LingXi: Diverse Research Applications

• LingXi is a term used for diverse research systems, including psychiatric diagnostics, Chinese poetry generation, software issue resolution, and adaptive video streaming.
• Each system adopts domain-specific methods—such as synthetic dialogue generation, diversity-aware decoding, procedural knowledge retrieval, or dynamic ABR optimization—to address its unique challenges.
• The artifacts reveal methodological diversity and practical insights, ranging from research benchmarks with clinical anchors to production systems optimized through real-user feedback.

Searching arXiv for the cited Lingxi papers and related entries. arxiv_search query: "Lingxi" LingXi, or Lingxi, is not a single model or platform in the arXiv literature. The name has been used for at least four distinct research artifacts: a Chinese psychiatric-diagnosis benchmark centered on LingxiDiagBench and LingxiDiag-16K; a diversity-aware Chinese modern poetry generation system; a repository-level issue resolution framework for software engineering; and a production system for personalized adaptive video streaming. These works are methodologically unrelated, but each uses the name for a technically specific system with its own datasets, objectives, and evaluation regime (Xu et al., 10 Feb 2026, Zhang et al., 2021, Yang et al., 13 Oct 2025, Jia et al., 22 Aug 2025).

1. Disambiguation and scope

The major arXiv usages of the name are summarized below.

Usage of LingXi	Domain	Core artifact
LingxiDiagBench / LingxiDiag-16K	Chinese psychiatric consultation and diagnosis	Multi-agent benchmark and synthetic dialogue dataset
Lingxi	Chinese modern poetry generation	GPT-2-like system with NS-RH decoding and semantic rejection sampling
Lingxi	Repository-level issue resolution	Procedural-knowledge-guided framework for fixing software issues
LingXi	Adaptive video streaming	Personalized objective-adjustment system for ABR

A recurring misconception is to treat LingXi as a single general-purpose model family. The published work does not support that interpretation. In one paper, “LingXi” refers to a psychiatric benchmarking effort centered on LingxiDiagBench and LingxiDiag-16K; in another, it denotes a poetry generation system; in a third, it denotes a software issue resolution framework; and in a fourth, it denotes a deployed personalized streaming system. The shared name therefore identifies different research programs rather than a unified architecture.

2. LingxiDiagBench: psychiatric consultation and diagnosis in Chinese

In psychiatric AI, “LingXi” refers to a benchmarking effort built around two artifacts: LingxiDiagBench, the evaluation framework, and LingxiDiag-16K, a dataset of 16,000 synthetic consultation dialogues with synthetic EMRs designed for Chinese-language, ICD-10-aligned psychiatric assessment (Xu et al., 10 Feb 2026). The stated target is substantially narrower than generic medical QA: models are evaluated on whether they can ask clinically useful questions, gather missing information over multiple turns, and then make diagnostic judgments under outpatient psychiatric conditions.

The benchmark is grounded in a real source dataset, LingxiDiag-Clinical, derived from about 4,500 outpatients seen at the Shanghai Mental Health Center (SMHC) between 2023 and 2024. After preprocessing, de-identification, automatic transcription, and manual verification, the authors curated 1,709 cases with well-annotated EMRs and verified transcripts. The synthetic benchmark is then generated from this real distributional anchor. Each synthetic case contains structured psychiatric record fields aligned to outpatient EMRs, including demographics, chief complaint, present illness history, past medical and psychiatric history, family history, and diagnostic conclusions.

The synthetic generation pipeline is explicitly distribution-matching. The authors extract demographic and clinical features from the real cases to build a knowledge graph, then generate synthetic EMRs by sampling from this graph according to the empirical distribution in the real data. Reported demographic matching includes 32.3% male / 67.7% female in the clinical data versus 32.4% male / 67.6% female in LingxiDiag-16K, and the appendix uses Jensen-Shannon divergence as evidence of high distributional similarity. The major diagnosis proportions are also closely matched, including F32 depressive episode 35.2% vs 34.8%, F41 anxiety disorders 22.1% vs 22.5%, F43 stress-related 12.3% vs 12.1%, F31 bipolar disorder 8.7% vs 8.9%, F42 OCD 5.4% vs 5.6%, F20 schizophrenia 4.8% vs 4.5%, F45 somatoform 3.9% vs 4.1%, F51 sleep disorders 3.2% vs 3.3%, F39 unspecified mood 2.1% vs 2.0%, F98 childhood-onset 1.5% vs 1.4%, and Others 0.8% vs 0.8%.

The benchmark’s full 12-way diagnosis task uses F20, F31, F32, F39, F41, F42, F43, F45, F51, F98, Z71, and Others. It is implemented as a multi-agent system with a Patient Agent, Doctor Agent, Diagnosis Agent, and LLM-as-a-Judge evaluators. In static diagnostic inference, the model receives a complete consultation transcript and predicts diagnoses; this isolates diagnostic reasoning over a fixed information set. In dynamic multi-turn consultation, the model itself plays the Doctor Agent and must decide what to ask next, when to stop, and what diagnosis to issue.

A major contribution is the patient simulator, LingxiDiag-Patient. It was designed because off-the-shelf LLM patient simulation was judged unrealistic in three ways: responses were too long, symptoms were dumped all at once, and language was too polished. The prompt therefore enforces brief, colloquial, emotionally restrained, fact-based answers together with a “Response Length Requirement.” On the doctor side, four interview strategies are supported: Free-form, Symptom-Tree, APA-Guided, and APA-Guided + MRD-RAG. The APA-guided scaffold has five phases: Screening (3–5 questions), Assessment (5–8 questions), Deep-dive (3–6 questions), Risk Assessment (2–4 questions), and Closure (1–2 questions).

The reported findings are clinically narrow but methodologically important. In static evaluation on LingxiDiag-16K, the best 2-class result is 0.854 accuracy by Gemini-3-Flash, but the 4-class task is led by traditional lexical baselines, including TF-IDF + RF at 0.479 accuracy and TF-IDF + LR at 0.458 macro-F1 / 0.480 weighted-F1. In the 12-class task, GPT-5-Mini reaches 0.409 accuracy, while the best macro-F1 among all methods is TF-IDF + LR with 0.295. On the real LingxiDiag-Clinical dataset, Qwen3-32B achieves the highest aggregate “Overall” score at 0.548, but simple lexical methods remain competitive. In dynamic evaluation, the strongest strategy is APA-Guided + MRD-RAG. Under that setting, DeepSeek-V3.2 reaches 92.3% accuracy on the 2-class task, while Grok-4.1-Fast reaches 43.0% accuracy on the 4-class task and 28.5% accuracy on the 12-class task. The benchmark’s central conclusion is that dynamic consultation often underperforms static evaluation: models can frequently reason from a complete transcript better than they can obtain that transcript themselves. Consultation quality and diagnostic accuracy are only moderately correlated, with reported correlation $r = 0.43$. The paper explicitly warns that the benchmark is “designed exclusively for research purposes” and “is not a clinical diagnostic tool.”

3. Lingxi as a Chinese modern poetry generation system

In natural language generation, “Lingxi” denotes a Chinese modern poetry generation system built around a GPT-2-like LLM backbone, a diversity-aware decoding method called NS-RH (“nucleus sampling with randomized head”), and a semantic-similarity-based rejection sampling stage for topic preservation (Zhang et al., 2021). The system addresses a task formulation in which the input is often only a short poetry title, so the generator must create a full poem from minimal context while avoiding repetition, generic over-fluent text, and topic drift.

The backbone is an auto-regressive Transformer decoder with 24 Transformer layers, 16 attention heads, embedding size 1,024, maximum context length 128, and 330 million parameters, trained on 8 NVIDIA GeForce RTX 2080 Ti GPUs. Training proceeds in two stages: pre-training on about 3,500 published books of Chinese novels, then fine-tuning on about 220,000 passages of Chinese modern poetry and lyrics, yielding the model called GPT-2 LyricCN. The preprocessing pipeline uses THULAC for Chinese Word Segmentation followed by vocabulary pruning and subword decomposition. The reported vocabulary sizes are 12,643 after the initial merge, 23,296 after subword decomposition, and 17,589 after final pruning.

The system encodes structural information using special tokens such as [TITLE], [START-OF-PASSAGE], [NEWLINE], and [END-OF-PASSAGE]. Training uses Adam, weight decay 0.01, linear decay, and 1% of total steps as warm-up. The best fine-tuning result is reported at epoch 4 with learning rate $2\times 10^{-5}$, and the validation perplexity moves from 17.58 after pre-training to 16.75 for the best fine-tuned model.

The core technical proposal is NS-RH. The paper defines the randomized head as

$V^{head} = NS(q)\cap top1ctrl(n),$

with

$top1ctrl(n)=\big\{x\mid p(x)\ge\max{p(x)}/n\big\}.$

Here $NS(q)$ is the nucleus set under threshold $q$, and $top1ctrl(n)$ keeps tokens whose probability is at least $\max p(x)/n$. Lingxi then evenly redistributes the probability mass over $V^{head}$, while still performing ordinary nucleus filtering with parameter $p$, where $2\times 10^{-5}$0. The stated intuition is that in poetry the source of low diversity is not only the low-probability tail but also the over-dominant high-probability head. In the appendix, the reported setting for NS-RH is $2\times 10^{-5}$1, $2\times 10^{-5}$2, with tunable $2\times 10^{-5}$3.

The second major component is semantic-similarity-based rejection sampling. The system generates $2\times 10^{-5}$4 short hypotheses of length $2\times 10^{-5}$5, scores them by semantic similarity to the title, accepts the most similar one, and then continues poem generation from that accepted prefix. In the appendix, the reported parameters are $2\times 10^{-5}$6 and $2\times 10^{-5}$7. Sentence embeddings are formed by averaging hidden states from the last 12 out of 24 layers and applying standard normalization calibration, and similarity is measured by cosine distance.

Evaluation uses both automatic and human judgments. Automatic metrics include Perplexity (PPL), Self-BLEU, Zipf coefficient, repetition entropy

$2\times 10^{-5}$8

rhyming entropy

$2\times 10^{-5}$9

and line length. The system generates 5,000 samples per method for metric computation. Human-system comparison uses 50 poems from each system under the title “spring water,” with 103 annotations from 9 annotators per system.

The main result is that NS-RH substantially increases novelty while preserving acceptable fluency. Representative settings include NS-RH, $V^{head} = NS(q)\cap top1ctrl(n),$0 with PPL 8.82 and Self-BLEU-4 0.50, NS-RH, $V^{head} = NS(q)\cap top1ctrl(n),$1 with PPL 28.71 and Self-BLEU-4 0.35, and NS-RH, $V^{head} = NS(q)\cap top1ctrl(n),$2 with PPL 41.54, Self-BLEU-4 0.30, and Self-BLEU-5 0.17. In human evaluation, Lingxi ($V^{head} = NS(q)\cap top1ctrl(n),$3) scores 4.43 for fluency, 4.31 for novelty, 4.23 for coherence, and 4.32 overall, compared with Youling at 4.24 overall and XiaoIce at 4.23 overall. The paper’s main claim is therefore not about likelihood optimization but about reshaping the probability distribution so that comparatively lower-frequency but still plausible words can compete during decoding.

4. Lingxi as a repository-level issue resolution framework

In software engineering, “Lingxi” refers to a framework for repository-level issue resolution that extracts transferable procedural knowledge from historical issue-fixing data and uses it to guide LLM agents on new issues (Yang et al., 13 Oct 2025). The motivating claim is that existing LLM software agents can navigate code and generate patches, but they still struggle when success depends on architecture, execution flow, design conventions, and implementation constraints across an entire repository.

The framework separates offline knowledge construction from online issue resolution. In the offline stage, Lingxi reconstructs issue-specific procedural knowledge by reverse reasoning over historical issue reports, patches, and repository snapshots, then abstracts that knowledge into transferable form. In the online stage, it summarizes the target issue, retrieves similar historical issues, runs multiple knowledge-guided analysis agents in parallel, synthesizes the resulting analyses, and then passes the synthesis to planning and fixing agents.

The reconstructed knowledge schema contains 14 information types. The issue understanding knowledge category contains relevant architecture, dependency analysis, call trace, bug categorization, example usage context, concrete root cause analysis, current vs. expected behavior, and feature or functionality. The issue fixing knowledge category contains fix pattern, concrete fix steps, fix checklist, test case, resolution conclusion, and design patterns / code practices. The paper later identifies design patterns / code practices as the most critical knowledge aspect overall.

Retrieval is context-aware. Each issue is summarized into four elements: the core problem, key technical details and error messages, expected vs. actual behavior, and relevant code components. Historical summaries are embedded with Qwen3-Embedding-8B, the top 20 nearest issues are retrieved by cosine similarity, they are reranked with Qwen3-Reranker-4B, and the top 3 are retained. Only issues created before the target issue are eligible, which avoids temporal leakage. Online scaling is applied to the analysis phase rather than to blind patch generation: each retrieved knowledge item drives one parallel analysis trajectory, and the analyses are later synthesized under four criteria, Supportive, Complementarity, Accuracy, and Completeness.

All agents use a ReAct-style architecture. Available tools are view_directory, view_file_content, search_files_by_keywords via ripgrep, and str_replace_editor. Context management compresses inter-agent communication by retaining observations and reasoning while discarding verbose tool outputs.

The main benchmark is SWE-bench Verified, a curated subset of 500 GitHub issues drawn from the original SWE-bench dataset of 2,294 issues, spanning 12 open-source projects. The main metric is Pass@1. On this benchmark, Lingxi + Claude 4 Sonnet achieves 74.6% Pass@1, compared with 70.40% for OpenHands + Claude 4 Sonnet, 70.40% for Augment Agent v1 + Claude 4 Sonnet, 70.08% for SWE-search + Claude 4 Sonnet, and 66.60% for SWE-agent + Claude 4 Sonnet. With weaker backbones, the framework still improves results: GPT-4.1 mini rises from 23.94% under mini-SWE-agent to 45.20% under Lingxi, and Kimi-K2 rises from 65.40% under OpenHands to 70.33% under Lingxi.

The ablation study, run on 218 instances at a 95% confidence level with a 5% margin of error, is central to the paper’s interpretation. The full system achieves 74.80%, compared with 72.48% for $V^{head} = NS(q)\cap top1ctrl(n),$4, 68.30% for $V^{head} = NS(q)\cap top1ctrl(n),$5, 67.74% for $V^{head} = NS(q)\cap top1ctrl(n),$6, and 67.43% for $V^{head} = NS(q)\cap top1ctrl(n),$7. The framework therefore attributes most of its gains to transferable procedural knowledge rather than to scaling alone. The paper also reports that, among 20 manually inspected success cases, the qualitative coding agreement reaches Cohen’s kappa = 0.66. Limitations remain substantial: even with Lingxi, 25.4% of SWE-bench Verified instances remain unresolved, and the failure analysis identifies Incomplete fixes (6/20), Incorrect fault localization (5/20), Semantic errors (5/20), Overfitting patches (2/20), and Side-effect errors (2/20).

5. LingXi as a personalized adaptive video streaming system

In multimedia systems, LingXi denotes a production system for user-level personalized adaptive video streaming deployed at Kuaishou (Jia et al., 22 Aug 2025). Its premise is that traditional optimization around system-wide QoS metrics such as bitrate, stall time, and smoothness has approached practical saturation in mature large-scale streaming systems, and that further gains require aligning optimization with actual user engagement.

The paper formalizes the conventional objective as

$V^{head} = NS(q)\cap top1ctrl(n),$8

but reports that in a production 5-day A/B test the differences among three QoS-optimized production variants were small: video quality improvement space around 0.5%, stall reduction around 2%, and no statistically significant difference in total watch time. In Kuaishou’s production environment, only about 10% of users have average bandwidth below the maximum video bitrate requirement, over 90% have stall-free playback in a day, and over 99% experience no more than two stall events per day. This motivates a personalized rather than system-wide approach.

LingXi uses segment-level exit rate as its operational proxy for user-level QoE. The paper’s empirical analysis reports markedly different effect scales for standard QoS components: video quality affects exit rate at around $V^{head} = NS(q)\cap top1ctrl(n),$9, video smoothness at around $top1ctrl(n)=\big\{x\mid p(x)\ge\max{p(x)}/n\big\}.$0, and stall time at around $top1ctrl(n)=\big\{x\mid p(x)\ge\max{p(x)}/n\big\}.$1. The deployed system therefore uses a hybrid exit-rate predictor in which stall-related effects are personalized and quality/smoothness effects are handled statistically: $top1ctrl(n)=\big\{x\mid p(x)\ge\max{p(x)}/n\big\}.$2 The predictor is trained with cross-entropy,

$top1ctrl(n)=\big\{x\mid p(x)\ge\max{p(x)}/n\big\}.$3

and uses five input dimensions: Bitrate, Throughput, Past stall time, Last stall interval, and Last stall-exit interval. These are stored in a matrix of length 8. The personalized stall predictor is a CNN with five 1D-CNN layers, each with 64 channels, followed by a 64-dimensional fully connected layer, a 2-dimensional fully connected layer, and a final softmax.

Training uses around 100,000 stall-event entries with an 80/20 train/test split. Because the continued-watch to exit ratio is about 4:1, the paper uses balanced sampling with random undersampling of the majority class. The ablation results justify the stall-centric design: using all playback states produces accuracy and recall just over 60% with precision and F1 around 20%; using only switch-plus-stall “relevant events” raises accuracy to around 90% but leaves precision around 40%; using stall-only data yields all metrics above 95%.

Optimization combines Monte Carlo simulation with online Bayesian optimization (OBO). The paper states the objective as finding

$top1ctrl(n)=\big\{x\mid p(x)\ge\max{p(x)}/n\big\}.$4

where $top1ctrl(n)=\big\{x\mid p(x)\ge\max{p(x)}/n\big\}.$5 is the vector of QoE or ABR parameters, $top1ctrl(n)=\big\{x\mid p(x)\ge\max{p(x)}/n\big\}.$6 the bandwidth model, $top1ctrl(n)=\big\{x\mid p(x)\ge\max{p(x)}/n\big\}.$7 the user state, and $top1ctrl(n)=\big\{x\mid p(x)\ge\max{p(x)}/n\big\}.$8 the player environment. Future bandwidth is sampled from

$top1ctrl(n)=\big\{x\mid p(x)\ge\max{p(x)}/n\big\}.$9

and the simulated playback dynamics update buffer state, waiting time, and segment choice under the candidate ABR parameters. In deployment with HYB, the optimized parameter is the control knob $NS(q)$0. The optimization is triggered only after two stall events, and the system uses two pruning strategies: early termination during virtual playback if a candidate’s predicted exit rate already exceeds the best observed, and pre-playback pruning when

$NS(q)$1

so that stalls are highly unlikely. With these controls, the paper states that a daily active user needs at most 1–2 LingXi invocations per day.

The deployment evidence comes from a 10-day difference-in-differences A/B test using 8% of total traffic, corresponding to about 30 million DAU. The main watch-time result is

$NS(q)$2

with $NS(q)$3 and $NS(q)$4, reported in the abstract as a 0.15% increase in total viewing time. QoS also improves, with bitrate

$NS(q)$5

and stall time

$NS(q)$6

corresponding to a 0.1% bitrate improvement and 1.3% stall-time reduction overall. The strongest subgroup effect is for users below 2000 kbps, for whom LingXi reduces stall time by up to 15% while keeping video quality comparable.

6. Comparative themes and distinctions

Across these four usages, the name LingXi marks very different technical objects. In psychiatry, it denotes a benchmark ecosystem with a synthetic dataset, patient simulator, doctor strategies, and diagnostic evaluation. In poetry generation, it denotes a decoding-oriented generation system built on GPT-2-style modeling. In software engineering, it denotes a knowledge-guided agent framework for repository-level issue resolution. In streaming, it denotes a production optimization layer for adaptive bitrate control.

The methodological commonality is limited. Each system performs a different kind of guided decision-making: LingxiDiagBench separates consultation behavior from diagnostic reasoning; the poetry system reshapes token probabilities through head randomization and topic-preserving rejection sampling; the software framework guides analysis using retrieved procedural knowledge; and the streaming system adjusts ABR objectives using predicted exit behavior. This suggests a shared naming preference rather than a shared algorithmic lineage.

The most important disambiguation concerns real-world status and intended use. The psychiatric LingXi is explicitly a research benchmark and not a clinical diagnostic tool. The streaming LingXi is explicitly a deployed production system at Kuaishou. The software Lingxi is evaluated on SWE-bench Verified rather than in a live development workflow. The poetry Lingxi is a generation system evaluated through automatic metrics and human annotation rather than deployment at scale.

Taken together, the arXiv record uses “LingXi” to name four distinct systems: a clinically grounded Chinese psychiatric benchmark exposing the gap between information gathering and diagnosis (Xu et al., 10 Feb 2026); a Chinese modern poetry generator centered on NS-RH and semantic rejection sampling (Zhang et al., 2021); a procedural-knowledge-guided framework for repository-level issue resolution (Yang et al., 13 Oct 2025); and a large-scale personalized adaptive streaming system optimized around exit rate and watch time (Jia et al., 22 Aug 2025).