liqiang
b119135836
Build latest book artifacts / build (push) Canceled after 0s
dependency resolution / resolve (3.11) (push) Canceled after 0s
dependency resolution / resolve (3.13) (push) Canceled after 0s
deploy-pages / build (push) Canceled after 0s
deploy-pages / deploy (push) Canceled after 0s
i18n consistency check / check (push) Canceled after 0s
provider adoption tests / test (chapter2/context-compression) (push) Canceled after 0s
provider adoption tests / test (chapter2/prompt-injection) (push) Canceled after 0s
provider adoption tests / test (chapter2/system-hint) (push) Canceled after 0s
provider adoption tests / test (chapter3/log-sanitization) (push) Canceled after 0s
web-search-agent tests / test (push) Canceled after 0s
web-search-agent tests / agentbook (push) Canceled after 0s
51 lines
10 KiB
XML
51 lines
10 KiB
XML
<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 40 780 490" width="780" height="490" style="background:#ffffff">
|
||
<defs><marker id="ah" markerWidth="12" markerHeight="8" refX="12" refY="4" orient="auto"><polygon points="0 0, 12 4, 0 8" fill="#333333"/></marker><marker id="ah-light" markerWidth="12" markerHeight="8" refX="12" refY="4" orient="auto"><polygon points="0 0, 12 4, 0 8" fill="#999999"/></marker></defs>
|
||
|
||
<rect x="30" y="55" width="340" height="165" rx="8" fill="#ffffff" stroke="#333333" stroke-width="2" stroke-dasharray="8,4"/>
|
||
<text x="42" y="73" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="16" fill="#666666" text-anchor="start" dominant-baseline="central" font-weight="bold">GP-L (Plain Text)</text>
|
||
<rect x="45" y="82" width="150" height="50" rx="6" fill="#f0f0f0" stroke="#333333" stroke-width="2"/>
|
||
<text x="120" y="100" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="16" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">Training Set</text>
|
||
<text x="120" y="116" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">J/Q/K = 10</text>
|
||
<line x1="197" y1="107" x2="218" y2="107" stroke="#333333" stroke-width="2" marker-end="url(#ah)"/>
|
||
<rect x="220" y="82" width="135" height="50" rx="4" fill="#f5f5f5" stroke="#333333" stroke-width="2"/>
|
||
<text x="287" y="100" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="16" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">Test Set (OOD)</text>
|
||
<text x="287" y="116" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">J=11, Q=12, K=13</text>
|
||
<text x="195" y="155" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">Input: "Use [7, J, 3, 5] to compute 24"</text>
|
||
<text x="195" y="175" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">Output: "(J-7)×(5+3) = 3×8 = 24"</text>
|
||
<text x="195" y="195" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">* OOD when J=11: "(11-7)×(5+3)=32≠24"</text>
|
||
<rect x="410" y="55" width="340" height="165" rx="8" fill="#ffffff" stroke="#333333" stroke-width="2" stroke-dasharray="8,4"/>
|
||
<text x="422" y="73" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="16" fill="#666666" text-anchor="start" dominant-baseline="central" font-weight="bold">GP-VL (Vision-Language)</text>
|
||
<rect x="425" y="82" width="150" height="50" rx="6" fill="#f0f0f0" stroke="#333333" stroke-width="2"/>
|
||
<text x="500" y="100" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="16" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">Training Set</text>
|
||
<text x="500" y="116" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">♠♣ Black Suits</text>
|
||
<line x1="577" y1="107" x2="598" y2="107" stroke="#333333" stroke-width="2" marker-end="url(#ah)"/>
|
||
<rect x="600" y="82" width="135" height="50" rx="4" fill="#f5f5f5" stroke="#333333" stroke-width="2"/>
|
||
<text x="667" y="100" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="16" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">Test Set (OOD)</text>
|
||
<text x="667" y="116" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">♥♦ Red Suits</text>
|
||
<text x="575" y="155" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">Input: Playing card image</text>
|
||
<text x="575" y="175" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">Visual Encoder → Digit Recognition → Reasoning</text>
|
||
<text x="575" y="195" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="12.5" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">* OOD: Suit appearance variation tests visual robustness</text>
|
||
<text x="390" y="245" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="20" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">Llama-3.2-Vision-11B</text>
|
||
<rect x="40" y="265" width="320" height="50" rx="6" fill="#f0f0f0" stroke="#333333" stroke-width="2"/>
|
||
<text x="200" y="282" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="16" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">SFT Path</text>
|
||
<text x="200" y="302" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">max Σ log P(y|x) | Memorize Training Distribution</text>
|
||
<rect x="420" y="265" width="320" height="50" rx="6" fill="#d0d0d0" stroke="#333333" stroke-width="2"/>
|
||
<text x="580" y="282" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="16" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">RL Path (PPO)</text>
|
||
<text x="580" y="302" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">max E[R(τ)] | Explore Generalization Strategies</text>
|
||
<rect x="30" y="340" width="720" height="160" rx="4" fill="#f5f5f5" stroke="#333333" stroke-width="2"/>
|
||
<text x="390" y="362" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="20" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">OOD Test Results (Relative to SFT Init Baseline)</text>
|
||
<text x="100" y="390" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold"></text>
|
||
<text x="245" y="390" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">Rule OOD (GP-L)</text>
|
||
<text x="415" y="390" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">Rule OOD (GP-VL)</text>
|
||
<text x="590" y="390" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">Visual OOD (GP-VL)</text>
|
||
<text x="100" y="415" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="normal">SFT Extension</text>
|
||
<text x="245" y="415" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">−8.1% (11.5→3.4)</text>
|
||
<text x="415" y="415" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">−5.6%</text>
|
||
<text x="590" y="415" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">−9.9% (23.6→13.7)</text>
|
||
<text x="100" y="440" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="normal">RL Extension</text>
|
||
<text x="245" y="440" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">+3.5% (11.5→15.0)</text>
|
||
<text x="415" y="440" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">+3.0%</text>
|
||
<text x="590" y="440" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">+17.6% (23.6→41.2)</text>
|
||
<text x="390" y="470" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="12.5" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">Conclusion: SFT consistently decreases out-of-distribution, RL consistently improves — "SFT memorizes, RL generalizes"</text>
|
||
<text x="390" y="490" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">Key premise: End-to-end RL without SFT completely fails (output is unstructured, reward cannot be computed)</text>
|
||
</svg> |
