Add dual-vector ablation experiment (z_short/z_long necessity)

- Add rag_vector_no_short (eta_short=0) and rag_vector_no_long (eta_long=0) configs
- Add ablation run script for 60p60s experiment
- Document ablation design and expected results in notes

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

4 files changed, 109 insertions, 0 deletions

diff --git a/collaborativeagents/adapters/personalized_llm_adapter.py b/collaborativeagents/adapters/personalized_llm_adapter.py
index 488b241..4189f31 100644
--- a/collaborativeagents/adapters/personalized_llm_adapter.py
+++ b/collaborativeagents/adapters/personalized_llm_adapter.py
@@ -672,6 +672,40 @@ def create_baseline_adapter(
                 "extractor": "cuda:2",
             },
         ),
+        # Ablation: RAG + Vector without z_short (only z_long, no within-session adaptation)
+        "rag_vector_no_short": AdapterConfig(
+            mode="full",
+            enable_preference_extraction=True,
+            enable_rl_updates=True,
+            use_user_vector=True,
+            llm_name=llm_name,
+            use_shared_models=use_shared_models,
+            enable_query_transform=True,
+            enable_global_preferences=True,
+            eta_short=0.0,  # Disable z_short learning
+            device_assignment={
+                "embed": "cuda:2",
+                "reranker": "cuda:3",
+                "extractor": "cuda:2",
+            },
+        ),
+        # Ablation: RAG + Vector without z_long (only z_short, no cross-session learning)
+        "rag_vector_no_long": AdapterConfig(
+            mode="full",
+            enable_preference_extraction=True,
+            enable_rl_updates=True,
+            use_user_vector=True,
+            llm_name=llm_name,
+            use_shared_models=use_shared_models,
+            enable_query_transform=True,
+            enable_global_preferences=True,
+            eta_long=0.0,  # Disable z_long learning
+            device_assignment={
+                "embed": "cuda:2",
+                "reranker": "cuda:3",
+                "extractor": "cuda:2",
+            },
+        ),
         # Baseline 7a: RAG + Vector + Preference Rewrite (combines best of both)
         "rag_rewrite_vector": AdapterConfig(
             mode="full",
diff --git a/collaborativeagents/scripts/queue_ablation_zshort.sh b/collaborativeagents/scripts/queue_ablation_zshort.sh
new file mode 100755
index 0000000..b4879c7
--- /dev/null
+++ b/collaborativeagents/scripts/queue_ablation_zshort.sh
@@ -0,0 +1,47 @@
+#!/bin/bash
+# Ablation experiment: z_short necessity
+# Run rag_vector_no_short (z_long only) and rag_vector_no_long (z_short only)
+# Same setup as 60s main experiment for direct comparison
+
+echo "Starting z_short ablation experiments..."
+echo "$(date '+%Y-%m-%d %H:%M:%S')"
+
+# First: run rag_vector_no_short (only z_long, no within-session adaptation)
+echo "Starting rag_vector_no_short..."
+python collaborativeagents/scripts/run_experiments.py \
+    --methods rag_vector_no_short \
+    --datasets math-hard,math-500,bigcodebench \
+    --n-profiles 60 \
+    --n-sessions 60 \
+    --max-turns 10 \
+    --use-vllm \
+    --vllm-agent-url http://localhost:8003/v1 \
+    --vllm-user-url http://localhost:8004/v1 \
+    --use-batch-processing \
+    --batch-size 4 \
+    --parallel-profiles 20 \
+    --profile-path collaborativeagents/data/complex_profiles_v2/profiles_200.jsonl \
+    --output-dir collaborativeagents/results/ablation_no_short_60s \
+    2>&1 | tee collaborativeagents/results/ablation_no_short_60s.log
+
+echo "rag_vector_no_short completed at $(date '+%Y-%m-%d %H:%M:%S')"
+
+# Then: run rag_vector_no_long (only z_short, no cross-session learning)
+echo "Starting rag_vector_no_long..."
+python collaborativeagents/scripts/run_experiments.py \
+    --methods rag_vector_no_long \
+    --datasets math-hard,math-500,bigcodebench \
+    --n-profiles 60 \
+    --n-sessions 60 \
+    --max-turns 10 \
+    --use-vllm \
+    --vllm-agent-url http://localhost:8003/v1 \
+    --vllm-user-url http://localhost:8004/v1 \
+    --use-batch-processing \
+    --batch-size 4 \
+    --parallel-profiles 20 \
+    --profile-path collaborativeagents/data/complex_profiles_v2/profiles_200.jsonl \
+    --output-dir collaborativeagents/results/ablation_no_long_60s \
+    2>&1 | tee collaborativeagents/results/ablation_no_long_60s.log
+
+echo "Both ablation experiments completed at $(date '+%Y-%m-%d %H:%M:%S')"
diff --git a/collaborativeagents/scripts/run_experiments.py b/collaborativeagents/scripts/run_experiments.py
index da3549b..806b959 100644
--- a/collaborativeagents/scripts/run_experiments.py
+++ b/collaborativeagents/scripts/run_experiments.py
@@ -118,6 +118,8 @@ AVAILABLE_METHODS = {
     "rag_rewrite": "Extractor + RAG with LLM preference rewrite/merge",
     "rag_rewrite_vector": "Extractor + RAG + user vector + LLM preference rewrite",
     "rag_vector": "Extractor + RAG + user vector (proposed method)",
+    "rag_vector_no_short": "Ablation: RAG + user vector without z_short (long-term only)",
+    "rag_vector_no_long": "Ablation: RAG + user vector without z_long (short-term only)",
     "rag_vector_fast": "Extractor + RAG + user vector with 10x learning rate",
     "rag_vector_consolidate": "Extractor + RAG + user vector with session-level preference consolidation",
     "rag_vector_balanced": "Extractor + RAG + user vector with balanced rewards (10x LR + positive signal for good turns)",
diff --git a/notes.md b/notes.md
index 5864092..919633c 100644
--- a/notes.md
+++ b/notes.md
@@ -97,6 +97,32 @@ z_short 不和长期 preference Jaccard 相关是**符合预期的**：
 
 **可视化**: `collaborativeagents/results/fig_main_results.png` panel (c)
 
+### z_short Ablation Experiment (02/10 启动)
+
+**目的**: 通过消融实验证明dual-vector中z_short的必要性
+
+**实验设计**:
+| 条件 | eta_long | eta_short | beta_long | beta_short | 说明 |
+|------|----------|-----------|-----------|------------|------|
+| Full (baseline) | 0.01 | 0.05 | 2.0 | 5.0 | 完整dual-vector (已完成的60s实验) |
+| No z_short | 0.01 | **0.0** | 2.0 | 5.0 | 禁用session内适应 |
+| No z_long | **0.0** | 0.05 | 2.0 | 5.0 | 禁用跨session学习 |
+
+**配置**: 60 profiles × 60 sessions, max 10 turns, same as main experiment
+**方法名**: `rag_vector_no_short`, `rag_vector_no_long`
+
+**预期结果**:
+1. **No z_short** (z_long only): 跨session学习正常，但session内第1-2 turn后如果agent犯错无法快速修正偏好权重 → 预计E/T会更高（更多enforcement）
+2. **No z_long** (z_short only): 每个session从零开始（z_short在session开始时reset），完全没有跨session记忆 → 预计后期session表现差，没有学习曲线
+3. **Full** (both): 最佳表现，z_long提供跨session基础，z_short提供session内微调
+
+**关键对比指标**:
+- **E/T per-session曲线**: No z_short应该在early turns有更高E/T
+- **Success rate over sessions**: No z_long应该flat，Full应该upward
+- **Late-session performance**: Full应该最好，No z_long应该和early一样
+
+**状态**: 🔄 Running (02/10 21:48 启动)
+
 ---
 
 ## RAG vs Reflection 分析