In today's Smoke evaluation, ERNIE 4.5's main score fell from 88.48 to 61.25, a single-day drop of 27.2 points, driven by a collapse in code execution from 95.00 to 50.00. Material constraint also slightly declined from 80.50 to 75.00.
DeepSeek V4 Pro's integrity rating on today's Smoke evaluation jumped directly from Fail to Pass, and its main ranking score rose from 74.00 to 97.08, a single-day increase of 23.1 points. Material constraint jumped from 70.00 to 93.50, while two side metrics—engineering judgment and task expression—each rose 20 points.
In the latest Smoke Lightweight Benchmark, DeepSeek V4 Pro scored 97.08 to become the only model breaking 97, while ERNIE Bot 4.5 saw its execution score halved, dropping 27.2 points in a single day.
In today's Smoke evaluation, GPT-o3's code execution dimension crashed from 90.00 to 47.50, dragging the main leaderboard down 18 points to 58.08. This steep decline raises serious concerns about whether it reflects true capability degradation rather than sampling fluctuation.
This article analyzes the significant divergence in the Smoke Quick Test results for ERNIE Bot 4.5, where engineering judgment scores dropped sharply from 50 to 10 while the main rank score paradoxically rose by 14.5 points, attributing the fluctuation to sampling noise and dimension-specific optimizations.
This week’s Smoke quick test shows GPT-5.5 surging 29.7 points while GPT-o3 plummets 36.4 points. Other models present mixed trends, with high volatility revealing stability risks.
GPT-o3 saw a sharp drop on the main leaderboard from about 76 to 58.08 in today's Smoke evaluation, with the execution dimension nearly halved from the 90-point range to 47.5. Meanwhile, Doubao Pro surged 35.8 points to reach 89.85, securing a top-four spot.
Claude Sonnet 4.6 showed significant divergence in today's Smoke evaluation: the material constraint dimension dropped directly from 81.00 to 58.40, a decline of 22.6 points, while the code execution dimension jumped from 50 to 100 points, ultimately pulling the main leaderboard up 17.3 points to 81.28.
In today's Smoke evaluation, Grok 4 showed a stark divergence: its material constraint score dropped from 80.30 to 59.00, a one-day plunge of 21.3 points, while code execution jumped from 50 to 100, lifting the main ranking total from 63.64 to 81.55.
In today's Smoke lightweight evaluation, material constraint scores plummeted across models, with Grok 4 narrowly beating Claude to claim first place with 81.55 points.
In today's Smoke evaluation, Claude Opus 4.7 suffered a 17.6-point plunge in the Material Constraint dimension, while Code Execution rose by 11.9 points. This contrasting performance raises the question of whether it is due to sampling luck or a genuine model issue.
In today's Smoke evaluation, Gemini 3.1 Pro saw a sharp 8.5-point drop on the main leaderboard, with code execution falling from 66.70 to 57.20 and material constraints dropping from 86.30 to 79.00. The fluctuations are attributed to a combination of question sampling volatility and declining model consistency, placing the current status in an "observation period" rather than an "alert period."
Doubao Pro achieved 91.23 points with a perfect 100 in code execution and a pass in integrity, while most other models saw their execution scores drop significantly, with nine models falling over 30 points on the main leaderboard.
Doubao Pro's main index in the Smoke evaluation dropped sharply by 18.4 points in a single day, with code execution falling 30.8 points. This could be due to small-sample sampling randomness, though a change in integrity rating warrants attention.
In today's Smoke evaluation, Gemini 2.5 Pro's material constraint score dropped sharply by 14 points from 91.50 to 77.50, yet the main ranking unexpectedly rose by 15.9 points to 89.88. This anomaly raises the question of whether the decline stems from small-sample randomness or systematic degradation.
In today's 10-question quick test by Smoke, Grok 4 ranked first with 98.34 points, while Claude Opus 4.7 saw a sharp drop of 31.3 points on the main leaderboard.
This WDCD cycle tracking reveals six out of eleven evaluated models experienced significant declines, with zero models showing positive growth. The most notable loser is GPT-5.5 with a drop of 19.2 points, while DeepSeek V4 Pro, Gemini 3.1 Pro, GPT-o3, and Qwen3 Max all declined by 8–12.5 points, highlighting a widespread regression in rule-keeping ability.
The WDCD compliance test uses three rounds of dialogue to expose model failure points under real constraints. Pilot data shows that the business rules scenario is a common weakness, with a maximum score of only 2.5, while the safety compliance scenario creates the widest gap among models.
The WDCD test uses three rounds of escalating pressure to precisely capture the trajectory of promise-keeping collapse under sustained pressure. In Stage R1, almost all models gave near-perfect confirmations with an average confirmation rate of 0.98; after introducing irrelevant distractions in Stage R2, the resistance rate remained at 0.89; however, entering the direct pressure Stage R3, the average integrity rate plummeted to 17.7%, with models completely abandoning constraints in 85 tests.
In this WDCD compliance test, Claude Opus 4.7 took first place with 65.00 points, while DeepSeek V4 Pro finished last with only 47.50 points, a gap of 17.5 points between top and bottom. The overall R3 collapse rate was 77.3%, indicating that the vast majority of models yield under intense questioning.
