| """ |
| ScratchpadLayer Module |
| |
| Persistent state tracking layer for multi-turn confessional reasoning. |
| Maintains a learnable scratchpad state that accumulates across reasoning steps. |
| """ |
|
|
| import torch |
| import torch.nn as nn |
|
|
|
|
| class ScratchpadLayer(nn.Module): |
| """ |
| Scratchpad layer for maintaining persistent state across confessional reasoning cycles. |
| """ |
| def __init__(self, d_model): |
| super().__init__() |
| self.pad_proj = nn.Linear(d_model, d_model) |
| self.reset = nn.Parameter(torch.zeros(1, d_model)) |
|
|
| def forward(self, x, prev_z=None): |
| """ |
| Update scratchpad state with new input. |
| |
| Args: |
| x: Input tensor (batch_size, sequence_length, d_model) |
| prev_z: Previous scratchpad state (batch_size, d_model), None for reset |
| |
| Returns: |
| Updated scratchpad state (batch_size, d_model) |
| """ |
| if prev_z is None: |
| prev_z = self.reset.expand(x.size(0), -1) |
| x_pooled = x.mean(dim=1) |
| z = self.pad_proj(x_pooled) + 0.7 * prev_z |
| return z |
|
|
