Accurately understanding user intent in context is essential for large language models (LLMs) to provide effective assistance. However, observed behavior may support multiple plausible intents, making the current context alone insufficient for identifying the user's underlying goal. A user's interaction history may contain recurring context-to-intent patterns that help disambiguate the current intent. Leveraging such historical patterns requires deciding whether to invoke retrieval for the current context and, if so, forming queries that express plausible intents rather than relying on contextual similarity alone. We present IntPro, a proxy agent that learns to either infer intent directly from the current context or generate candidate intent explanations to retrieve intent-aligned records from the user's history. We design intent explanations as a shared intent-oriented natural-language form that expresses the observed action, its functional goal, and a cue-to-intent rationale. Label-conditioned explanations represent historical records, while candidate explanations generated from the current context serve as retrieval queries in the same form. IntPro first learns direct and retrieval-conditioned inference through Direct-and-Retrieval supervised fine-tuning (SFT). IntPro then uses multi-turn GRPO with a retrieval-decision reward to jointly optimize intent prediction and selective retrieval mechanism. Experiments across three scenarios and three model backbones demonstrate that IntPro consistently improves context-aware intent inference.
Figures & tables
Fig. 1: Conceptual comparison of current Human–LLM collaboration and the Human–Proxy–LLM pipeline. The proxy chooses between direct inference from the current context and retrieval-conditioned inference using the per-user intent library, then supplies the inferred structured intent to the downstream LLM.
Fig. 2: Construction of the per-user intent library and supervised trajectories. Historical records are represented by label-conditioned intent explanations, while candidate intent explanations generated from C serve as retrieval queries. Ground truth supports label validation and final teacher target construction.
Fig. 3: Two-stage training of IntPro. SFT learns direct and retrieval-conditioned inference from validated teacher trajectories; GRPO then samples complete trajectories following either path and uses final intent correctness to optimize intent prediction and the retrieval decision.
Dataset
Total
Intent library
Training set
Test set
MIntRec2.0
8,383
2,340
4,363
1,680
Weibo Post-Sync
3,184
775
1,923
486
Highlight-Intent
3,297
980
1,660
657
TABLE I: Dataset partitions after data cleaning.
Model
Highlight-Intent
MIntRec2.0
Weibo Post-Sync
Acc (%)
M-F1 (%)
W-F1 (%)
Acc (%)
M-F1 (%)
W-F1 (%)
Acc (%)
M-F1 (%)
W-F1 (%)
LLM and Discriminative Baselines
Gemini-3-flash (Direct)
32.42
16.27
29.35
41.43
29.68
43.19
46.30
41.56
47.63
Gemini-3-flash (Retrieval-Augmented)
40.18
18.70
35.42
46.13
31.96
47.12
67.70
44.49
62.08
Qwen3-30B-A3B (Direct)
29.22
13.82
26.33
38.93
26.06
39.59
42.80
32.15
37.88
Qwen3-30B-A3B (Retrieval-Augmented)
36.53
17.50
31.14
41.96
30.83
43.46
63.99
40.06
57.27
TABLE II: Overall context-aware intent-inference performance across three datasets. IntPro and Naive GRPO share the Direct-and-Retrieval SFT initialization; bold marks the best result within each backbone block.
Fig. 4: Four supervised fine-tuning formats across two models and three datasets. The Generalization Gap axis is inverted for radial scaling.
Dataset
Input Form
Intra-Sim ↑
Inter-Sim ↓
User LOO Acc. ↑
Global R@1 ↑
MIntRec2.0
Action Only
0.5218
0.4843
0.2295
0.3171
Full Context
0.5306
0.5239
0.0848
0.0940
Free-form Context-to-Intent Description
0.5979
0.5584
0.3386
0.4376
Intent explanation
0.6067
0.5598
0.3818
0.5051
Weibo Post-Sync
Action Only
0.4633
0.4045
0.3690
0.6438
Full Context
0.5581
0.5173
0.3781
0.6215
TABLE III: Comparison of retrieval representations for annotated historical records across datasets.
Fig. 5: t-SNE projections of Full Context (top) and the intent explanation (bottom) on MIntRec2.0, Weibo Post-Sync, and Highlight-Intent. Within each dataset, the two rows use the same sampled records; each panel is projected independently and is shown descriptively.
Condition
MIntRec2.0
Weibo Post-Sync
Highlight- Intent
Action Only
52.86
78.19
56.32
Full Context
27.14
76.95
56.62
Intent explanation
56.43
80.45
58.90
TABLE IV: Evaluating intent explanations as retrieval queries for intent inference.
Policy
Dataset
Acc (%)
M-F1 (%)
No Retrieval
Highlight-Intent
48.71
31.29
MIntRec2.0
67.32
56.21
Weibo Post-Sync
75.51
61.74
Always Retrieval
Highlight-Intent
45.36
28.04
MIntRec2.0
66.85
53.28
Weibo Post-Sync
77.16
64.05
TABLE V: Context-aware intent-inference performance under fixed and selective retrieval policies using Qwen3-4B.
Fig. 6: Per-class Accuracy of the Qwen3-4B student variants and Qwen3-30B-A3B (Direct) on the two least-frequent and two most-frequent MIntRec2.0 intents.
Quantity
Highlight- Intent
MIntRec 2.0
Weibo Post-Sync
Retrieval rate (%)
12.02
19.82
45.06
Direct-inference Acc. (%)
51.38
69.93
79.03
Retrieval-conditioned Acc. (%)
51.90
68.77
80.37
Rate-matched random (%)
48.30
67.23
76.26
Selective Retrieval (%)
51.45
69.70
79.63
Path oracle (%)
57.53
76.49
85.60
TABLE VI: Retrieval frequency and decision quality for Selective Retrieval using Qwen3-4B.
Fig. 7: Retrieval-choice dynamics of IntPro during GRPO training on MIntRec2.0. IntPro’s observed group-level decisions are compared with a rate-matched independent-choice reference. (a) Mean retrieval-choice entropy. (b) Distribution of the number of retrieval trajectories in each group.
Fig. 8: Paired examples from IntPro with the Qwen3-4B backbone across MIntRec2.0, Weibo Post-Sync, and Highlight-Intent. In each dataset, retrieval corrects a direct error in the upper case and changes a correct direct prediction into an error in the lower case.
Budget
Dataset
Acc (%)
M-F1 (%)
K=1
Highlight-Intent
47.34
31.06
MIntRec2.0
66.43
58.79
Weibo Post-Sync
79.22
63.88
K=2
Highlight-Intent
50.23
32.81
MIntRec2.0
68.27
59.04
Weibo Post-Sync
79.63
64.83
TABLE VII: Retrieval-budget sensitivity of Selective Retrieval based on Qwen3-4B.