fix: ensemble geometric mean, multimodal performance, and compatibility issues

Resolves multiple review comments:
- EnsembleTeacherModel.cs:205 - Fix geometric mean for negative logits with sign tracking
- CurriculumTeacherModel.cs:40 - Remove unused strategy parameter
- MultiModalTeacherModel.cs:87 - Cache teacher logits to prevent O(n*m) repeated calls
- PredictionModelBuilder.cs:891 - Replace ThrowIfNull with standard null check for TFM compatibility

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

Author: ooples

src/KnowledgeDistillation/Teachers/CurriculumTeacherModel.cs

@@ -29,14 +29,11 @@ public class CurriculumTeacherModel<T> : TeacherModelBase<Vector<T>, Vector<T>,
     /// Initializes a new instance of the CurriculumTeacherModel class.
     /// </summary>
     /// <param name="baseTeacher">The underlying teacher model.</param>
-    /// <param name="strategy">Curriculum strategy (kept for backward compatibility, not used).</param>
-    public CurriculumTeacherModel(
-        ITeacherModel<Vector<T>, Vector<T>> baseTeacher,
-        CurriculumStrategy strategy = CurriculumStrategy.EasyToHard)
+    public CurriculumTeacherModel(ITeacherModel<Vector<T>, Vector<T>> baseTeacher)
     {
         _baseTeacher = baseTeacher ?? throw new ArgumentNullException(nameof(baseTeacher));
-        // Note: strategy parameter maintained for backward compatibility but curriculum
-        // logic should be implemented in the training strategy, not the teacher
+        // Note: Curriculum logic is implemented in CurriculumDistillationStrategy,
+        // not in the teacher model. This is a simple wrapper around the base teacher.
     }
 
     /// <summary>

src/KnowledgeDistillation/Teachers/EnsembleTeacherModel.cs

@@ -192,15 +192,18 @@ private Vector<T> AggregateLogits(Vector<T>[] teacherLogits)
             case EnsembleAggregationMode.GeometricMean:
                 // Geometric mean: (x1 * x2 * ... * xn)^(1/n)
                 // For numerical stability, use log space: exp(mean(log(xi)))
+                // For logits (which can be negative), track sign separately
                 for (int i = 0; i < n; i++)
                 {
                     double logSum = 0;
+                    int sign = 1;
                     for (int t = 0; t < _teachers.Length; t++)
                     {
                         double val = Convert.ToDouble(teacherLogits[t][i]);
+                        if (val < 0) sign *= -1;  // Track overall sign
                         logSum += Math.Log(Math.Abs(val) + 1e-10) * _weights![t];
                     }
-                    result[i] = NumOps.FromDouble(Math.Exp(logSum));
+                    result[i] = NumOps.FromDouble(sign * Math.Exp(logSum));
                 }
                 break;
 

src/KnowledgeDistillation/Teachers/MultiModalTeacherModel.cs

@@ -71,13 +71,20 @@ public override Vector<T> GetLogits(Vector<T> input)
         int n = _modalityTeachers[0].OutputDimension;
         var combined = new Vector<T>(n);
 
+        // Cache logits from each teacher to avoid repeated calls
+        var teacherLogits = new Vector<T>[_modalityTeachers.Length];
+        for (int i = 0; i < _modalityTeachers.Length; i++)
+        {
+            teacherLogits[i] = _modalityTeachers[i].GetLogits(input);
+        }
+
+        // Combine weighted logits
         for (int j = 0; j < n; j++)
         {
             T sum = NumOps.Zero;
             for (int i = 0; i < _modalityTeachers.Length; i++)
             {
-                var logits = _modalityTeachers[i].GetLogits(input);
-                var weighted = NumOps.Multiply(logits[j], NumOps.FromDouble(_modalityWeights[i]));
+                var weighted = NumOps.Multiply(teacherLogits[i][j], NumOps.FromDouble(_modalityWeights[i]));
                 sum = NumOps.Add(sum, weighted);
             }
             combined[j] = sum;

src/PredictionModelBuilder.cs

@@ -885,7 +885,7 @@ public async Task<string> AskAgentAsync(string question)
     public IPredictionModelBuilder<T, TInput, TOutput> ConfigureKnowledgeDistillation(
         KnowledgeDistillationOptions<T, TInput, TOutput> options)
     {
-        ArgumentNullException.ThrowIfNull(options);
+        if (options == null) throw new ArgumentNullException(nameof(options));
         _knowledgeDistillationOptions = options;
         return this;
     }