研究报告 - Winzheng Research Lab

WDCD Run #185: Average Instruction Decay Hits -57.5% Across 11 Models, Qwen3 Max Leads at 92.5 Points

WDCD Run #185 (2026-06-17) measured multi-turn commitment across 11 models, recording an average instruction decay of -57.5% from Round 1 to Round 3. Qwen3 Max topped the run at 92.5 points, while 文心一言 4.5 showed the strongest decay resistance.

Research Lab

5大模型翻译对决：第25周质量评测，passthrough 以 9 分领跑

本周共翻译 443 篇文章，覆盖 5 个AI模型。经抽样盲评，passthrough 综合得分最高（9/10）。报告详细对比各模型在准确性、流畅性、术语一致性方面的表现差异。

Research Lab

WDCD Run #171: Average Instruction Decay Hits -37.9% Across 11 Models, Qwen3 Max Leads Despite Steep Drop

WDCD Run #171 (2026-06-14) measured multi-turn commitment across 11 frontier models, recording an average instruction decay of -37.9% from Round 1 to Round 3. Qwen3 Max topped the ranking at 84.4 points despite a -59% decay curve.

Research Lab

WDCD Run #169: Grok 4 Leads Multi-Turn Commitment Test as Average Instruction Decay Drops to 4.5%

WDCD Run #169 (2026-06-13) evaluated 11 AI models on multi-turn commitment integrity, with Grok 4 topping the leaderboard at 74.2 points and the cross-model average instruction decay measured at 4.5% from Round 1 to Round 3.

Research Lab

Instruction Decay Measured: LLM Compliance Falls from 95.8% to 68.3% Under Three Rounds of Pressure

In WDCD Run #164 (June 11, 2026), 11 frontier LLMs acknowledged user constraints 95.8% of the time, but only 68.3% still honored those constraints after distraction and direct social-engineering pressure. 73 of 330 tests (22.1%) ended in complete integrity collapse. General capability did not predict commitment: Claude Opus 4.7, ranked #2 on the capability leaderboard, finished second-to-last on commitment keeping.

Research Lab

WDCD Run #164: Average Instruction Decay Hits -44.3% Across 11 Frontier Models

WDCD Run #164 (2026-06-11) evaluated 11 frontier models across three dialogue rounds, recording an average commitment decay of -44.3% from R1 to R3. GPT-5.5 led the leaderboard at 88.3 points, while 豆包 Pro showed the strongest decay resistance.

Research Lab

WDCD Run #161: Average Instruction Decay Hits -48.6% Across 11 Models, GPT-5.5 Leads at 89.2 Points

WDCD Run #161 (2026-06-11) evaluated 11 large language models on multi-turn commitment integrity, recording an average instruction decay of -48.6% from Round 1 to Round 3. GPT-5.5 led the ranking with 89.2 points, while 豆包 Pro showed the strongest decay resistance.

Research Lab

WDCD Run #157: Average Instruction Decay Hits 47.7% Across 11 Models, Three-Way Tie at the Top

WDCD Run #157 (2026-06-10) recorded a 47.7% average commitment decay across 11 models, with Claude Sonnet 4.6, Gemini 2.5 Pro, and Qwen3 Max tying for first at 67.5 points.

Research Lab

3大模型翻译对决：第24周质量评测，passthrough 以 9 分领跑

本周共翻译 2425 篇文章，覆盖 3 个AI模型。经抽样盲评，passthrough 综合得分最高（9/10）。报告详细对比各模型在准确性、流畅性、术语一致性方面的表现差异。

Research Lab

WDCD Run #146: Average Instruction Decay Hits 24.7% Across 11 Models, Claude Opus 4.7 and GPT-5.5 Tie at Top

WDCD Run #146 (2026-06-03) tested 11 frontier models on multi-turn commitment integrity, recording an average instruction decay of 24.7% from Round 1 to Round 3. Claude Opus 4.7, GPT-5.5, and GPT-o3 tied for first place at 70 points with only -10% decay each.

Research Lab

3大模型翻译对决：第23周质量评测，gpt-o3 以 9 分领跑

本周共翻译 270 篇文章，覆盖 3 个AI模型。经抽样盲评，gpt-o3 综合得分最高（9/10）。报告详细对比各模型在准确性、流畅性、术语一致性方面的表现差异。

Research Lab

WDCD Run #140: Qwen3 Max Leads with 17% Instruction Decay as Average Hits 36.5%

WDCD Run #140 (2026-05-31) evaluated 11 frontier models on multi-turn commitment integrity, finding an average instruction decay of 36.5% from Round 1 to Round 3. Qwen3 Max led the field with 70.8 points and the lowest decay rate at 17%.