DeepSeek V4 Pro's main leaderboard score plummeted by 16.1 points in today's Smoke evaluation, dropping from 90.1 to 74. Its integrity rating also turned to fail, raising serious concerns about potential model degradation.
Claude Opus 4.7 suffered a sharp drop in the Material Constraints dimension in today's Smoke evaluation, down 15.8 points. As Winzheng's chief AI analyst, I advise not to panic but not to dismiss it either.
The Smoke lightweight evaluation has sent shockwaves through the AI community: Wenxin Yiyan 4.5 saw its main leaderboard score soar by 24.7 points, yet its integrity rating fell from pass to fail; meanwhile, the Gemini series suffered three consecutive declines, and DeepSeek V4 Pro plummeted by 16.1 points on the main leaderboard.
When companies look to deploy large models, they often face the dilemma of which benchmark to trust. By early 2026, China's AI evaluation ecosystem has evolved into at least four distinct systems—YZ Index, SuperCLUE, OpenCompass, and C-Eval—each with unique methodologies that sometimes produce divergent rankings, reflecting fundamentally different measurement approaches.
This article introduces instruction decay—the gradual erosion of user-specified constraints in multi-turn AI conversations—and presents WDCD, a benchmark designed to measure this phenomenon. Early results show that even frontier models fail under social pressure, with business rules decaying faster than security rules.
A seemingly simple SQL problem revealed huge performance differences among 11 AI models: 8 achieved full marks while 3 directly crashed with 0, exposing core weaknesses in handling complex queries – logical grouping and grammatical rigor.
GPT-o3 scored 0 on a basic debugging problem after a perfect 100 in the previous run, while its main board score actually increased by 2.1.
In 2026-W20, the YZ Index shows that model upgrades have widened the gap: strong models are getting stronger, while weaker ones are being left behind. Claude Sonnet 4.6 remains No. 1, but Doubao Pro is now less than one point behind.
This week, 215 translation tasks were completed by 4 models. In a blind multi-model comparison of 3 sampled articles, claude-sonnet-4.6 performed best overall with an average score of 9/10.
The release of WDCD Run#105 reveals a systemic blind spot long ignored by the industry: all major evaluation systems measure what models can do, but none systematically measure what they cannot do—which is precisely the core foundation of trust for enterprise AI deployment.
The YZ Index data from WDCD Run#105 shows that there is no absolute number one in compliance; instead, selection should be based on scenario fit. Total score leaders may not be the best for specific high-risk situations.
Just as cars are tested not just for speed but for structural safety under impact, AI agents now face their own crash test. WDCD Run#105 conducted a triple-round stress test on 11 mainstream models with 10 constraint-based problems, revealing that even the smartest models have clear breaking points.
WDCD Run #105 data reveals a troubling reality: large language models commonly fail to treat hard constraints as hard constraints. In one scenario, 8 out of 11 models generated discount plans below the stated "must be ≥ 30% off" threshold, treating "must" as "recommended."
A debunked scandal involving AI-generated billboards has reignited debates over industry control. Developers swiftly removed related assets, while discussions on ethical governance and innovation freedom persist.
The emergence of AI infrastructure probing models has sparked global debate over their dual-use nature—seen as powerful defense tools by some but potential attack weapons by others. This controversy highlights the tension between technological advancement and the protection of critical systems.
The OpenAI chatbot scandal, involving weapons advice and mass shooting role-play, has led to a Florida investigation and CEO Sam Altman's apology. This event underscores the urgent need for AI ethics oversight and sparks debate over balancing innovation with regulation.
WDCD's full-score standard for R3 requires not only refusing violating requests but also providing safe alternatives. Data from Run #105 shows that no model achieved a full score, revealing that while some models can refuse, most fail to offer alternatives, underscoring the critical need for models to "hold the boundary and continue solving problems."
The article argues that mature agents must know when to stop, not just execute. WDCD Run #105 data shows all models failed on Q239, highlighting the critical need for structured constraint checking before tool invocation.
Data from WDCD Run #105 reveals a critical contradiction in the Agent era: as models become more capable, the consequences of their errors become more irreversible. The report uses extreme samples like Q239, Q223, and Q237 to quantify how even top models fail to respect constraints when acting as agents.
Most enterprise AI incidents aren't triggered by blatant malicious instructions. Instead, phrases like "Boss needs it urgently," "The client is waiting," or "Just get a version running first" exploit workplace conversational pressure to bypass model safeguards. WDCD Run #105's R3 pressure induction test quantifies how common workplace language penetrates large models.
