Agent decisions depend on empirical relationships that deployment observations reveal and task descriptions alone may leave unresolved. We introduce Decoupled Predictive Decision Making (DPDM), an architecture that formulates heterogeneous agent decisions as conditional prediction with a shared pretrained learner. Language models provide semantic representations of current and historical records, while a general-purpose conditional predictor predicts task-defined targets from the represented evidence. Pretraining supplies reusable inference capability; deployment evidence instantiates problem-specific predictions without updating either model's neural parameters. Tool judgments, candidate preferences, and reasoning assessments share an evidence-and-query interface, with predictions mapped to choices by a decision rule. This architecture establishes complementary sources of agent capability: semantic knowledge acquired through language pretraining and empirical inference acquired through structured-data pretraining. A single frozen LimiX-2 predictor serves all five task categories. With Qwen3.5-27B as the semantic encoder, DPDM raises mean accuracy across 50 benchmarks from 78.80% with few-shot prompting to 84.92% under their respective evaluated history budgets, and reaches 82.33% with identical histories and queries. On ToolPRMBench's BFCL and ToolSandbox tasks, absolute accuracy gains are 34.26% and 28.87%, respectively, at the methods' respective evaluated history budgets. Accuracy improves as history grows. With identical imbalanced histories, mean balanced accuracy improves by an absolute 14.51% across five reasoning-step assessment tasks and four encoders. DPDM makes learning from experience an independently improvable foundation-model capability, creating a shared route for advances in conditional prediction to improve decisions across agent tasks.
Figures & tables
Figure 1: DPDM uses historical feedback to make the current decision. A semantic encoder represents each record, and a shared conditional predictor, instantiated here with LimiX-2, estimates the current target from history. The diagram follows the main benchmark: historical targets are supplied separately from record representations, and the highest-probability candidate answer is selected.
Encoder
Zero-shot
Few-shot
DPDM
Qwen3.5-4B
64.56
60.51
80.78 ± 0.11
Qwen3.5-9B
69.92
65.03
82.27 ± 0.15
Qwen3.5-27B
74.02
78.80
84.92 ± 0.14
Llama-3.1-8B
53.49
41.87
77.57 ± 0.13
Jev API
75.94
79.30
—
Table 1: Decision accuracy (%) across 50 benchmarks. Each language model is evaluated through direct prompting and as a semantic encoder in DPDM. Benchmark scores average task configurations; the overall score weights benchmarks equally, and category scores use the same rule within each category. DPDM results average three sampled-history runs; ± denotes their sample standard deviation.
Figure 2: DPDM improves accuracy across Qwen and Llama encoders. Left: mean accuracy over 50 benchmarks; connected points compare few-shot prompting with DPDM using the same encoder, and short marks show zero-shot accuracy. Right: numbers of datasets with higher, equal, or lower accuracy than few-shot prompting.
Figure 3: Decision accuracy across five task categories. Each panel shows one category. Open and filled circles compare few-shot prompting with DPDM using the same encoder; triangles show zero-shot accuracy. The bottom row gives the Jev API reference. Within each panel, the shaded region above 80% uses five times the horizontal scale; no intervals are omitted.
Figure 4: DPDM makes more accurate decisions as history grows. Each curve pairs LimiX-2 with the frozen semantic encoder named in the legend, averaging accuracy over the same 21 datasets. The horizontal axis counts records on a logarithmic scale.
Figure 5: Additional history improves four agent-related judgments. DPDM uses Qwen3.5-27B representations; bands show standard deviations across three history samples. Dashed curves show measured Qwen few-shot accuracy; dotted lines show Jev few-shot references on the same queries. The displayed histories start at 16 records. History sizes and reference budgets are listed in Appendix C.2 .
Figure 6: Decisions under imbalanced histories in BIG-Bench Mistake. Each task averages the three Qwen3.5 encoders and Llama-3.1-8B. The overall mean weights the five tasks equally. The dashed line marks the 50% BA of a constant answer.
Condition
Accuracy (%)
DPDM_Qwen3.5-27B
80.71 ± 0.41
Permuted historical labels
50.13 ± 0.80
Most-frequent-label rule
50.52 ± 1.50
Table 2: Correct historical input–target pairings improve decisions on six tasks. Each history contains 128 records. DPDM_Qwen3.5-27B uses the original input–target pairs. Permuted historical labels retain the same inputs and class counts but shuffle their target assignments. The most-frequent-label rule always returns the modal historical target without conditioning on the query.
Historical records
Rows
Accuracy (%)
16 distinct
16
60.31 ± 3.49
16 repeated 8 times
128
64.79 ± 2.02
128 distinct (DPDM)
128
80.71 ± 0.41
Table 3: Distinct observations provide more useful history than repeated examples on the same six tasks.
History ceiling
DPDM
No semantic features
Permuted labels
16
58.20 ± 2.55
43.04 ± 2.88
38.93 ± 3.09
32
67.80 ± 1.09
51.63 ± 2.08
42.95 ± 2.00
128
75.90 ± 0.44
57.21 ± 1.63
46.45 ± 1.06
512
80.42 ± 0.33
63.95 ± 0.31
48.78 ± 1.99
Table 4: Accuracy (%) on eight tasks with semantic-feature and historical-label ablations.
Figure 7: Measured GPU time relative to cached few-shot prompting. Both totals include historical preparation and processing of the same 256 queries per task. The last row gives the equal-task arithmetic mean.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Encoder/service
Prompt records
DPDM records
StartLux-0.8B
32–4,839
58–8,192
StartLux-4B
24–4,839
58–8,192
StartLux-9B
32–4,839
58–8,192
StartLux-27B
24–2,048
58–8,192
Qwen3.5-4B
32–8,192
58–8,192
Qwen3.5-9B
32–8,192
58–8,192
Appendix
Table 7: History ranges for the main and supplementary comparisons, measured in labeled records.
Task
DPDM
Qwen
Jev
endpoint
few-shot
few-shot
ToolPRMBench BFCL
233
64
16
ToolPRMBench ToolSandbox
228
128
16
CodeCritic Code
2,238
128
32
CodeCritic QA
773
64
16
Appendix
Table 8: Historical record counts for Figure 5 .
Encoder
Zero-shot
Few-shot
DPDM
StartLux-0.8B
65.19
54.91
78.80 ± 0.04
StartLux-4B
72.32
72.97
84.06 ± 0.02
StartLux-9B
74.92
75.20
85.25 ± 0.14
StartLux-27B
78.66
83.85
86.57 ± 0.11
Appendix
Table 9: Decision accuracy (%) with StartLux. The final column pairs each frozen StartLux encoder with LimiX-2.
Few-shot
DPDM
Encoder
M
Hist. correct
Acc.
BA
Acorrect
Aincorrect
Acc.
BA
Acorrect
Aincorrect
Dyck languages
Qwen3.5-4B
1024
94.92
90.74
54.04
95.11
12.96
96.00
62.76
99.96
25.56
Qwen3.5-9B
256
95.31
94.88
54.82
99.65
10.00
92.45
52.84
97.17
8.52
Qwen3.5-27B
1024
94.92
95.05
53.52
100.00
7.04
95.68
60.32
99.90
20.74
Llama-3.1-8B
256
95.31
94.09
49.86
99.35
0.37
93.54
52.19
98.46
5.93
Appendix
Table 10: Encoder–task details for the imbalanced-history analysis. M counts historical records; Hist. correct is their correct-step percentage. Acc. and BA denote ordinary and balanced accuracy. Acorrect and Aincorrect are class-specific accuracies. All scores are percentages.
Task
M
DPDM
Cached few-shot
GPU-h
GPU-h
AG News
1,024
0.036234
0.170823
BeaverTails
1,024
0.036845
0.206056
FaithDial
512
0.023341
0.184489
CommonsenseQA
2,048
0.072727
0.206743
CodeCritic Code
128
0.021397
0.184467
Appendix
Table 11: Full-pipeline GPU-hours on five tasks, with 256 queries per task. Both methods include historical preparation. M counts historical records.
LLM decision-making studies how language models choose, act, seek information, and allocate resources under goals, uncertainty, constraints, and social pressures.