RQ5: Mitigation and software workflow
Mitigation can act during dataset construction, model training, or generation. The intervention should match the diagnosed property and be evaluated against its intended output quality and possible regressions.
| Family | Useful conditions | Benefit | Cost, trade-off, and evidence limit |
|---|---|---|---|
| Deterministic validation and filtering | A checkable sample defect and another training run | Removes specified invalid, vulnerable, duplicate, or unsuitable material | Rules can discard useful samples; bundled pipelines rarely isolate each filter |
| Scoring, sampling, and expansion | Quality signals, a target mixture, or verified new sources | Selects graded quality and broadens relevant coverage | Labels and mixture targets require judgment; scoring and synthesis add cost and can repeat bias |
| Rewriting and transformation | Useful samples have recoverable presentation, alignment, or implementation defects | Retains information while changing a defined property | Requires validation; improvements can be task-dependent and can cause regressions |
| Model adaptation and reward-guided training | Weights or prefixes can be updated with relevant examples or rewards | Amortizes a target behavior across requests | Training and regression evaluation cost; reward proxies can conflict with functionality |
| Prompt, retrieval, and decoding control | Constraints or current context vary by request | Supplies targeted guidance or checker feedback at inference | Retrieval and decoding add latency; poor context can harm otherwise familiar tasks |
| Feedback-guided or deterministic repair | A compiler, test, analyzer, API map, or rule localizes a defect | Reuses development tools to repair candidates | Repeated calls and tests cost time; repairs can introduce defects or plateau |
Generated-code mitigation
Data-level mitigation is detailed in the next section. Its effect on generated code requires training and held-out output evaluation.
| Method | Path in taxonomy | Papers |
|---|---|---|
| Model Adaptation and Fine-Tuning | Model-Level Mitigation > Model Adaptation and Fine-Tuning | 6 |
| Reward-Guided Training | Model-Level Mitigation > Reward-Guided Training | 3 |
| Training-Time Repetition Control | Model-Level Mitigation > Training-Time Repetition Control | 1 |
| Prompt Constraints | Generation-Level Mitigation > Pre-Generation > Prompt Constraints | 4 |
| Privacy-Preserving Input Transformation | Generation-Level Mitigation > Pre-Generation > Privacy-Preserving Input Transformation | 1 |
| Retrieval-Augmented Generation | Generation-Level Mitigation > Pre-Generation > Retrieval-Augmented Generation | 5 |
| Decoding-Time Intervention | Generation-Level Mitigation > In-Generation > Decoding-Time Intervention | 4 |
| Feedback-Guided Regeneration | Generation-Level Mitigation > Post-Generation > Feedback-Guided Regeneration | 3 |
| Deterministic Repair and Transformation | Generation-Level Mitigation > Post-Generation > Deterministic Repair and Transformation | 2 |
Select a paper count to open its supporting papers. Counts are multi-label literature memberships. Download the method definitions and paper–method memberships to trace each entry.
Training-data mitigation
| Method | Path in taxonomy | Papers |
|---|---|---|
| Execution-Feedback-Based Filtering | Data Cleaning and Filtering > Execution-Feedback-Based Filtering | 6 |
| Deterministic Validation, Filtering, Sanitization, and Deduplication | Data Cleaning and Filtering > Deterministic Validation, Filtering, Sanitization, and Deduplication | 15 |
| LLM Scoring or Judging | Data Cleaning and Filtering > LLM Scoring or Judging | 8 |
| Learned Quality or Relevance Scorers | Data Cleaning and Filtering > Learned Quality or Relevance Scorers | 5 |
| Manual Review and Curation | Data Cleaning and Filtering > Manual Review and Curation | 4 |
| Mixture Design and Quality-Aware Sampling | Mixture Design and Quality-Aware Sampling | 6 |
| LLM-Based Rewriting and Semantic Correction | Data Enhancement > LLM-Based Rewriting and Semantic Correction | 3 |
| Rule- or Tool-Based Transformation | Data Enhancement > Rule- or Tool-Based Transformation | 1 |
| Curated-Source Expansion | Data Augmentation > Curated-Source Expansion | 2 |
| Synthetic Data Generation | Data Augmentation > Synthetic Data Generation | 13 |
Select a paper count to open its supporting papers. Counts are multi-label literature memberships. Download the method definitions and paper–method memberships to trace each entry.
Applying controls in a project
The following workflow is an author synthesis of separately evaluated controls. Included papers do not evaluate the complete arrangement as an integrated production system.
| Workflow point | Controls and retained evidence |
|---|---|
| Dataset intake | Record sources, licensing, provenance, versions, and coverage; check samples and training–evaluation overlap |
| Dependencies and build | Pin the environment; validate APIs and versions; compile or resolve packages |
| Local review | Run tests and static checks; examine intent, security, and maintainability in context |
| CI/CD and release | Apply project-specific gates; measure workloads and boundaries; return repaired candidates through applicable checks |
| Repository maintenance | Retain model, prompt, retrieval, artifact, and tool versions; investigate regressions as dependencies and requirements change |
Begin with deterministic checks, add execution and stress testing where an oracle exists, and use contextual triage and repair for residual cases. Projects set their own thresholds and escalation rules. The automation crosswalk records narrow exemplars and scope limits; it does not establish full issue coverage or production readiness.
Supporting papers
Browse 48 distinct papers by mitigation method, separated by the artifact being changed. Expand a method to see its taxonomy path and supporting papers, or select its paper count in the tables above. Membership is multi-label: a paper can appear under several methods. The paper–method memberships retain the exact assignments.
Generated-code mitigation
Model Adaptation and Fine-Tuning 6 papers
Model-Level Mitigation > Model Adaptation and Fine-Tuning
package hallucinations
We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs
Large Language Models for Code
Large Language Models for Code: Security Hardening and Adversarial Testing
ReCode
ReCode: Updating Code API Knowledge with Reinforcement Learning
CRPE
CRPE: Expanding The Reasoning Capability of Large Language Model for Code Generation
UCD-Training
Unseen-Codebases-Domain Data Synthesis and Training Based on Code Graphs
ShortCoder
ShortCoder: Knowledge-Augmented Syntax Optimization for Token-Efficient Code GenerationPreprint
Reward-Guided Training 3 papers
Model-Level Mitigation > Reward-Guided Training
Codequal Analyzer
Improving LLM-Generated Code Quality with GRPO
REAL
Training Language Models to Generate Quality Code with Program Analysis Feedback
ReCode
ReCode: Updating Code API Knowledge with Reinforcement Learning
Training-Time Repetition Control 1 paper
Model-Level Mitigation > Training-Time Repetition Control
Repetition In Repetition Out
Repetition In Repetition Out: Towards Understanding Neural Text Degeneration from the Data Perspective
Prompt Constraints 4 papers
Generation-Level Mitigation > Pre-Generation > Prompt Constraints
LLMs Meet Library Evolution
LLMs Meet Library Evolution: Evaluating Deprecated API Usage in LLM-based Code Completion
SStuBs
Large Language Models and Simple, Stupid Bugs
Syntactic Robustness
Syntactic Robustness for LLM-based Code Generation
Beyond Functional Correctness
Beyond functional correctness: Investigating coding style inconsistencies in large language models
Privacy-Preserving Input Transformation 1 paper
Generation-Level Mitigation > Pre-Generation > Privacy-Preserving Input Transformation
CodeCipher
CodeCipher: Learning to Obfuscate Source Code Against LLMs
Retrieval-Augmented Generation 5 papers
Generation-Level Mitigation > Pre-Generation > Retrieval-Augmented Generation
package hallucinations
We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs
CloudAPIBench
On Mitigating Code LLM Hallucinations with API Documentation
AllianceCoder
What to retrieve for effective retrieval-augmented code generation? an empirical study and beyond
RustEvo^ 2
RustEvo^ 2: An Evolving Benchmark for API Evolution in LLM-based Rust Code Generation
Llm Hallucinations in Practical Code Generation
Llm hallucinations in practical code generation: Phenomena, mechanism, and mitigation
Decoding-Time Intervention 4 papers
Generation-Level Mitigation > In-Generation > Decoding-Time Intervention
CodeIP
CodeIP: A Grammar-Guided Multi-Bit Watermark for Large Language Models of Code
RPG
Rethinking Repetition Problems of LLMs in Code Generation
Adadec
Adadec: Uncertainty-guided adaptive decoding for llm-based code generation
RobGen
A Preliminary Study on the Robustness of Code Generation by Large Language Models
Feedback-Guided Regeneration 3 papers
Generation-Level Mitigation > Post-Generation > Feedback-Guided Regeneration
package hallucinations
We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs
Refining ChatGPT-Generated Code
Refining ChatGPT-Generated Code: Characterizing and Mitigating Code Quality Issues
MultiCodeIF
A hierarchical and evolvable benchmark for fine-grained code instruction following with multi-turn feedback
Deterministic Repair and Transformation 2 papers
Generation-Level Mitigation > Post-Generation > Deterministic Repair and Transformation
LLMs Meet Library Evolution
LLMs Meet Library Evolution: Evaluating Deprecated API Usage in LLM-based Code Completion
Code Copycat Conundrum
Code Copycat Conundrum: Demystifying Repetition in LLM-based Code Generation
Training-data mitigation
Execution-Feedback-Based Filtering 6 papers
Data Cleaning and Filtering > Execution-Feedback-Based Filtering
Qwen2
Qwen2 Technical Report
Qwen2.5
Qwen2.5 Technical Report
TeleChat
Technical Report of TeleChat2, TeleChat2.5 and T1
Kimi K2
Kimi K2: Open Agentic Intelligence
Seed-Coder
Seed-Coder: Let the Code Model Curate Data for Itself
Code Llama
Code Llama: Open Foundation Models for Code
Deterministic Validation, Filtering, Sanitization, and Deduplication 15 papers
Data Cleaning and Filtering > Deterministic Validation, Filtering, Sanitization, and Deduplication
Less is More
Less is More: On the Importance of Data Quality for Unit Test Generation
Qwen
Qwen Technical Report
Qwen2
Qwen2 Technical Report
Quality In, Quality Out
Quality In, Quality Out: Investigating Training Data's Role in AI Code Generation
SwallowCode
Rewriting Pre-Training Data Boosts LLM Performance in Math and Code
Qwen2.5
Qwen2.5 Technical Report
TeleChat
Technical Report of TeleChat2, TeleChat2.5 and T1
Seed-Coder
Seed-Coder: Let the Code Model Curate Data for Itself
DeepSeek-Coder
DeepSeek-Coder: When the Large Language Model Meets Programming -- The Rise of Code Intelligence
StarCoder 2 and The Stack v2
StarCoder 2 and The Stack v2: The Next Generation
DataComp-LM
DataComp-LM: In search of the next generation of training sets for language models
RedStone
RedStone: Curating General, Code, Math, and QA Data for Large Language Models
Code Llama
Code Llama: Open Foundation Models for Code
MG-Verilog
MG-Verilog: Multi-grained Dataset Towards Enhanced LLM-assisted Verilog Generation
aiXcoder-7B
aiXcoder-7B: A Lightweight and Effective Large Language Model for Code Processing
LLM Scoring or Judging 8 papers
Data Cleaning and Filtering > LLM Scoring or Judging
Qwen
Qwen Technical Report
Qwen2
Qwen2 Technical Report
DataMan
DataMan: Data Manager for Pre-training Large Language Models
Qwen3
Qwen3 Technical Report
Qwen2.5
Qwen2.5 Technical Report
TeleChat
Technical Report of TeleChat2, TeleChat2.5 and T1
Kimi K2
Kimi K2: Open Agentic Intelligence
Seed-Coder
Seed-Coder: Let the Code Model Curate Data for Itself
Learned Quality or Relevance Scorers 5 papers
Data Cleaning and Filtering > Learned Quality or Relevance Scorers
Less is More
Less is More: On the Importance of Data Quality for Unit Test Generation
DataMan
DataMan: Data Manager for Pre-training Large Language Models
CIDRe
CIDRe: A Reference-Free Multi-Aspect Criterion for Code Comment Quality Measurement
Seed-Coder
Seed-Coder: Let the Code Model Curate Data for Itself
DataComp-LM
DataComp-LM: In search of the next generation of training sets for language models
Manual Review and Curation 4 papers
Data Cleaning and Filtering > Manual Review and Curation
Qwen
Qwen Technical Report
TeleChat
Technical Report of TeleChat2, TeleChat2.5 and T1
ReCode
ReCode: Updating Code API Knowledge with Reinforcement Learning
StarCoder 2 and The Stack v2
StarCoder 2 and The Stack v2: The Next Generation
Mixture Design and Quality-Aware Sampling 6 papers
Mixture Design and Quality-Aware Sampling
Qwen
Qwen Technical Report
DataMan
DataMan: Data Manager for Pre-training Large Language Models
Qwen3
Qwen3 Technical Report
TeleChat
Technical Report of TeleChat2, TeleChat2.5 and T1
Code Pretraining
How Does Code Pretraining Affect Language Model Task Performance?
Code Llama
Code Llama: Open Foundation Models for Code
LLM-Based Rewriting and Semantic Correction 3 papers
Data Enhancement > LLM-Based Rewriting and Semantic Correction
SwallowCode
Rewriting Pre-Training Data Boosts LLM Performance in Math and Code
Qwen3
Qwen3 Technical Report
DataRecipe
DataRecipe --- How to Cook the Data for CodeLLM?
Rule- or Tool-Based Transformation 1 paper
Data Enhancement > Rule- or Tool-Based Transformation
DataRecipe
DataRecipe --- How to Cook the Data for CodeLLM?
Curated-Source Expansion 2 papers
Data Augmentation > Curated-Source Expansion
StarCoder 2 and The Stack v2
StarCoder 2 and The Stack v2: The Next Generation
Code Llama
Code Llama: Open Foundation Models for Code
Synthetic Data Generation 13 papers
Data Augmentation > Synthetic Data Generation
Qwen2
Qwen2 Technical Report
Phi-4
Phi-4 Technical Report
Qwen3
Qwen3 Technical Report
Qwen2.5
Qwen2.5 Technical Report
Kimi K2
Kimi K2: Open Agentic Intelligence
ReCode
ReCode: Updating Code API Knowledge with Reinforcement Learning
Seed-Coder
Seed-Coder: Let the Code Model Curate Data for Itself
CRPE
CRPE: Expanding The Reasoning Capability of Large Language Model for Code Generation
Code Llama
Code Llama: Open Foundation Models for Code
MG-Verilog
MG-Verilog: Multi-grained Dataset Towards Enhanced LLM-assisted Verilog Generation
UCD-Training
Unseen-Codebases-Domain Data Synthesis and Training Based on Code Graphs
ShortCoder
ShortCoder: Knowledge-Augmented Syntax Optimization for Token-Efficient Code GenerationPreprint
APIKG4SYN
Framework-Aware Code Generation with API Knowledge Graph-Constructed Data: A Study on HarmonyOS