Update BLOCK_WORLD.md

BLOCK_WORLD.md

@@ -126,82 +126,6 @@ N,K,start_state,goal_state,current_state,next_state,move,num_moves
 
 ---
 
-## 🧮 Complexity Analysis
-
-### Solution Length Formula
-
-For N blocks across K stacks, the **worst-case optimal solution length** is:
-
-```
-L(N) = O(N)
-```
-
-**Explanation**: In the worst case, each block must be moved exactly once from its initial position to its goal position. When K ≥ 3 stacks are available, there is always at least one empty stack (or a stack whose top block is already correctly positioned) that can serve as a temporary buffer.
-
-### Transition Verification
-
-Checking whether a proposed move is legal requires:
-
-```
-C_verify = O(1)
-```
-
-**Why O(1)?** To validate a move `[block, source, dest]`:
-1. Check that `source` stack is non-empty (the block to be moved exists)
-2. Check that `dest` stack either exists or is an empty position
-
-Both checks involve reading only the **top element** of two stacks, independent of N.
-
-### Action Space
-
-The branching factor (number of possible moves per state):
-
-```
-b = O(K²)
-```
-
-With K=3 stacks, at most 3×2 = 6 moves per state (from each non-empty stack to the other two stacks).
-
----
-
-## 📈 Benchmark Results
-
-### Model Performance (from paper)
-
-| Model | Condition | Val Accuracy (N=1-7) | OOD Accuracy (N=8-10) |
-|-------|-----------|---------------------|---------------------|
-| **T5-small (60M)** | Pre-trained + Fine-tuned | **97.27%** | **81.00%** ✅ |
-| GPT-2 (124M) | Pre-trained + Fine-tuned | 24.55% | 0.00% |
-| T5-small | Trained from scratch | 0.00% | 0.00% |
-| GPT-2 | Trained from scratch | 21.82% | 0.00% |
-
-### Key Findings
-
-1. **Architecture > Scale**: T5 (60M parameters) dramatically outperforms GPT-2 (124M parameters)
-   - T5's bidirectional encoder provides full goal access at every decoding step
-   - GPT-2's causal mask creates a planning bottleneck
-
-2. **Pre-training Helps**: Pre-trained T5 gains +97.27 percentage points over training from scratch
-   - Pre-training transfers because Block World has **O(1) local transition structure**
-
-3. **Gradual OOD Degradation**: T5 maintains strong performance even on harder instances
-   - N=8: 84% accuracy
-   - N=9: 84% accuracy  
-   - N=10: 75% accuracy
-   - This gradual decay (not abrupt collapse) indicates genuine rule learning
-
-### Failure Mode Analysis
-
-**T5 Fine-tuned failures** (on the 3% validation set that failed):
-- **Loops**: 67% - Model generates valid moves but revisits previous states
-- **Premature Stop**: 33% - Model halts before reaching goal
-
-**GPT-2 Fine-tuned failures**:
-- **Validation**: 81% loops (model learns legal moves but no goal direction)
-- **OOD**: 91% invalid moves (systematic shift! Constraint generalization fails)
-
----
-
 ## 💡 Usage Tips
 
 ### For Model Training