Understanding a conversation requires connecting what is said to the speakers' relationships, expectations, and earlier exchanges. Prior work shows that supplying relevant context can improve models' social judgments, but much of this information may already be present in the conversation. This motivates training models to discover the cues that matter and use them to support an interpretation. We propose CLUE, a two-stage training pipeline that combines supervised evidence reasoning with GRPO guided by Grounded Information Utilization (IUG). This reward credits new, evidence-supported social inferences while accounting for reasoning effort. We train on a small set of conversational examples to assess effectiveness and generalization across tasks, languages, and genres. At inference time, the trained model reasons over the supplied interaction in a self-contained manner. Compared with base Qwen, CLUE achieves relative gains of 17.3% in-domain and 9.3% out-of-domain in average weighted F1, with higher F1 on four of five out-of-domain evaluations. These gains are accompanied by over 80% fewer reasoning tokens on average in both domains. Human evaluation further finds that CLUE more often grounds its explanations in specific conversational content, with closer alignment between its quoted evidence and human-selected context.
Figures & tables
Figure 1: The model trained with CLUE correctly identifies surprise in this real conversation excerpt by connecting the earlier playful expression ( E1 ) with the later gift discussion ( E2 ); the base model predicts anger. English translations appear in gray italics; key turns are shown in Chinese and English.
Figure 2: Stage 1 of CLUE : supervised initialization using refined reasoning traces and two added evidence tokens. The resulting model undergoes GRPO training in stage 2, illustrated in Figure 3 .
Figure 3: Stage 2 of CLUE : GRPO training with task correctness and IUG . The illustrated trace receives credit for two new, evidence-supported inferences, which contribute to its completion-level reward. Verified quotation tokens receive a bounded exemption from the effort cost.
In-Domain ZH
Out-of-Domain
Model
Emotion
Dialog Act
Norm Violation
Japanese Emo ∗
English Emo †
Chinese Emo ‡
Chinese Dialog Act ‡
English Power Pred.
Avg.
Qwen3.5 (9B)
43.46 ± 0.77
69.14 ± 1.48
89.34 ± 0.82
45.55 ± 1.11
56.47 ± 1.00
18.38 ± 0.41
52.44 ± 0.30
68.00 ± 0.50
55.35
+ Label SFT
58.02 ± 1.05
66.17 ± 1.03
85.83 ± 0.69
55.61 ± 0.94
56.05 ± 0.88
16.70 ± 0.75
41.78 ± 1.37
55.61 ± 1.50
54.47
+ Reasoning SFT
55.09 ± 0.75
67.72 ± 0.64
96.34 ± 0.15
56.12 ± 0.75
54.01 ± 1.22
18.61 ± 0.49
38.91 ± 0.14
67.62 ± 0.89
56.80
Accuracy-Only RL
60.98 ± 0.18
72.33 ± 1.25
97.99 ± 0.08
62.59 ± 0.94
58.75 ± 1.17
20.77 ± 1.24
43.75 ± 0.41
67.67 ± 0.41
60.60
CLUE
64.83 ± 0.55
73.81 ± 0.92
98.15 ± 0.00
66.47 ± 0.92
59.40 ± 0.42
21.69 ± 0.50
47.37 ± 0.32
68.35 ± 1.09
62.51
Table 1: Weighted F1 across social and pragmatic tasks (mean ± sample SD). All entries are in percent. Best performance per group in bold ; second best underlined . ∗ : CCU Japanese; † : MELD; ‡ : CPED.
Figure 4: Relative gains in average weighted F1 and reasoning token savings over base Qwen. Tasks are weighted equally within each domain. Values are percentages.
Figure 5: Mean completion length during RL for CLUE and Accuracy-Only RL.
Task
CLUE − Base Qwen
CLUE − Reasoning SFT
Japanese Emo ∗
+20.92
+10.36
English Emo †
+2.93
+5.39
Chinese Emo ‡
+3.31
+3.08
Chinese Dialog Act ‡
−5.07
+8.46
English Power Pred.
+0.35
+0.73
Table 2: Mean weighted F1 differences on out-of-domain tasks. Values are in F1 points. ∗ : CCU Japanese; † : MELD; ‡ : CPED.
CLUE − Accuracy-Only RL
Task
Weighted F1
Macro F1
Emotion ZH
+3.85
+3.97
Dialog Act ZH
+1.48
+1.35
Norm Violation ZH
+0.16
+0.08
Japanese Emo ∗
+3.88
+2.87
English Emo †
+0.65
+1.63
Table 3: Mean F1 differences. Values are in F1 points. ∗ : CCU Japanese; † : MELD; ‡ : CPED.
Model
Macro F1
Qwen3.5 (9B)
69.64
+ Label SFT
68.54
+ Reasoning SFT
69.79
Accuracy-Only RL
69.90
CLUE
70.64
Table 4: Macro F1 on MMLU (%) using greedy decoding.
Figure 6: Cause identification and weighted emotion F1 on emotion-cause dataset, averaged over three sampled evaluations. Brackets show absolute gains.
Figure 7: Rates of grounded excerpts, Evidence Alignment F1, and final-answer accuracy on 50 paired instances. Values are percentages.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Setting
SFT
RL
LoRA rank / scaling
32 / 64
32 / 32
LoRA dropout
0.05
0
Peak learning rate
5×10−5
5×10−6
Gradient clipping
0.3
1.0
Token limits
32,768 total
6,144 prompt 5,120 output
Appendix
Table 5: Training settings for supervised initialization and RL.
Parameter
Value
Label-error factor γ
0.3
Quotation allowance Qmax
200 tokens
Saturation scale κ
3
Effort budget B
795 tokens
Effort penalty λ
0.5
Reward weight wIU
0.2
Appendix
Table 6: Reward parameters for CLUE .
Parameter
Value
Temperature
1.0
Top- p
1.0
Top- k
Disabled
Min- p
0
Presence penalty
0
Repetition penalty
1.0
Appendix
Table 7: Sampling parameters for social and pragmatic tasks.
Dataset
Task
Examples
CCU Mandarin
Emotion
1,094
Dialog Act
1,189
Norm Violation
1,212
CCU Japanese
Emotion
1,121
MELD
Emotion
1,086
CPED
Emotion
1,419
Appendix
Table 8: Evaluation set sizes. MMLU counts questions, power prediction counts conversations, and other tasks count target turns.
Figure 8: Annotation interface for selecting conversational evidence and rating anonymized reasoning excerpts.
Pragmatic reasoning in language models studies how models infer and communicate meanings shaped by context, intent, shared knowledge, and social norms.