feat: [GDPval-AA v2 Updates 1 / n] - GDPval Multi-Reference Model Support (#1663)

Adds multi-reference model support and Bradley-Terry MLE ELO computation
with multiple reference models in order to directly match AA's ELO score
calcualtion algorithm: [GDPval-AA
v2](https://artificialanalysis.ai/methodology/intelligence-benchmarking#gdpval-aa)

> After collecting pairwise rankings, we fit them to a Bradley-Terry
model via maximum likelihood estimation and compute confidence intervals
using the sandwich estimator to establish our final Elo metric

This is one of many changes needed to match AA's v2 benchmark.

---------

Signed-off-by: Alex Gronskiy <agronskiy@nvidia.com>
Co-authored-by: Virginia Adams <vadams@oci-hsg-cs-001-vscode-02.cm.cluster>
Co-authored-by: Alex Gronskiy <agronskiy@nvidia.com>

Author: 3 people

benchmarks/gdpval/config.yaml

@@ -10,6 +10,15 @@
 # Comparison mode (pairwise ELO vs a reference model's deliverables):
 #   ++gdpval_resources_server.resources_servers.gdpval.reward_mode=comparison \
 #   ++gdpval_resources_server.resources_servers.gdpval.reference_deliverables_dir=/path/to/fork
+#
+# Multi-reference comparison mode (anchored Bradley-Terry MLE ELO over N
+# references — each reference is judged separately and the eval model's rating
+# is fit globally, with per-reference + total win stats reported):
+#   ++gdpval_resources_server.resources_servers.gdpval.reward_mode=comparison \
+#   ++gdpval_resources_server.resources_servers.gdpval.reference_models.kimi.deliverables_dir=/gdpval/refs/kimi \
+#   ++gdpval_resources_server.resources_servers.gdpval.reference_models.kimi.elo=1290 \
+#   ++gdpval_resources_server.resources_servers.gdpval.reference_models.gpt5.deliverables_dir=/gdpval/refs/gpt5 \
+#   ++gdpval_resources_server.resources_servers.gdpval.reference_models.gpt5.elo=1320
 
 # Judge model — proxy to NVIDIA inference API for Gemini 3.1 Pro.
 gdpval_judge_model:
@@ -28,9 +37,14 @@ gdpval_resources_server:
       domain: other
       verified: false
       reward_mode: rubric
+      # Single-reference (legacy) comparison fields.
       reference_deliverables_dir: null
-      num_comparison_trials: 4
       reference_elo: 1000.0
+      # Multi-reference comparison: map of reference id -> {deliverables_dir, elo}.
+      # When set, the eval model's ELO is fit via anchored Bradley-Terry MLE
+      # across all references and per-reference win stats are reported.
+      reference_models: {}
+      num_comparison_trials: 4
       preconvert_office_to_pdf: true
       preconvert_max_concurrent: 1
       judge_model_server: