talks
|
works
LABLOG
December 6, 2025
Press D to download as a PDF and F to toggle presentation mode. Navigate up and down with K and J when presenting
LABLOG
Noah Syrkis
December 6, 2025
1 |
Parabellum
2 |
Dimensions
3 |
Perception
4 |
Evaluation
1 |
Parabellum
Aske is an intelligent guy
1 of
7
1 |
Parabellum
▶
We can now generate both intelligence and state counterfactuals:
▶
mask = mask[None, :] & random.bernoulli...
←
for masking state
▶
mask = mask[None, :] | random.bernoulli...
←
for masking intel
▶
We now query
Gemma4b
with an image of the battle space
▶
Next step is (probably) top down test (i.e., send units left or right) …
▶
… or have
Gemma
attempt to infer enemy targets.
▶
… and move unit types and teams to
context
▶
… Add a third level of hierarchy
2 of
7
1 |
Parabellum
▶
Parabellum absolutely needs hierarchy
▶
Tentative ORBAT: Unit
∈
squad
∈
platoon
▶
Simulate
𝑛
𝑙
locations throughout Switzerland
▶
𝑛
𝑡
low-level unit steps for 1 global step (
Gemma
)
Place (i.e., Zurich)
Buildings
Boundaries
Points (UTM)
Rasters (ENU)
States (ECEF)
[
[
[
1
1
4
3
2
3
4
4
]
]
]
[
[
[
[
0
0
0
0
1
0
0
0
1
0
0
0
1
1
0
0
1
0
0
0
0
0
0
0
0
]
]
]
]
3 of
7
2 |
Dimensions
▶
As per
Eq. 1
, we run
𝑛
𝑠
parallel base sims…
▶
… divided into
𝑛
𝑐
time chunks at
𝑛
𝑙
locations …
▶
… of
𝑛
𝑡
time steps each.
▶
Per
Eq. 2
, for each team we compute
𝑛
𝑘
unique
counter factuals, each with
𝑛
𝑟
random seeds
𝑛
𝑠
×
𝑛
𝑐
×
new
⏞
𝑛
𝑙
×
𝑛
𝑡
(1)
𝑛
𝑠
×
𝑛
𝑐
×
𝑛
𝑙
×
2
×
𝑛
𝑘
×
𝑛
𝑟
⏟
⏟
⏟
⏟
⏟
counter factuals
×
𝑛
𝑡
(2)
Figure 2: State sequence dimensions of ground
truths (top) and counter factuals (bottom)
4 of
7
Figure 3:
𝑛
𝑙
⏟
4
×
𝑛
𝑐
⋅
𝑛
𝑡
⏟
1
0
0
concurrently run simulation steps throughout Zurich
Figure 4:
4
⏞
𝑛
𝑙
×
2
⏞
𝑛
𝑘
×
3
⏞
𝑛
𝑟
counterfactuals from the
blue
perspective,
with
red
positions sampled / inferred
3 |
Perception
▶
Gemma270M
successfully processes the intel
▶
Every digit is always assigned to a unique token
▶
Intel coords. are thus mapped to
[
0
.
.
9
]
▶
For map size of 256 we thus have
𝔼
(
𝛿
)
≈
1
0
▶
𝑠
is made from intel randomly masked by
𝑀
̂
𝑠
𝑡
Gemma
𝑠
𝑡
masked
𝔼
[
𝑠
𝑡
]
baseline
5
1
0
1
5
2
0
Distance
𝛿
from
𝑠
𝑡
Figure 5: Distance
𝛿
to
𝑠
𝑡
from best (left), masked
(middle), and baseline (right) guess
5 of
7
4 |
Evaluation
▶
For multiple
𝑠
sims we log divergence from
̂
𝑠
▶
Each
𝑠
being based on a different mask
𝑀
▶
We get importance by solving
𝑀
𝛽
=
𝛿
(
̂
𝑠
,
𝑠
)
Troop
Armor
Plane
Medic
Civil
-2
-1
0
1
2
𝛽
value by unit type
6 of
7
References
7 of
7
LABLOG
Noah Syrkis
December 6, 2025
1 |
Parabellum
2 |
Dimensions
3 |
Perception
4 |
Evaluation
1 |
Parabellum
Aske is an intelligent guy
1 of
7
1 |
Parabellum
▶
We can now generate both intelligence and state counterfactuals:
2 of
7
1 |
Parabellum
▶
We can now generate both intelligence and state counterfactuals:
▶
mask = mask[None, :] & random.bernoulli...
←
for masking state
2 of
7
1 |
Parabellum
▶
We can now generate both intelligence and state counterfactuals:
▶
mask = mask[None, :] & random.bernoulli...
←
for masking state
▶
mask = mask[None, :] | random.bernoulli...
←
for masking intel
2 of
7
1 |
Parabellum
▶
We can now generate both intelligence and state counterfactuals:
▶
mask = mask[None, :] & random.bernoulli...
←
for masking state
▶
mask = mask[None, :] | random.bernoulli...
←
for masking intel
▶
We now query
Gemma4b
with an image of the battle space
2 of
7
1 |
Parabellum
▶
We can now generate both intelligence and state counterfactuals:
▶
mask = mask[None, :] & random.bernoulli...
←
for masking state
▶
mask = mask[None, :] | random.bernoulli...
←
for masking intel
▶
We now query
Gemma4b
with an image of the battle space
▶
Next step is (probably) top down test (i.e., send units left or right) …
2 of
7
1 |
Parabellum
▶
We can now generate both intelligence and state counterfactuals:
▶
mask = mask[None, :] & random.bernoulli...
←
for masking state
▶
mask = mask[None, :] | random.bernoulli...
←
for masking intel
▶
We now query
Gemma4b
with an image of the battle space
▶
Next step is (probably) top down test (i.e., send units left or right) …
▶
… or have
Gemma
attempt to infer enemy targets.
▶
… and move unit types and teams to
context
▶
… Add a third level of hierarchy
2 of
7
1 |
Parabellum
▶
Parabellum absolutely needs hierarchy
Place (i.e., Zurich)
Buildings
Boundaries
Points (UTM)
Rasters (ENU)
States (ECEF)
[
[
[
1
1
4
3
2
3
4
4
]
]
]
[
[
[
[
0
0
0
0
1
0
0
0
1
0
0
0
1
1
0
0
1
0
0
0
0
0
0
0
0
]
]
]
]
3 of
7
1 |
Parabellum
▶
Parabellum absolutely needs hierarchy
▶
Tentative ORBAT: Unit
∈
squad
∈
platoon
Place (i.e., Zurich)
Buildings
Boundaries
Points (UTM)
Rasters (ENU)
States (ECEF)
[
[
[
1
1
4
3
2
3
4
4
]
]
]
[
[
[
[
0
0
0
0
1
0
0
0
1
0
0
0
1
1
0
0
1
0
0
0
0
0
0
0
0
]
]
]
]
3 of
7
1 |
Parabellum
▶
Parabellum absolutely needs hierarchy
▶
Tentative ORBAT: Unit
∈
squad
∈
platoon
▶
Simulate
𝑛
𝑙
locations throughout Switzerland
Place (i.e., Zurich)
Buildings
Boundaries
Points (UTM)
Rasters (ENU)
States (ECEF)
[
[
[
1
1
4
3
2
3
4
4
]
]
]
[
[
[
[
0
0
0
0
1
0
0
0
1
0
0
0
1
1
0
0
1
0
0
0
0
0
0
0
0
]
]
]
]
3 of
7
1 |
Parabellum
▶
Parabellum absolutely needs hierarchy
▶
Tentative ORBAT: Unit
∈
squad
∈
platoon
▶
Simulate
𝑛
𝑙
locations throughout Switzerland
▶
𝑛
𝑡
low-level unit steps for 1 global step (
Gemma
)
Place (i.e., Zurich)
Buildings
Boundaries
Points (UTM)
Rasters (ENU)
States (ECEF)
[
[
[
1
1
4
3
2
3
4
4
]
]
]
[
[
[
[
0
0
0
0
1
0
0
0
1
0
0
0
1
1
0
0
1
0
0
0
0
0
0
0
0
]
]
]
]
3 of
7
2 |
Dimensions
▶
As per
Eq. 1
, we run
𝑛
𝑠
parallel base sims…
▶
… divided into
𝑛
𝑐
time chunks at
𝑛
𝑙
locations …
▶
… of
𝑛
𝑡
time steps each.
▶
Per
Eq. 2
, for each team we compute
𝑛
𝑘
unique
counter factuals, each with
𝑛
𝑟
random seeds
𝑛
𝑠
×
𝑛
𝑐
×
new
⏞
𝑛
𝑙
×
𝑛
𝑡
(1)
𝑛
𝑠
×
𝑛
𝑐
×
𝑛
𝑙
×
2
×
𝑛
𝑘
×
𝑛
𝑟
⏟
⏟
⏟
⏟
⏟
counter factuals
×
𝑛
𝑡
(2)
Figure 2: State sequence dimensions of ground
truths (top) and counter factuals (bottom)
4 of
7
Figure 3:
𝑛
𝑙
⏟
4
×
𝑛
𝑐
⋅
𝑛
𝑡
⏟
1
0
0
concurrently run simulation steps throughout Zurich
Figure 4:
4
⏞
𝑛
𝑙
×
2
⏞
𝑛
𝑘
×
3
⏞
𝑛
𝑟
counterfactuals from the
blue
perspective,
with
red
positions sampled / inferred
3 |
Perception
▶
Gemma270M
successfully processes the intel
▶
Every digit is always assigned to a unique token
▶
Intel coords. are thus mapped to
[
0
.
.
9
]
▶
For map size of 256 we thus have
𝔼
(
𝛿
)
≈
1
0
▶
𝑠
is made from intel randomly masked by
𝑀
̂
𝑠
𝑡
Gemma
𝑠
𝑡
masked
𝔼
[
𝑠
𝑡
]
baseline
5
1
0
1
5
2
0
Distance
𝛿
from
𝑠
𝑡
Figure 5: Distance
𝛿
to
𝑠
𝑡
from best (left), masked
(middle), and baseline (right) guess
5 of
7
4 |
Evaluation
▶
For multiple
𝑠
sims we log divergence from
̂
𝑠
▶
Each
𝑠
being based on a different mask
𝑀
▶
We get importance by solving
𝑀
𝛽
=
𝛿
(
̂
𝑠
,
𝑠
)
Troop
Armor
Plane
Medic
Civil
-2
-1
0
1
2
𝛽
value by unit type
6 of
7
References
7 of
7
