File size: 10,589 Bytes
3e363e2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 | """Output heads: answer selection, relation, support, contradiction, confidence.
The selection head is the substantive addition. Previously ``reason`` returned
*every* path the verifier accepted, so the answer set was "everything reachable"
and exact-match against a specific gold set was unattainable by construction.
Selection is the model's job: :class:`PathSelectionHead` emits a per-path logit
and a learned decision threshold, so the model commits to a subset and is
scored on that commitment.
"""
from __future__ import annotations
from dataclasses import dataclass
import torch
from torch import Tensor, nn
@dataclass(frozen=True)
class SelectionOutput:
"""Per-path selection logits and the resulting boolean decision."""
logits: Tensor
keep: Tensor
threshold: Tensor
class PathSelectionHead(nn.Module):
"""Score candidate paths and decide which ones support the answer.
The threshold is **query-dependent**: a global bias plus an offset predicted
from the query and the pooled candidate evidence. A single global scalar
cannot express both "this query has one answer" and "this one has seven", and
the measured symptom was a threshold that drifted without becoming
discriminative -- selection precision flat near 0.20 while path recall fell
from 0.761 to 0.368, and a wider beam scoring *worse* than greedy.
"""
def __init__(self, hidden_dim: int, dropout: float = 0.1) -> None:
super().__init__()
self.score = nn.Sequential(
nn.Linear(hidden_dim * 3, hidden_dim),
nn.GELU(),
nn.Dropout(dropout),
nn.Linear(hidden_dim, hidden_dim // 2),
nn.GELU(),
nn.Linear(hidden_dim // 2, 1),
)
self.threshold = nn.Parameter(torch.zeros(1))
self.threshold_offset = nn.Sequential(
nn.Linear(hidden_dim * 2, hidden_dim // 2),
nn.GELU(),
nn.Linear(hidden_dim // 2, 1),
)
def forward(
self, reasoning_states: Tensor, path_states: Tensor, query_state: Tensor
) -> SelectionOutput:
if query_state.ndim == 1:
query_state = query_state.unsqueeze(0)
if query_state.shape[0] == 1 and reasoning_states.shape[0] != 1:
query_state = query_state.expand(reasoning_states.shape[0], -1)
features = torch.cat((reasoning_states, path_states, query_state), dim=-1)
logits = self.score(features).squeeze(-1)
# One threshold per query, from the query plus the pooled evidence.
context = torch.cat(
(query_state[:1], path_states.mean(0, keepdim=True)), dim=-1
)
threshold = self.threshold + self.threshold_offset(context).reshape(())
return SelectionOutput(logits, logits > threshold, threshold)
class NextHopHead(nn.Module):
"""Score each candidate transition for beam expansion (PLAN section 14)."""
def __init__(self, hidden_dim: int, dropout: float = 0.1) -> None:
super().__init__()
self.score = nn.Sequential(
nn.Linear(hidden_dim * 4, hidden_dim),
nn.GELU(),
nn.Dropout(dropout),
nn.Linear(hidden_dim, 1),
)
def forward(
self,
reasoning_states: Tensor,
source_states: Tensor,
relation_states: Tensor,
target_states: Tensor,
) -> Tensor:
features = torch.cat(
(reasoning_states, source_states, relation_states, target_states), dim=-1
)
return self.score(features).squeeze(-1)
class EntityAnswerHead(nn.Module):
"""Rank candidate answer entities against the reasoning state."""
def __init__(self, hidden_dim: int) -> None:
super().__init__()
self.project = nn.Linear(hidden_dim, hidden_dim)
self.scale = nn.Parameter(torch.ones(1))
def forward(self, reasoning_state: Tensor, entity_states: Tensor) -> Tensor:
if reasoning_state.ndim == 1:
reasoning_state = reasoning_state.unsqueeze(0)
projected = self.project(reasoning_state)
return self.scale * (entity_states @ projected.transpose(-2, -1)).squeeze(-1)
class RelationAnswerHead(nn.Module):
"""Predict the relation that answers a relation-mode query."""
def __init__(self, hidden_dim: int, num_relations: int) -> None:
super().__init__()
self.score = nn.Sequential(
nn.Linear(hidden_dim * 2, hidden_dim),
nn.GELU(),
nn.Linear(hidden_dim, num_relations),
)
def forward(self, reasoning_state: Tensor, relation_context: Tensor) -> Tensor:
return self.score(torch.cat((reasoning_state, relation_context), dim=-1))
class VerificationStack(nn.Module):
"""Shared neural verification trunk consumed by the support-side heads.
PLAN section 20 describes neural verification as consuming the query, the PAT
path representation, the graph representation, the candidate answer and the
constraint state. That is a real inference problem, not a linear readout, so
it gets a residual MLP trunk of configurable depth and width whose output is
shared by the support, contradiction, count and confidence heads.
Sharing one trunk is also what makes the heads bucket reach its section 6
budget honestly: the capacity does inference work rather than padding a
per-head linear layer to hit a number.
"""
def __init__(
self,
hidden_dim: int,
depth: int = 2,
width_multiplier: int = 4,
dropout: float = 0.1,
num_signals: int = 0,
) -> None:
super().__init__()
if depth <= 0 or width_multiplier <= 0:
raise ValueError("verification depth and width_multiplier must be positive")
if num_signals < 0:
raise ValueError("num_signals must not be negative")
inner = hidden_dim * width_multiplier
self.num_signals = num_signals
self.input_projection = nn.Linear(hidden_dim * 3, hidden_dim)
# Scalar evidence summaries, projected and added rather than concatenated so
# the trunk width is unchanged. These exist because the pooled evidence mean
# this trunk otherwise receives is built from *different distributions* at
# training and inference -- gold positives plus sampled negatives during
# training, the beam's own candidates at inference -- and the support head
# learned a rule that does not transfer: it abstained on 25% of answerable
# examples and only 12.5% of unanswerable ones, worse than chance in the
# direction that matters. A signal computed identically in both regimes fixes
# the mismatch rather than reweighting around it.
self.signal_projection = (
nn.Linear(num_signals, hidden_dim) if num_signals else None
)
self.blocks = nn.ModuleList(
[
nn.Sequential(
nn.LayerNorm(hidden_dim),
nn.Linear(hidden_dim, inner),
nn.GELU(),
nn.Dropout(dropout),
nn.Linear(inner, hidden_dim),
)
for _ in range(depth)
]
)
self.final_norm = nn.LayerNorm(hidden_dim)
def forward(
self,
reasoning_state: Tensor,
evidence_state: Tensor,
query_state: Tensor,
signals: Tensor | None = None,
) -> Tensor:
states = self.input_projection(
torch.cat((reasoning_state, evidence_state, query_state), dim=-1)
)
if self.signal_projection is not None:
if signals is None:
raise ValueError(
f"this verification stack expects {self.num_signals} evidence "
"signals but none were supplied"
)
if signals.ndim == 1:
signals = signals.unsqueeze(0)
if signals.shape[-1] != self.num_signals:
raise ValueError(
f"expected {self.num_signals} evidence signals, "
f"got {signals.shape[-1]}"
)
states = states + self.signal_projection(signals.to(states.dtype))
for block in self.blocks:
states = states + block(states)
return self.final_norm(states)
class SupportHead(nn.Module):
"""Estimate ``P(answer supported | query, evidence, graph)``."""
def __init__(self, hidden_dim: int) -> None:
super().__init__()
self.score = nn.Sequential(
nn.Linear(hidden_dim, hidden_dim // 2),
nn.GELU(),
nn.Linear(hidden_dim // 2, 1),
)
def forward(self, verified_state: Tensor) -> Tensor:
return self.score(verified_state).squeeze(-1)
class ContradictionHead(nn.Module):
"""Four-way status classifier over the PLAN section 19 output states."""
def __init__(self, hidden_dim: int) -> None:
super().__init__()
self.score = nn.Sequential(
nn.Linear(hidden_dim, hidden_dim // 2),
nn.GELU(),
nn.Linear(hidden_dim // 2, 4),
)
def forward(self, verified_state: Tensor) -> Tensor:
return self.score(verified_state)
class CountHead(nn.Module):
"""Predict an aggregate count for count/aggregate answer modes."""
def __init__(self, hidden_dim: int, max_count: int = 32) -> None:
super().__init__()
self.max_count = max_count
self.score = nn.Sequential(
nn.Linear(hidden_dim, hidden_dim // 2),
nn.GELU(),
nn.Linear(hidden_dim // 2, max_count + 1),
)
def forward(self, verified_state: Tensor) -> Tensor:
return self.score(verified_state)
class ConfidenceHead(nn.Module):
"""Emit a calibrated-in-training confidence logit (PLAN section 47)."""
def __init__(self, hidden_dim: int) -> None:
super().__init__()
self.score = nn.Sequential(
nn.Linear(hidden_dim + 2, hidden_dim // 2),
nn.GELU(),
nn.Linear(hidden_dim // 2, 1),
)
def forward(
self,
verified_state: Tensor,
support_logit: Tensor,
selected_fraction: Tensor,
) -> Tensor:
features = torch.cat(
(
verified_state.reshape(1, -1),
support_logit.reshape(1, -1),
selected_fraction.reshape(1, -1),
),
dim=-1,
)
return self.score(features).squeeze(-1)
|