Grok 4 scored 79.30 in today's Smoke evaluation, down from yesterday's 87.66, a drop of 8.4 points.
Score Breakdown and Key Changes
The Code Execution dimension edged down from 94.50 to 93.70, a decrease of 0.8 points. The Material Constraint dimension plummeted from 79.30 to 61.70, a drop of 17.6 points. Engineering Judgment rose from 70.80 to 86.10, an increase of 15.3 points. Task Expression rose from 70.80 to 75.00, an increase of 4.2 points. The Integrity Rating changed from pass to warn.
Possible Cause Analysis
The one-day decline in the Material Constraint dimension far exceeds that in Code Execution, indicating that this fluctuation is primarily concentrated in questions related to material fidelity. The Smoke evaluation consists of only 10 questions per day, 2 per dimension, so differences in sampled questions can directly cause large swings in the Material Constraint dimension score. Code Execution remained relatively stable, suggesting no systemic degradation in the model's fundamental coding ability. Engineering Judgment and Task Expression both rose simultaneously, indicating that the model's secondary capabilities did not weaken in parallel.
The Integrity Rating changed from pass to warn, pointing to at least one instance of clear inconsistency or avoidance behavior in the model's responses. This change coincides with the decline in Material Constraint, suggesting that some questions may have triggered deviations from constraints, rather than merely reflecting changes in question difficulty.
Practical Implications for Users
For scenarios that rely on material constraints—such as long-document summarization, contract clause extraction, and knowledge base Q&A—Grok 4's performance today has fallen below yesterday's baseline. Teams that depend heavily on code execution are less affected, with Code Execution dropping only 0.8 points, so routine script generation and debugging can still maintain a high level. For prototype design tasks that require strong engineering judgment, today's scores are actually higher, making it a viable temporary alternative.
For models with an Integrity Rating of warn, manual review steps should be increased in production environments, especially for outputs that involve external material citations.
Strategic Assessment
This decline in the main score is primarily driven by a single dimension—Material Constraint—while Code Execution and secondary capabilities have not deteriorated in parallel, making it more likely a fluctuation due to question sampling rather than a genuine model regression. However, the simultaneous downgrade in the Integrity Rating constitutes a signal that requires confirmation in the next round. If Material Constraint recovers and the Integrity Rating returns to pass next time, it can be judged as a single sampling anomaly; if Material Constraint remains below 70 and the Integrity Rating stays at warn, the suitability of Grok 4 for material-sensitive scenarios should be reassessed.
Current data does not support a long-term downgrade of the model's overall capability; it only supports short-term attention to the Material Constraint and Integrity metrics.
Data source: YZ Index | Run #224 | View raw data
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接