Organizations: Microsoft, India · Singapore Institute of Technology, Singapore · IIIT Delhi, India · Nanyang Technological University, Singapore · MIDAS Lab, IIT Kanpur, India · IIT Hyderabad, India
Organisations increasingly deploy large language model agents that must follow organisational values. We check whether activation steering, which adds vectors to a model's internal activations during generation without changing its weights, can improve compliance beyond prompting alone. Since most scenarios involve several values at once, we focus on multi-value steering. A simple way to build these vectors is to contrast responses that follow all values relevant to a scenario with responses that violate them. However, this gives nearly identical vectors for different values: they capture a general compliance signal rather than individual values, which limits their use when combining several values. We propose Value-Isolating Activation Steering (VIAS), which contrasts responses to the same scenario under two prompts, one listing all relevant values and one omitting a single target value. We keep a pair only if the target value changes from compliant to non-compliant and both responses still meet a minimum compliance level on the other values. We evaluate VIAS on PROTECT, a workplace compliance dataset where each scenario is tagged with relevant values from a 21-value taxonomy. VIAS reduces overlap between value vectors, making them better suited for multi-value steering, although this alone does not lead to independent control over each value. Combined with prompting, VIAS complies with every tagged value on 69.6% of scenarios for Qwen2.5-7B, 75.2% for Mistral-7B, and 72.6% for Phi-3.5-mini, outperforming prompting alone for all three models.
Figures & tables
Figure 1: Overview of VIAS. (a) Each scenario is tagged with a subset of the 21 values; dInt , dCom , dAcc are the Integrity, Compliance, and Accountability directions. (b) VIAS drops one value V from the prompt and keeps the pair only if the judge’s verdict on V changes from complies to not complies; afull and a∖V are aV+ and aV− in Eq. ( 2 ). (c) All-values directions are nearly collinear (cosine 0.91 – 0.94 ); VIAS directions overlap less ( ∣cos∣0.09 – 0.45 ) but are not orthogonal. One combined vector v steers all test scenarios.
Qwen2.5-7B
Mistral-7B
Phi-3.5-mini
Method
sample
micro
macro
sample
micro
macro
sample
micro
macro
Baseline
60.8
86.7
82.7
63.9
87.4
83.6
61.6
87.9
83.8
Values-prompt (VP)
67.4
89.4
86.5
72.1
91.5
89.1
60.5
85.9
81.7
SFT (LoRA, dev-selected)
63.8
88.3
84.4
62.0
87.1
83.1
63.5
88.4
85.6
All-values contrasts, lenient prompts
All-values (lenient)
60.4
87.0
83.0
65.8
89.1
85.5
61.6
87.8
83.8
Table 1: Compliance on the PROTECT test set (%). sample is the strict all-tagged-values metric, micro is per pair, macro averages over values. All-values uses our multi-value adaptation of CAA. Persyn uses synthetic single-value pairs; VIAS uses value-isolating contrasts on PROTECT scenarios. Table 2 defines the variant labels; Appendix A gives the prompt templates. Bold marks the best VIAS-based result per model. SFT uses development-selected checkpoints and separate evaluation settings (Appendix B ).
Figure 2: Per-value compliance for each model’s baseline and best VIAS+VP configuration (raw for Qwen and Phi; weighted for Mistral), shown in alternating rows and sorted by baseline compliance. Baseline compliance is high for many values but lower for Innovation, Leadership, and Competence. VIAS+VP improves most values but lowers a few, e.g. Reliability on Mistral and Phi.
Figure 3: (a) Strict sample compliance; dashed lines mark the baseline. (b) Hard-subset recovery: the fraction of (scenario, value) pairs the baseline got wrong that each method fixes. Colour marks the method, a lighter shade the weighted variant, and hatching the values-prompt (VP). All-values without VP uses the better of the lenient and strict extraction prompts per model. SFT ∗ recovery is measured against its separately evaluated base model.
Label
Meaning
Values-prompt (VP)
Lists the scenario’s tagged values and asks the model to follow them at inference.
Lenient
Extraction prompts list value names without their definitions.
Strict
Extraction prompts also give value definitions and explicitly require every value to be addressed.
Raw
Averages all retained pair differences equally into one global vector.
Weighted
Combines usable value directions with reliability weights and Gram–Schmidt; global unless targeted.
Targeted
Uses weighted combination of only the current scenario’s tagged-value directions, rather than one global vector.
Table 2: Variant labels can be combined; they do not name independent methods. A global vector is shared across scenarios for a given model, layer, and construction.
Model
Contrasts
cˉ
∣c∣
reff
Qwen2.5-7B (L23)
All-values
0.938
0.938
1.13
VIAS
−0.005
0.093
16.53
Mistral-7B (L24)
All-values
0.911
0.911
1.19
VIAS
0.046
0.099
16.06
Phi-3.5-mini (L16)
All-values
0.940
0.940
1.12
VIAS
0.046
0.453
3.57
Table 3: Mean signed/absolute cosines and participation-ratio rank of 21 raw directions before Gram–Schmidt. Cosines average 210 distinct pairs; all-values uses strict prompts. VIAS reduces, but does not eliminate, overlap.
Figure 4: Pairwise cosine of the 21 per-value steering directions at each model’s VIAS layer. The left panels use all-values contrasts and the right panels use VIAS contrasts, before Gram–Schmidt. Cosines below −0.2 share the darkest colour.
Figure 5: VIAS strict sample compliance (%) vs. fraction of clean value-isolating pairs used to build steering vectors, at fixed coefficients (raw/weighted α : Qwen 12/32 , Mistral 6/8 , Phi 12/32 ) that differ from Table 1 , so the 100% points need not match it. With VP, raw VIAS with 25% of pairs slightly exceeds the values-prompt on Qwen and Mistral. Dashed lines mark baseline (black) and values-prompt (gray).
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Source
Yes / 20
Yes (%)
Intended assignment (fixed reference)
20/20
100.0
Human annotator 1
17/20
85.0
Human annotator 2
16/20
80.0
Both annotators
14/20
70.0
At least one annotator
19/20
95.0
Appendix
Table 4: Persyn scenario-relevance check on 20 sampled scenarios. Human–human agreement is 75% .
Model
Method
agree
κ3
κ2
Qwen2.5-7B
Baseline
0.871
0.561
0.568
VP
0.882
0.543
0.550
All-values
0.864
0.550
0.555
VIAS+VP
0.876
0.522
0.524
Mistral-7B
Baseline
0.879
0.569
0.576
VP
0.878
0.482
0.489
Appendix
Table 5: Agreement between the two judges on individual (scenario, value) pairs. κ3 uses the three labels complies / violates / unclear; κ2 collapses to complies versus not. Each cell covers 3,448 pairs; All-values here uses strict extraction prompts.
Model
family of
J1
J2
gap
Qwen2.5-7B
judge 1
60.8
44.8
+16.0
Mistral-7B
judge 2
63.9
46.6
+17.3
Phi-3.5-mini
neither
61.6
44.4
+17.2
Appendix
Table 6: Strict sample compliance of the no-values baseline under judge 1 (Qwen2.5-32B) and judge 2 (Mistral-Small-24B). These baseline gaps show no obvious matching-family advantage.
Judge 1
Judge 2
Model
Comp.
Δ
p
Δ
p
Qwen2.5-7B
VP
+2.3
.108
+1.8
.268
Base
+8.9
.000
+6.2
.000
All-values
+9.8
.000
+6.9
.000
Mistral-7B
VP
+3.1
.054
+1.9
.277
Base
+11.4
.000
+11.1
.000
Appendix
Table 7: Paired bootstrap ( 10,000 resamples over all 800 scenarios, seed 42 ) for VIAS+VP against each reference method, computed separately under each judge on identical responses. The All-values reference is the strict-prompt run. Δ is the difference between the strict sample compliance percentages, computed on unrounded rates, so it can differ by 0.1 from the difference of the rounded entries in Table 1 (Qwen VIAS+VP − VP is exactly 2.25 ). These nine contrasts retain their signs and nominal 5% significance decisions
Pair
agree (%)
κ
Annotator 1 – Annotator 2
76.6
0.25
Annotator 1 – Judge 1
80.6
0.40
Annotator 2 – Judge 1
81.5
0.43
Appendix
Table 8: Agreement on 248 (scenario, value) labels. 95% scenario-bootstrap intervals for κ are [0.11,0.40] , [0.17,0.60] , and [0.27,0.55] .
Figure 6: Layer-wise label prediction and steering performance. (a) Probe macro-F1 for value tags. (b) Development macro compliance for weighted All-values at matched injected norm. Relative depth runs from the first to the last layer; the best layer for prediction need not be best for steering.
Qwen2.5-7B
Mistral-7B
Phi-3.5-mini
Combination and selection
sample
micro
macro
sample
micro
macro
sample
micro
macro
Raw, original
59.9
86.6
83.0
64.1
87.7
83.5
61.8
87.8
83.7
Raw, dev-selected
57.1
83.9
79.1
54.0
82.6
76.6
57.1
84.0
79.2
Weighted, original
60.9
86.9
83.0
63.6
87.8
84.2
62.7
88.1
84.2
Weighted, dev-selected
56.5
83.7
79.2
57.4
83.6
78.8
55.1
83.8
79.0
Appendix
Table 9: All-values compliance (%) under two layer-and-strength selection procedures, using strict extraction prompts and no values-prompt. The original rows repeat the corresponding strict All-values results in Table 1 .
Figure 7: Development macro compliance (solid) and coherence (dotted) versus injected norm τ for the weighted All-values vector. The red line is the coherence gate at 80 .
Figure 8: Negated steering (push toward violation) with the lenient All-values vectors on 400 test scenarios. Only Mistral’s compliance drops; Qwen and Phi are robust.
Figure 9: VIAS coefficient sweeps for all three models. Each panel shows sample compliance vs. α for VIAS (blue) and VIAS, weighted (orange), with (solid) and without (dotted) the values-prompt. For Qwen, VIAS+VP peaks at α=24 ( 69.6% ) and weighted VIAS+VP reaches 69.1% at α=64 . For Mistral (L24), VIAS+VP peaks at α=6 ( 75.2% ) and weighted VIAS+VP peaks at α=10 ( 75.5% ). For Phi (L16), VIAS alone at α=24 reaches 66.2% without VP; VIAS+VP peaks at α=32 ( 72.6% ) and weighted VIAS+VP peaks at α=32 ( 69.2% ).
Figure 10: Qwen2.5-7B. Both responses de-escalate and propose support for the older team members; only VIAS+VP frames the joint search for solutions as a source of innovative ideas, which satisfies Innovation. VIAS without the values-prompt also complies on Innovation.
Figure 11: Qwen2.5-7B. The two responses share the same plan; both name Competence in the opening sentence, but only VIAS+VP acts on it by checking skills and filling gaps. Competence is unclear under all seven other settings.
Figure 12: Mistral-7B. Both responses route the project through the privacy team. The VP response names only Privacy, Compliance, and Integrity and treats the project purely as a compliance matter; VIAS+VP also encourages innovative solutions within the same safeguards. VIAS without the values-prompt also complies on Innovation.
Figure 13: Mistral-7B. Both responses refuse the request. VP mainly points to proper channels and documentation; VIAS+VP proposes concrete alternatives for faster workflows, such as system improvements and new workflow strategies. Innovation is unclear under all seven other settings.
Figure 14: Phi-3.5-mini. Both responses reject the request. The VP response names three of the four values and treats the question as a breach to correct; VIAS+VP also redirects the profit motive toward legitimate improvements. Innovation is labelled violates or unclear under all seven other settings.
Figure 15: Phi-3.5-mini. The VP response lists five of the six tagged values, omits Flexibility, and follows that list; VIAS+VP adds a Flexibility step. Flexibility complies in every setting without the values-prompt but is lost in every setting with it except VIAS+VP.