Upload RESULTS.md with huggingface_hub
Browse files- RESULTS.md +49 -0
RESULTS.md
CHANGED
|
@@ -164,6 +164,55 @@ false-confidence figures are claimed. That axis remains to be measured, and
|
|
| 164 |
it is a natural next step given how badly exact retrieval fails from a dense
|
| 165 |
image.
|
| 166 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 167 |
## Synthesis
|
| 168 |
|
| 169 |
The trade-off that the founding document set out to characterise is real and
|
|
|
|
| 164 |
it is a natural next step given how badly exact retrieval fails from a dense
|
| 165 |
image.
|
| 166 |
|
| 167 |
+
## Core evaluation at proper scale
|
| 168 |
+
|
| 169 |
+
The findings above were first seen in a very small pilot. They have since been
|
| 170 |
+
reproduced on the core subset, which holds 322 instances weighted towards the
|
| 171 |
+
harder tiers, for all three current models under every modality condition at
|
| 172 |
+
the regular density. At this scale the central result is no longer merely
|
| 173 |
+
directional. The collapse of exact retrieval when content is delivered as an
|
| 174 |
+
image is overwhelming, with every one of the discordant paired outcomes in the
|
| 175 |
+
retrieval families flipping in the same direction for all three models, at an
|
| 176 |
+
exact McNemar probability below ten to the power minus twelve.
|
| 177 |
+
|
| 178 |
+
The table gives the exact-match rate by condition, where the two letters are
|
| 179 |
+
the instruction channel and the content channel, and t and i denote text and
|
| 180 |
+
image.
|
| 181 |
+
|
| 182 |
+
| model | text / text | text / image | image / text | image / image |
|
| 183 |
+
|---|---|---|---|---|
|
| 184 |
+
| GPT 5.5 | 1.00 | 0.80 | 0.81 | 0.69 |
|
| 185 |
+
| Gemini 3.5 Flash | 0.95 | 0.80 | 0.94 | 0.85 |
|
| 186 |
+
| Kimi K2.6 | 0.90 | 0.75 | 0.91 | 0.78 |
|
| 187 |
+
|
| 188 |
+
Every model reads best from plain text and worst from a single combined
|
| 189 |
+
image, with the two mixed conditions in between. The one clear model
|
| 190 |
+
difference is that GPT 5.5 follows an instruction rendered into an image less
|
| 191 |
+
reliably than the other two, whose image-instruction accuracy stays close to
|
| 192 |
+
their text baseline.
|
| 193 |
+
|
| 194 |
+
The figures below show the same story. The first plots the modality gradient.
|
| 195 |
+
|
| 196 |
+

|
| 197 |
+
|
| 198 |
+
The second shows which families lose accuracy when content is imaged, namely
|
| 199 |
+
the retrieval families rather than the reasoning ones.
|
| 200 |
+
|
| 201 |
+

|
| 202 |
+
|
| 203 |
+
The third shows the retrieval loss deepening as documents lengthen, since
|
| 204 |
+
tiers four and five are the long documents.
|
| 205 |
+
|
| 206 |
+

|
| 207 |
+
|
| 208 |
+
The fourth places each task family on the trade-off between the tokens saved
|
| 209 |
+
and the accuracy lost on long dense documents. Reasoning and instruction sit
|
| 210 |
+
near the top, saving tokens without losing accuracy, whereas extraction and
|
| 211 |
+
transcription fall to the bottom, saving the same tokens but losing almost
|
| 212 |
+
everything.
|
| 213 |
+
|
| 214 |
+

|
| 215 |
+
|
| 216 |
## Synthesis
|
| 217 |
|
| 218 |
The trade-off that the founding document set out to characterise is real and
|