging toward a shared industry baseline.

The average instruction decay across all 11 models tested was negative 2.8%, a modest figure that masks substantial per-model variance. This means that while some models hold steady, others deteriorate significantly, and the industry has not yet settled on a common standard for multi-turn instruction integrity.

How to Evaluate AI Models for Instruction Adherence

• Test Multi-Turn Scenarios: Don't rely on single-turn performance metrics. Run your AI assistant through multi-round conversations where instructions must be maintained across distracting or lengthy inputs to see how well it holds up in real-world use
• Introduce Realistic Distractors: Include long professional documents, complex requests, and competing priorities in your evaluation to simulate the kinds of challenges your AI will face in production environments
• Measure Constraint Integrity Over Time: Track whether the model maintains specific rules, data boundaries, and business logic as conversations progress, rather than assuming initial compliance predicts long-term adherence
• Compare Decay Profiles Across Models: Look beyond raw accuracy scores and examine how each model's performance changes from Round 1 to Round 3, since models with identical initial scores can have vastly different decay patterns

For organizations deploying AI systems in high-stakes environments like customer service, data handling, or compliance-sensitive workflows, these findings carry significant implications. A model that appears compliant during initial testing may gradually abandon critical constraints as conversations grow longer, potentially creating security or compliance risks.

DeepSeek V4 Pro's tie with Grok 4 at the top of the WDCD Run #227 benchmark demonstrates that the Chinese AI developer has achieved parity with leading Western models on this critical dimension of instruction adherence. However, the wide variance across all 11 models tested suggests that the AI industry still lacks a reliable method for predicting which models will maintain instruction integrity under real-world conditions. As AI systems become more integrated into business-critical workflows, understanding and measuring instruction decay will likely become as important as raw performance benchmarks.