SFL-MTSC: Leveraging Semantic Frame-Level Multi-Task Self-Consistency for Robust Multi-Intent Spoken Language Understanding

# SFL-MTSC: Leveraging Semantic Frame-Level Multi-Task Self-Consistency for Robust Multi-Intent Spoken Language Understanding
Source: [https://arxiv.org/html/2606.25552](https://arxiv.org/html/2606.25552)
Chen Chen

###### Abstract

Prompt\-based spoken language understanding \(SLU\) with large language models \(LLMs\) often suffers from inconsistent intent–slot structures due to decoding stochasticity, particularly in multi\-intent scenarios\. In view of this, we propose Semantic Frame\-Level Multi\-Task Self\-Consistency \(SFL\-MTSC\), a novel structured aggregation framework operating at the semantic frame level\. Instead of output\-level majority voting, SFL\-MTSC decomposes predictions into intent\-specific frames, applies domain–intent grouping and slot\-level clustering, and evaluates cluster reliability using path support scoring\. Reliable frames are retained and re\-integrated to form the final prediction\. Zero\-shot experiments on the MAC\-SLU benchmark dataset show improved slot F1 and overall accuracy over single\-path inference, while intent accuracy remains largely stable across most settings\.

###### keywords:

Spoken Language Understanding \(SLU\), Large Language Models, Multi\-Intent Self\-Consistency

## 1Introduction

Spoken Language Understanding \(SLU\) is a fundamental paradigm for task\-oriented spoken semantics extraction, widely applied in scenarios such as smart homes and in\-vehicle systems to interpret user spoken commands for executing downstream tasks\[tur2011spoken\]\. It typically consists of two core subtasks: intent detection and slot filling\[qinSurveySpokenLanguage2021\]\. For example, given the utterance \`\`I want to know the weather in Taipei'', intent detection identifies the overall semantic intent \(e\.g\.,GetWeather\), while slot filling extracts the associated arguments \(e\.g\.,city\_name=Taipei\)\. Recently, with the advancement of large language models \(LLMs\), prompt\-based methods relying on LLMs have emerged as a promising paradigm for SLU, enabling zero\-shot or few\-shot inference without task\-specific fine\-tuning\[panPreliminaryEvaluationChatGPT2023,gpt\-slu,croprompt,xingDXANetDualTaskCrossLingual2026,liHowChatGPTRobust,10447804,aroraIntentDetectionAge2024,Gao,Lib\]\.

However, prompt\-based SLU with LLMs often suffers from inconsistent intent–slot structures due to decoding stochasticity, particularly in multi\-intent scenarios where a single utterance may simultaneously express multiple intents across different domains\[macslu\]\. In such cases, different decoding runs may produce conflicting semantic frames, leading to unreliable predictions that are difficult to aggregate at the output level\.

Self\-consistency\[self\-consistency,taubenfeldConfidenceImprovesSelfConsistency2025,chenUniversalSelfConsistencyLarge,Ahmed,Nowak\]has been proposed to improve LLM reasoning by sampling multiple paths and aggregating via majority voting\. While prior work has explored its application to SLU\[croprompt,HIT\-SCIR\], these methods were not designed for multi\-intent settings with semantic frame\-structured predictions\. LLM\-as\-a\-judge\[gu2026survey,li\-etal\-2025\-generation\]offers an alternative, but requires additional LLM calls, carries hallucination risk\[limitationofLLMasajudge\], and operates at the output level, making fine\-grained frame\-level filtering infeasible\.

In this work, we proposeSemantic Frame\-Level Multi\-Task Self\-Consistency \(SFL\-MTSC\), a structured aggregation framework designed for robust multi\-intent SLU\. Instead of output\-level majority voting, SFL\-MTSC decomposes predictions into intent\-specific semantic frames, applies domain–intent clustering and slot\-level clustering, and evaluates cluster reliability using path support inspired by association rule mining\[zhaoAssociationRuleMining2003,ref1\]\. Reliable frames are retained and re\-integrated to form the final prediction\. Our main contributions are as follows:

- •We propose SFL\-MTSC, a frame\-level self\-consistency framework for multi\-intent SLU\. It works below the output level, which allows it to better remove false intents and noisy slot predictions\.
- •We introduce Hybrid Jaccard similarity for slot clustering\. It combines key\-based and value\-based matching, making the framework more robust when slot key names vary across reasoning paths\.
- •We conduct zero\-shot experiments on MAC\-SLU\[macslu\], a challenging Chinese multi\-intent SLU benchmark, across three model configurations spanning text\-only LLMs\[yang2025qwen3technicalreport\], ASR\+LLM pipelines\[radford2023robust,yang2025qwen3technicalreport\], and end\-to\-end large audio\-language models \(LALMs\)\[xu2025qwen25omnitechnicalreport\], demonstrating consistent improvements in overall accuracy and Slot F1\.

All of code and data used in this work will be publicly available at[https://github\.com/boyan1001/SFL\-MTSC](https://github.com/boyan1001/SFL-MTSC)\.

## 2Proposed Approach

In this section, we illustrate the workflow of SFL\-MTSC as Figure[1](https://arxiv.org/html/2606.25552#S2.F1)\. Unlike standard majority voting, which operates on the final output level, SFL\-MTSC decomposes each reasoning path into intent\-specific semantic frames and evaluates structural consistency at the frame level\. This allows the framework to selectively retain stable frames supported by multiple reasoning paths while discarding hallucinated or structurally conflicting ones, thereby producing more reliable multi\-intent predictions\.

![Refer to caption](https://arxiv.org/html/2606.25552v1/x1.png)Figure 1:Overview of the SFL\-MTSC framework\. Given an input utteranceXX, multiple reasoning paths are sampled to construct a frame pool\. Frames are then grouped via domain–intent clustering and slot\-level clustering, filtered by path support, and re\-integrated into the final multi\-intent predictionY∗Y^\{\*\}\.### 2\.1Multi\-Paths Reasoning

Given an input utteranceXX, we perform inference over the same LLM or LALM usingKKdistinct reasoning paths, each differing in sampling temperature\. Each semantic frameFFis a structured triple consisting of a domaindd, an intentii, and a set of slot–value pairsss:

F=\(d,i,s\),s=\{\(k1,v1\),\(k2,v2\),…\}F=\(d,i,s\),\\quad s=\\\{\(k\_\{1\},v\_\{1\}\),\(k\_\{2\},v\_\{2\}\),\\dots\\\}\(1\)The prediction of pathkkconsists of a set of such frames:

Y\(k\)=\{F1\(k\),F2\(k\),…,Fmk\(k\)\}Y^\{\(k\)\}=\\\{F^\{\(k\)\}\_\{1\},F^\{\(k\)\}\_\{2\},\\dots,F^\{\(k\)\}\_\{m\_\{k\}\}\\\}\(2\)wheremkm\_\{k\}is the number of frames predicted by pathkk\. Due to model stochasticity, different paths may produce conflicting intent–slot structures for the same utterance, particularly in multi\-intent scenarios\. The subsequent stages of SFL\-MTSC are designed to aggregate these diverse predictions and resolve such inconsistencies at the frame level\.

### 2\.2Frame Pool Construction

All semantic frames acrossKKreasoning paths are collected into a unified frame pool:

ℱ=⋃k=1KY\(k\)=\{Fj\(k\)∣k∈\{1,…,K\},j∈\{1,…,mk\}\}\\mathcal\{F\}=\\bigcup\_\{k=1\}^\{K\}Y^\{\(k\)\}=\\left\\\{F\_\{j\}^\{\(k\)\}\\mid k\\in\\\{1,\\dots,K\\\},\\ j\\in\\\{1,\\dots,m\_\{k\}\\\}\\right\\\}\(3\)Frames with missingdomainorintentfields are discarded prior to pooling\. Each retained frame is annotated with its source path indexppfor use in the subsequent self\-consistency scoring stage\.

### 2\.3Frame\-Level Clustering

Withℱ\\mathcal\{F\}constructed, we identify frames referring to the same underlying semantic intent via a coarse\-to\-fine clustering procedure\.

#### 2\.3\.1Domain–Intent Clustering

We first partitionℱ\\mathcal\{F\}into buckets by\(domain, intent\)pair:

ℱd,i=\{Fj\(k\)∈ℱ∣dj\(k\)=d,ij\(k\)=i\}\\mathcal\{F\}\_\{d,i\}=\\left\\\{F\_\{j\}^\{\(k\)\}\\in\\mathcal\{F\}\\mid d\_\{j\}^\{\(k\)\}=d,\\ i\_\{j\}^\{\(k\)\}=i\\right\\\}\(4\)This prevents frames with different intents from being merged and narrows the search space for slot\-level comparison\. In multi\-intent scenarios, the same intent label may appear in multiple frames within a single path, motivating the need for further slot\-level clustering within each bucket\.

#### 2\.3\.2Slot Clustering

Within each bucketℱd,i\\mathcal\{F\}\_\{d,i\}, we cluster frames by slot similarity using a threshold similarity graph whose connected components are taken as clusters, inspired by\[widdowsGraphModelUnsupervised2002\]\. The similarity between frames is measured by Hybrid Jaccard, inspired by the Jaccard index\[jaccard1901etude\], which interpolates Key–Value Jaccardsimk​v\\text\{sim\}\_\{kv\}and Value\-Based Jaccardsimv​a​l\\text\{sim\}\_\{val\}with coefficientα∈\[0,1\]\\alpha\\in\[0,1\], wherev​\(s\)v\(s\)denotes the set of all values in slot setss\.

simk​v​\(Fa,Fb\)=\|sa∩sb\|\|sa∪sb\|,simv​a​l​\(Fa,Fb\)=\|v​\(sa\)∩v​\(sb\)\|\|v​\(sa\)∪v​\(sb\)\|\\text\{sim\}\_\{kv\}\(F\_\{a\},F\_\{b\}\)=\\frac\{\|s\_\{a\}\\cap s\_\{b\}\|\}\{\|s\_\{a\}\\cup s\_\{b\}\|\},\\quad\\text\{sim\}\_\{val\}\(F\_\{a\},F\_\{b\}\)=\\frac\{\|v\(s\_\{a\}\)\\cap v\(s\_\{b\}\)\|\}\{\|v\(s\_\{a\}\)\\cup v\(s\_\{b\}\)\|\}\(5\)simh​y​b​\(Fa,Fb\)=α⋅simk​v​\(Fa,Fb\)\+\(1−α\)⋅simv​a​l​\(Fa,Fb\)\\text\{sim\}\_\{hyb\}\(F\_\{a\},F\_\{b\}\)=\\alpha\\cdot\\text\{sim\}\_\{kv\}\(F\_\{a\},F\_\{b\}\)\+\(1\-\\alpha\)\\cdot\\text\{sim\}\_\{val\}\(F\_\{a\},F\_\{b\}\)\(6\)simk​v\\text\{sim\}\_\{kv\}is structurally precise but sensitive to key naming variation;simv​a​l\\text\{sim\}\_\{val\}is more robust but may conflate frames with coincidentally overlapping values\. Hybrid Jaccard balances both\. We then construct a threshold similarity graphGd,i=\(ℱd,i,Ed,i\)G\_\{d,i\}=\(\\mathcal\{F\}\_\{d,i\},E\_\{d,i\}\):

Ed,i=\{\(Fa,Fb\)∣Fa,Fb∈ℱd,i,simh​y​b​\(Fa,Fb\)≥τ\}E\_\{d,i\}=\\left\\\{\(F\_\{a\},F\_\{b\}\)\\mid F\_\{a\},F\_\{b\}\\in\\mathcal\{F\}\_\{d,i\},\\ \\text\{sim\}\_\{hyb\}\(F\_\{a\},F\_\{b\}\)\\geq\\tau\\right\\\}\(7\)The connected components ofGd,iG\_\{d,i\}are taken as the final slot clusters, yielding the overall set of frame\-level clusters:

𝒞d,i=CC​\(Gd,i\),𝒞=⋃\(d,i\)𝒞d,i\\mathcal\{C\}\_\{d,i\}=\\mathrm\{CC\}\(G\_\{d,i\}\),\\quad\\mathcal\{C\}=\\bigcup\_\{\(d,i\)\}\\mathcal\{C\}\_\{d,i\}\(8\)where eachC⊆ℱd,iC\\subseteq\\mathcal\{F\}\_\{d,i\}represents a distinct semantic frame instance\.

MethodsText\-NLUASR \+ NLU \(Pipeline\)End\-to\-EndQwen3\-4B\-Instruct\-2507Whisper \+ Qwen3\-4B\-Instruct\-2507Qwen2\.5\-Omni\-7BOverall Acc\.Intent Acc\.Slot F1Overall Acc\.Intent Acc\.Slot F1Overall Acc\.Intent Acc\.Slot F1Vanilla Prompting\[macslu\]2\.0746\.8429\.392\.4343\.6322\.150\.6421\.464\.86\+SFL\-MTSC3\.30\(\+1\.23\)44\.22\(\-2\.62\)58\.25\(\+28\.86\)3\.39\(\+0\.4\)38\.31\(\-5\.32\)49\.53\(\+27\.38\)2\.09\(\+1\.45\)13\.90\(\-7\.56\)4\.57\(\-0\.29\)CroPropmt \(Intent→\\rightarrowSlot\)\[croprompt\]1\.5148\.8436\.931\.8143\.6823\.770\.4225\.305\.46\+SFL\-MTSC1\.91\(\+0\.40\)46\.13\(\-2\.71\)46\.24\(\+9\.31\)2\.17\(\+0\.36\)41\.44\(\-2\.24\)34\.48\(\+10\.71\)2\.17\(\+1\.75\)18\.59\(\-6\.71\)4\.79\(\-0\.67\)CroPropmt \(Slot→\\rightarrowIntent\)\[croprompt\]4\.1655\.1252\.704\.4950\.8639\.010\.9412\.140\.37\+SFL\-MTSC4\.52\(\+0\.36\)53\.26\(\-1\.86\)52\.23\(\-0\.47\)4\.69\(\+0\.20\)49\.61\(\-1\.25\)39\.54\(\+0\.53\)2\.43\(\+1\.49\)6\.78\(\-5\.36\)3\.93\(\+3\.56\)GPT\-SLU\[gpt\-slu\]4\.0758\.0549\.064\.2353\.5243\.160\.1814\.1211\.24\+SFL\-MTSC4\.10\(\+0\.03\)56\.82\(\-1\.23\)47\.33\(\-1\.73\)4\.26\(\+0\.03\)50\.13\(\-3\.39\)40\.02\(\-3\.14\)2\.13\(\+1\.32\)8\.16\(\-5\.96\)3\.91\(\-7\.33\)

Table 1:Main results for LLMs and LALMs on MAC\-SLU dataset by using different temperature and using SFL\-MTSC for each prompting methods\. ForVanilla Prompting, we follow MAC\-SLU methods\[macslu\]to directly utilize a sample single\-round prompting method for SLU\. Performance gains/drops compared to baseline are highlighted withblue/red\. The best results are shown inbold\.

### 2\.4Path Support Scoring

After clustering, each clusterC∈𝒞C\\in\\mathcal\{C\}represents a candidate semantic frame instance\. However, clusters vary in reliability: a single\-path cluster may reflect a hallucinated intent, while one with conflicting slot values indicates structural inconsistency\. Inspired by the support measure in association rule mining\[zhaoAssociationRuleMining2003,ref1\], we quantify cluster reliability by counting the number of distinct reasoning paths that contribute frames to the cluster:

supp​\(C\)=\|\{p∣∃Fj\(p\)∈C\}\|\\text\{sup\}\_\{p\}\(C\)=\\left\|\\\{p\\mid\\exists\\,F\_\{j\}^\{\(p\)\}\\in C\\\}\\right\|\(9\)
Unlike response probability in CISC\[taubenfeldConfidenceImprovesSelfConsistency2025\], which scores each reasoning path as a whole, our approach applies support\-based filtering at the slot\-cluster level, retaining only clusters backed by enough distinct paths\. This allows the framework to filter out hallucinated intents and noisy slot predictions at a finer granularity than path\-level scoring\.

### 2\.5Support\-Based Filtering

Based on the computed support score, we filter out unreliable clusters\. Similar to the path\-level filtering explored in reasoning\-aware self\-consistency frameworks\[wan\-etal\-2025\-reasoning\], we apply a minimum support threshold to discard clusters that lack sufficient cross\-path agreement\. Specifically, clusters are retained only if they satisfy:

𝒞∗=\{C∈𝒞∣supp​\(C\)≥⌈K2⌉\}\\mathcal\{C\}^\{\*\}=\\left\\\{C\\in\\mathcal\{C\}\\mid\\text\{sup\}\_\{p\}\(C\)\\geq\\left\\lceil\\frac\{K\}\{2\}\\right\\rceil\\right\\\}\(10\)The support criterion discards hallucinated intents supported by fewer than half the paths\.

### 2\.6Re\-Integration

To re\-integrate the clusters into the frames of the final prediction, we directly use the frame domain and intent\. For slots, we propose a Value\-First re\-integration strategy inspired by Token\-level Self\-consistency from CroPrompt\[croprompt\], where representative slot values are first identified by support, and the corresponding slot keys are then determined by majority vote\.

We compute thevalue support score, which counts the number of distinct frames containing valuevvat least once\. Then we filter out the values less than half frames support\.

supV​\(v\)=\|\{F∈C∣∃k:\(k,v\)∈F\}\|\\mathrm\{sup\}\_\{V\}\(v\)=\\left\|\\left\\\{F\\in C\\mid\\exists\\,k:\(k,v\)\\in F\\right\\\}\\right\|\(11\)V∗=\{v∣supV​\(v\)≥⌈\|C\|2⌉\}V^\{\*\}=\\left\\\{v\\mid\\mathrm\{sup\}\_\{V\}\(v\)\\geq\\left\\lceil\\frac\{\|C\|\}\{2\}\\right\\rceil\\right\\\}\(12\)
In the second step, we determine the corresponding slot keys by majority vote for each retained valuev∈V∗v\\in V^\{\*\}:

k∗​\(v\)=arg⁡max𝑘​∑F∈C𝟙​\(\(k,v\)∈F\)k^\{\*\}\(v\)=\\underset\{k\}\{\\arg\\max\}\\sum\_\{F\\in C\}\\mathds\{1\}\(\(k,v\)\\in F\)\(13\)
A representative frameF∗F^\{\*\}is then produced by fixing domain and intent from the bucket key, with the final slot set constructed as:

F∗=\(d,i,s∗\),s∗=\{\(k∗​\(v\),v\)∣v∈V∗\}F^\{\*\}=\(d,\\ i,\\ s^\{\*\}\),\\quad s^\{\*\}=\\\{\(k^\{\*\}\(v\),v\)\\mid v\\in V^\{\*\}\\\}\(14\)The final multi\-intent prediction is:

Y∗=\{FC∗∣C∈𝒞∗\}Y^\{\*\}=\\\{F^\{\*\}\_\{C\}\\mid C\\in\\mathcal\{C\}^\{\*\}\\\}\(15\)

### 2\.7Handling Empty Semantics

Reasoning paths may occasionally produce malformed or empty outputs\. Frames with missingdomainorintentfields are discarded prior to frame pool construction\. If the resulting frame pool is empty, the system returns an empty predictionY∗=∅Y^\{\*\}=\\emptyset\.

## 3Experiments

### 3\.1Experiment Setting

#### 3\.1\.1Datasets

We evaluate on MAC\-SLU\[macslu\], a Chinese multi\-intent SLU dataset for automotive cabin scenarios, spanning 8 domains, 81 intents, and 192 slot types with up to 4 simultaneous intents per utterance\. Semantic annotations are structured as semantic frames, consistent with the frame\-level design of SFL\-MTSC\.

#### 3\.1\.2Models

We evaluate three model configurations: \(1\) Qwen3\-4B\-Instruct\[yang2025qwen3technicalreport\]for text\-based inference, \(2\) a pipeline system combining Whisper\-Large\-V3\-Turbo\[radford2023robust\]\(CER = 12\.83% on the MAC\-SLU test set\) for ASR with Qwen3\-4B\-Instruct for NLU, and \(3\) Qwen2\.5\-Omni\-7B\[xu2025qwen25omnitechnicalreport\]as an end\-to\-end LALM for direct speech\-to\-semantics inference\.

#### 3\.1\.3Implementation Details

All experiments were conducted on an NVIDIA Titan RTX GPU\. For LLM inference, we employed vLLM\[kwonEfficientMemoryManagement2023\]for accelerated deployment\. For Qwen2\.5\-Omni\-7B, we used vLLM\-Omni\[yinVLLMOmniFullyDisaggregated2026\], an inference framework that extends vLLM with support for omni\-modal language models\.

#### 3\.1\.4Baselines

We compare SFL\-MTSC against three prompting baselines: Vanilla Prompting\[macslu\], which directly extracts intents and slots in a single prompt; CroPrompt\[croprompt\], which first performs intent detection and then leverages the results for slot filling; and GPT\-SLU\[gpt\-slu\], which jointly predicts intents and slots in the first stage and refines predictions via cross\-task information exchange in the second stage\. Since CroPrompt and GPT\-SLU were originally designed for zero\-shot single\-intent scenarios, we extend their prompt templates to support multiple semantic frames per utterance\.

#### 3\.1\.5Evaluation Metric

We followed the standard metrics for SLU tasks\[gpt\-slu\]\. For intent detection, we calculated accuracy, and for slot filling, we calculated the F1 score\. The Overall Acc\. measures the proportion of utterances for which both intent detection and slot filling are simultaneously correct, serving as the strictest end\-to\-end evaluation criterion\.

### 3\.2Main Results

All experiments are conducted in a zero\-shot setting, where no task\-specific training or fine\-tuning is performed\. The main results are shown in Table[1](https://arxiv.org/html/2606.25552#S2.T1)\. For each prompting method, we generate K = 5 reasoning paths at sampling temperaturesT∈\{0,0\.3,0\.5,0\.7,1\.0\}T\\in\\\{0,0\.3,0\.5,0\.7,1\.0\\\}\. Each path produces an independent set of semantic frame predictions for the same input utterance\. The five predictions are then aggregated by SFL\-MTSC, which performs frame\-level clustering, path support scoring, filtering, and re\-integration to produce the final multi\-intent predictionY∗Y^\{\*\}\. In this work, we set the hyperparameters as follows: similarity thresholdτ=0\.55\\tau=0\.55and hybrid Jaccard mixing coefficientα=0\.3\\alpha=0\.3, which balances Overall Accuracy and Slot F1 performance\.

Our observations are as follows:

\(1\) SFL\-MTSC consistently improves Overall Acc\. across all zero\-shot settings,demonstrating the effectiveness of frame\-level aggregation as a general post\-inference strategy that requires no additional supervision or fine\-tuning\.

\(2\) SFL\-MTSC yields the largest gains under zero\-shot Vanilla Prompting\.We observe improvements of 1\.23%, 0\.4%, and 1\.45% in Overall Acc\. for the three respective models\. This is because Vanilla Prompting provides the least structural guidance, leading to higher cross\-path variability in zero\-shot decoding, which SFL\-MTSC can effectively resolve through frame\-level aggregation\.

\(3\) Gains diminish for prompting methods with stronger structural guidance\.For CroPrompt and GPT\-SLU, which impose more structured zero\-shot prompting strategies, Overall Acc\. gains are more modest, indicating that frame\-level filtering may over\-regularize already stable predictions\.

\(4\) SFL\-MTSC regularizes slot structure more effectively than intent classification\.Across all zero\-shot configurations, Slot F1 shows more consistent improvements than Intent Acc\., which fluctuates within 1–7% and occasionally decreases slightly\. This suggests that in zero\-shot inference, cross\-path inconsistency primarily manifests at the slot level, while intent predictions tend to be relatively stable across reasoning paths even without aggregation\.

Table 2:Ablation results on support\-based filtering configurations\. All experiments use Qwen3\-4B\-Instruct\-2507\[yang2025qwen3technicalreport\]with vanilla prompting\[macslu\]\. The best results are shown inbold\.
### 3\.3Effect of Support\-Based Filtering

To examine the effect of support\-based filtering, we conduct an ablation study varying the placement of the support filter across two levels: domain–intent and slot cluster\. We evaluate four configurations: no filtering, intent\-level only, slot\-level only, and both levels combined, and intent\-level filtering is installed between domain\-intent clustering and slot clustering\. All experiments are conducted using Qwen3\-4B\-Instruct\-2507 with vanilla prompting\.

The results are illustrated in Table[2](https://arxiv.org/html/2606.25552#S3.T2)\. We observe that the slot\-level filter alone is sufficient to achieve the highest Overall Acc\. of 3\.30% and Intent Acc\. of 44\.22%, with the combined configuration leads to identical results, indicating that the slot\-level filter is the primary driver of performance gains, while adding intent\-level filtering alone contributes only marginal improvement\. Furthermore, removing all filtering leads to a substantial drop of 13\.03% in Intent Acc\., demonstrating the necessity of support\-based filtering for reliable multi\-intent prediction\.

Table 3:The results on using different Hybrid Jaccard mixing coefficient\. All experiments use Qwen3\-4B\-Instruct\-2507\[yang2025qwen3technicalreport\]with vanilla prompting\[macslu\]\. The best results are shown inbold\.
### 3\.4Different Hybrid Jaccard Mixing Coefficient

In slot clustering, we use Hybrid Jaccard with mixing coefficientα\\alphato interpolate between Key–Value Jaccardsimk​v\\mathrm\{sim\}\_\{kv\}and Value\-Based Jaccardsimv​a​l\\mathrm\{sim\}\_\{val\}\(see Equation \(6\)\)\. To examine the effect ofα\\alpha, we evaluate a range of valuesα\\alphaon the MAC\-SLU test set using Qwen3\-4B\-Instruct\-2507 with vanilla prompting\.

The results are shown in Table[3](https://arxiv.org/html/2606.25552#S3.T3)\. We observe that Overall Acc\. remains stable at 3\.30% across a wide range ofα∈\[0\.0,0\.7\]\\alpha\\in\[0\.0,0\.7\], and only begins to decline atα=0\.9\\alpha=0\.9, indicating that Overall Accuracy is largely insensitive to the mixing coefficient\. For Slot F1, the best performance of 58\.47% is achieved atα=0\.1\\alpha=0\.1, with a decrease asα\\alphaincreases\. This suggests that slot clustering benefits more from value\-based similarity than from key\-value\-based similarity\.

Based on these observations, we selectα=0\.3\\alpha=0\.3as the default setting, which maintains peak Overall Acc\. while preserving competitive Intent Acc\. and Slot F1 performance\.

## 4Conclusion

We proposed SFL\-MTSC, a semantic frame\-level self\-consistency framework for robust multi\-intent SLU\. It decomposes predictions into intent\-specific frames, applies Hybrid Jaccard slot clustering, and filters unreliable frames via path support scoring\. Zero\-shot experiments on MAC\-SLU show consistent improvements in Overall Acc\. and Slot F1, with the largest gains under Vanilla Prompting \(Slot F1 up to \+28\.86%\)\. Limitations include occasional drops in Intent Accuracy, limited gains in LALM settings due to high decoding variance, and evaluation on a single dataset\. Future work will explore finer\-grained intent clustering and aggregation mechanisms better suited for LALMs\.

## 5Acknowledgments

This work was supported in part by Realtek Semiconductor Corporation under Grant Numbers 113KK01103 and 114KK01005\. Any findings and implications in the paper do not necessarily reflect those of the sponsors\.

## 6Generative AI Use Disclosure

We used Claude Sonnet 4\.6 to assist with editing and polishing the manuscript\. We also used AI tools to assist with code development, with strict human review and verification to ensure correctness\.

## References