We consider the question: when a large language reasoning model makes a choice, did it think first and then decide to, or decide first and then think? In this paper, we present evidence that detectable, early-encoded decisions shape chain-of-thought in reasoning models. Specifically, we show that a simple linear probe successfully decodes tool-calling decisions from pre-generation activations with very high confidence, and in some cases, even before a single reasoning token is produced. Activation steering supports this causally: perturbing the decision direction leads to inflated deliberation, and flips behavior in many examples (between 7 - 79% depending on model and benchmark). We also show through behavioral analysis that, when steering changes the decision, the chain-of-thought process often rationalizes the flip rather than resisting it. Together, these results suggest that reasoning models can encode action choices before they begin to deliberate in text.
Figures & tables
Figure 1: Overview of our methodology. Linear probes detect action decisions. We apply steering vectors, and measure quantitative as well as behavioral impact on CoT.
Figure 2: Decision predictability using probes at layer 20 for Qwen3-4B and GLM-Z1-9B. Both models exhibit a dip at around 5% of the reasoning trace.
Figure 3: Agreement ratio between decisions detected by probe at layer 20 for various stages and think_end tokens, and correctness, for Qwen3-4B and GLM-Z1-9B.
When2Call
BFCL
Model
α(±)
Suppress (flip %)
Inject (flip %)
α(±)
Suppress (flip %)
Inject (flip %)
(T)
(NT)
(T)
(NT)
(T)
(NT)
(T)
(NT)
Qwen3 4B
4
10
2
16
14
4
6
5
11
9
8
26
6
38
18
8
12
12
32
18
12
49
18
62
29
12
26
20
53
27
GLM Z1-9B
4
10
–
3
–
10
9
–
29
–
Table 1: Steering flip rates on held-out examples ( n=100 per direction). (T) refers to thinking mode, (NT) refers to no-think (unavailable for GLM).
When2Call
BFCL
Model
Dir.
Out.
n
Avg Baseline CoT
Avg Steered CoT
Avg Ratio
n
Avg Baseline CoT
Avg Steered CoT
Avg Ratio
Qwen3 4B
Supp.
Flip
45
537
741
1.38
26
441.7
791.2
1.79
Resist.
55
208
477
2.30
74
266.1
473.5
1.78
Inj.
Flip
62
420
735
1.75
53
430.6
597.6
1.39
Resist.
38
158
156
0.98
47
306.8
305.3
1.00
GLM-Z1 9B
Supp.
Flip
11
1062.4
1605.5
1.51
58
623.2
1261.6
2.02
Table 2: Average CoT token inflation at α=12 (except, α=30 for GLM+BFCL) on held-out examples, grouped by suppress or inject direction, and flipped or resisted outcome.
Figure 5: Probe AUROC across sampled layers and generation positions on When2Call for the two main models, Qwen3-4B and GLM-Z1-9B. In both cases, the strongest probes appear in mid-to-late layers, with strong pre_gen predictability and a dip around 5% to 10% of the reasoning trace.
Figure 6: Probe AUROC across sampled layers and generation positions on BFCL for the two main models, Qwen3-4B and GLM-Z1-9B. Both models preserve strong pre_gen predictability, show the early dip in the reasoning trace, and recover in later positions.
Figure 7: Probe AUROC across sampled layers and generation positions on When2Call for GPT-OSS-20B with medium and high reasoning. Both variants show the same overall pattern as the main models: strong pre_gen predictability, a dip early in the reasoning trace, and recovery toward the end of thinking.
Figure 8: Probe AUROC across sampled layers and generation positions on BFCL for GPT-OSS-20B with medium and high reasoning. As on When2Call, both variants retain strong early predictability, followed by an early dip in the reasoning trace and recovery toward later positions.
Figure 9: Decision predictability across positions on When2Call for GPT-OSS-20B under medium and high reasoning. Across both settings, GPT-OSS exhibits strong pre_gen predictability and a dip early in the reasoning trace.
Figure 10: Decision predictability across positions on BFCL for GPT-OSS-20B under medium and high reasoning. The same overall pattern persists across both reasoning settings, with strong early predictability and an early dip in the reasoning trace.
Figure 11: Agreement with the final think_end probe on When2Call for GPT-OSS-20B under medium and high reasoning. Early-position agreement remains lower than for the two main models, before recovering toward later positions.
Figure 12: Agreement with the final think_end probe on BFCL for GPT-OSS-20B under medium and high reasoning. Agreement likewise strengthens toward later positions in the reasoning trace.
Benchmark
Model
Dir.
Disagree rate
Most common disagreement
When2Call
GLM-Z1 9B
Inject
11.0%
No Meaningful Difference / Inflated Deliberation (4)
Suppress
28.0%
Inflated Deliberation / No Meaningful Difference (13)
Qwen3 4B
Inject
7.0%
No Meaningful Difference / Inflated Deliberation (2)
Suppress
26.3%
Inflated Deliberation / Seamless Divergence (7)
BFCL
GLM-Z1 9B
Inject
38.0%
Confabulated Support / Constraint Override (16)
Suppress
31.0%
Inflated Deliberation / Decision Instability (16)
Appendix
Table 5: Judge disagreement statistics for the behavioral analysis. Each row reports cases where Claude Sonnet 4.6 and GPT-5.4 assigned different buckets. The final column gives the most frequent disagreement pair, with the count in parentheses.
LLM decision-making studies how language models choose, act, seek information, and allocate resources under goals, uncertainty, constraints, and social pressures.