RAG wastes tokens. We propose Knowledge Packs: pre-computed KV caches that deliver the same knowledge at zero token cost. For causal transformers, the KV cache from a forward pass on text F is identical to what a joint pass on F+q would produce - this follows directly from the causal mask. The equivalence is exact but fragile: wrong chat template formatting causes 6-7pp degradation, which we believe explains prior claims of KV outperforming RAG. With correct formatting: zero divergences across 700 questions on Qwen3-8B and Llama-3.1-8B, up to 95% token savings. The KV interface also enables behavioral steering that RAG cannot do. Because RoPE rotates keys but leaves values untouched, contrastive deltas on cached values can nudge model behavior while key arithmetic destroys coherence. The effect sits in mid-layer values (33-66%), independent directions are nearly orthogonal (cos~0) and compose, and both channels - knowledge and steering - run simultaneously at alpha<=0.7 without interference. No training, no weight modification.
Figures & tables
Model
Method
Overall
Bridge
Comp.
Tok.
Diverg.
Qwen3-8B ( N=500 )
Baseline
28.4%
17.1%
62.4%
0
—
RAG (BGE top-2)
65.2%
60.5%
79.2%
284
—
KV-chat
65.2%
60.5%
79.2%
0
0/500
KV-raw (no template)
59.2%
54.1%
74.4%
0
—
Llama-3.1-8B ( N=200 )
Baseline
29.5%
18.7%
62.0%
0
—
RAG (BGE top-2)
61.5%
56.7%
76.0%
290
—
Table 1: HotpotQA results. KV-chat and RAG use identical system messages with identical chat template formatting; the only difference is the delivery mechanism.
Searches
Model
KV EM
RAG EM
KV tok
RAG tok
Savings
1
Qwen3-8B
43.0%
43.0%
35
176
141 (80%)
Llama-3.1-8B
42.0%
42.0%
31
188
157 (84%)
2
Qwen3-8B
59.0%
59.0%
35
299
264 (88%)
Llama-3.1-8B
59.0%
58.0%
31
305
274 (90%)
3
Qwen3-8B
65.0%
65.0%
35
438
404 (92%)
Llama-3.1-8B
61.0%
61.0%
31
437
406 (93%)
Table 2: Accumulation scaling ( N=100 per level per model). As retrieval steps increase, RAG token cost grows linearly while KV remains constant. Accuracy is identical at every level on both models.
Figure 1: Token cost scaling with retrieval steps. RAG prompt cost grows linearly ( ∼ 140–150 tokens per step) while KV injection remains constant ( ∼ 31–35 tokens). Shaded area shows cumulative token savings.
Model
Method
Overall
Bridge
Comp.
Qwen3-8B
KV-chat
65.2%
60.5%
79.2%
KV-raw
59.2%
54.1%
74.4%
Difference
+6.0pp
+6.4pp
+4.8pp
Llama-3.1-8B
KV-chat
61.5%
56.7%
76.0%
KV-raw
56.0%
49.3%
76.0%
Difference
+5.5pp
+7.4pp
0.0pp
Table 3: Chat template effect by model and question type.
Table 6: Value steering alpha sweep with 30 contrastive pairs on two models. Optimal α and coherence cliff differ, but both models show significant improvement over baseline at 100% coherence.
Qwen3-8B ( α=2.0 )
Llama-3.1-8B ( α=1.5 )
Layer Range
n
Score
n
Score
All (0–100%)
36
1.93
32
2.47
Early (0–33%)
11
1.20
10
1.33
Mid (33–66%)
12
1.93
11
2.27
Late (66–100%)
13
0.67
11
1.47
Baseline
—
0.60
—
1.47
Table 7: Layer-selective value steering on two models (each at its optimal α , 30 pairs). Mid layers (33–66%) carry the steering effect on both architectures.
Figure 2: Layer-selective value steering on two models. Mid layers (33–66%, green highlight) capture the full steering effect, matching all-layer steering. Late layers produce scores indistinguishable from baseline.
Qwen3-8B ( α=2.0 )
Llama-3.1-8B ( α=1.5 )
Condition
Def.
Doc.
Coh.
Def.
Doc.
Coh.
Baseline
0.40
0.00
100%
1.47
—
100%
A only (defensive)
1.27
1.00
100%
1.87
3.80
100%
B only (documented)
0.73
3.80
100%
1.93
3.53
100%
A+B composed
1.67
3.40
100%
2.13
4.07
100%
Def. retention
131%
114%
Table 8: Composable value steering on two models. “Def.” counts error-handling patterns (0–9), “Doc.” counts docstring/type-hint patterns (0–11). Retention shows A+B score as percentage of single-direction score.
Condition
EM
Form.
Words
A
No knowledge, no steering
20.5%
0.14
16
B
Knowledge only (KV)
72.5%
0.20
35
C
Steering only (V-delta)
15.0%
0.32
52
D
Knowledge + Steering ( α=2.0 )
40.5%
0.32
55
E
Text knowledge (RAG)
73.0%
0.13
11
F
Text knowledge + text steering
77.5%
0.26
48
Table 9: Dual-channel experiment: knowledge delivery + value steering simultaneously (Qwen3-8B, N=200 HotpotQA bridge questions). EM = exact match, Form. = formality score, Words = average response length.
α
EM
Formality
Words
0.0
72.5%
0.20
35
0.1
73.0%
0.21
37
0.2
73.0%
0.21
39
0.5
72.0%
0.23
43
0.7
73.0%
0.24
45
1.0
67.0%
0.25
49
Table 10: Alpha sweep for dual mode (knowledge + steering). At α≤0.7 , factual accuracy is fully preserved while steering effect is measurable. The Pareto frontier is governed by a single scalar.
Figure 3: Dual-channel trade-off (Qwen3-8B, N=200 ). At α≤0.7 (green zone), factual accuracy is fully preserved while behavioral steering is measurable. Higher α trades accuracy for stronger steering.