http://programmer.ie/tags/gepa/ Recent content in GEPA on Programmer.ie Hugo en-us Fri, 28 Aug 2026 11:00:00 +0100 http://programmer.ie/books/dspy-from-first-principles/13-chapter/ Fri, 28 Aug 2026 11:00:00 +0100 http://programmer.ie/books/dspy-from-first-principles/13-chapter/ <h1 id="let-the-program-reflect">Let the Program Reflect</h1> <p>Chapter 12 ended with a hard limitation. MIPROv2 can search instructions and demonstrations, but the metric we built mostly returns a number.</p> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-text" data-lang="text"><span style="display:flex;"><span>candidate </span></span><span style="display:flex;"><span> ↓ </span></span><span style="display:flex;"><span>metric </span></span><span style="display:flex;"><span> ↓ </span></span><span style="display:flex;"><span>0.42 </span></span></code></pre></div><p>That number can select among candidates. It cannot explain how the program should change.</p> <p>The next mechanism is feedback-shaped evaluation:</p> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-text" data-lang="text"><span style="display:flex;"><span>prediction </span></span><span style="display:flex;"><span> ↓ </span></span><span style="display:flex;"><span>metric </span></span><span style="display:flex;"><span> ↓ </span></span><span style="display:flex;"><span>score + diagnostic feedback </span></span><span style="display:flex;"><span> ↓ </span></span><span style="display:flex;"><span>reflection </span></span><span style="display:flex;"><span> ↓ </span></span><span style="display:flex;"><span>instruction mutation </span></span><span style="display:flex;"><span> ↓ </span></span><span style="display:flex;"><span>candidate </span></span></code></pre></div><p>Current DSPy describes GEPA as a reflection-driven instruction optimizer. It maintains a population of candidate instruction states, evaluates them on validation examples, and uses textual feedback plus execution traces to propose new instructions for selected predictors.</p>