Skip to main content ホーム クリエイター adu2021 skillxiv tabdsr-tabular-reasoning-llms
tabdsr-tabular-reasoning-llms Improve language model performance on complex numerical reasoning over tables through three-stage pipeline: decompose questions into sub-questions, sanitize and clean table data, then generate executable Python code for precise computation.
インストールへ移動 Skills Marketplace コミュニティが作成したAIスキルを発見・探索
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/ADu2021/skillXiv --skill tabdsr-tabular-reasoning-llmsコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
Zipをダウンロード ダウンロード中... name tabdsr-tabular-reasoning-llms title TabDSR: Decompose, Sanitize, Reason for Tabular Data version 0.0.2 engine skillxiv-v0.0.2-claude-opus-4.6 license MIT url https://arxiv.org/abs/2511.02219 keywords ["Tabular QA","Table Understanding","Program Synthesis","Numerical Reasoning","Data Cleaning"] description Improve language model performance on complex numerical reasoning over tables through three-stage pipeline: decompose questions into sub-questions, sanitize and clean table data, then generate executable Python code for precise computation.
Title: Enable Accurate Tabular Reasoning Through Decomposition and Code Generation
Language models struggle with numerical reasoning over tables because they hallucinate calculations and misalign rows/columns. TabDSR tackles this with a three-stage pipeline: (1) decompose complex questions into simpler sub-queries, (2) sanitize table data (fix headers, standardize formats, remove noise), and (3) use code generation to compute answers precisely. Each agent specializes in one task, enabling modular improvement.
The key insight is using code as the reasoning substrate for numerical operations.
Core Concept
Three-Stage Tabular Reasoning Pipeline :
Query Decomposer : Break complex questions into tractable sub-questionsTable Sanitizer : Clean structural and content issues, standardize formatsProgram-of-Thoughts (PoT) Reasoner : Generate Python code for precise computationModular Design : Each component can be improved independentlyEvaluation : New CalTab151 benchmark prevents data leakage
Architecture Overview
Query Decomposer Agent : LLM-based question analyzer using linguistic cuesTable Sanitizer Agent : Multi-step cleaning (header reconstruction, value standardization)PoT Reasoner : Code generation using Pandas DataFramesExecution Environment : Sandboxed Python for safe code executionDataset : CalTab151 with multi-hop questions and human verification
Implementation Steps
1. Implement Query Decomposer
Analyze questions to identify sub-problems.
class QueryDecomposer :
def __init__ (self, llm_model ):
self .llm = llm_model
def analyze_query_complexity (self, query ):
"""Determine if question needs decomposition"""
complex_patterns = [
'compare' , 'how many' , 'what percentage'
'ratio'
'more than'
'less than'
'between'
'and then'
any
in
for
in
return
def
decompose_query
self, query, table_context
"""Break complex query into sub-questions"""
if
not
self
return
f"""Break this question into simpler sub-questions:
Question: {query}
Table context: {self.summarize_table(table_context)}
Sub-questions (numbered):"""
self
self
return
def
parse_subquestions
self, text
"""Extract numbered sub-questions from LLM output"""
import
r'\d+\.\s*([^0-9]+?)(?=\d+\.|$)'
return
for
in
def
summarize_table
self, table
"""Provide table summary for context"""
return
f"Columns: {list (table.columns)} , Rows: {len (table)} "
2. Implement Table Sanitizer
Clean structural and content issues in tables.
class TableSanitizer :
def __init__ (self ):
self .null_indicators = ['N/A' , 'NA' , 'null' , 'None' , '-' , '?' ]
def sanitize_headers (self, table ):
"""Fix and standardize column headers"""
if isinstance (table.columns, pd.MultiIndex):
table.columns = ['_' .join(col).strip() for col in table.columns]
table.columns = [col.lower().strip().replace(' ' , '_' ) for col in table.columns]
table.columns = [f"col_{i} " if col == "" else col for i, col in enumerate (table.columns)]
return table
def sanitize_content (self, table ):
"""Clean cell values"""
for col in table.columns:
try :
table[col] = pd.to_numeric(table[col], errors='coerce' )
except :
table[col] = table[col].astype(str )
for null_ind in self .null_indicators:
table[col] = table[col].replace(null_ind, None )
table = table.dropna(how='all' )
table = table.dropna(axis=1 , how='all' )
return table
def reconstruct_segmented_tables (self, table ):
"""Handle tables split across multiple sections"""
null_rows = table.isnull().all (axis=1 ).values
break_indices = np.where(null_rows)[0 ]
if len (break_indices) > 0 :
segments = []
start = 0
for break_idx in break_indices:
segment = table.iloc[start:break_idx]
segments.append(segment)
start = break_idx + 1
segments.append(table.iloc[start:])
return max (segments, key=lambda x: x.shape[0 ])
return table
def sanitize (self, table ):
"""Full sanitization pipeline"""
table = self .sanitize_headers(table)
table = self .sanitize_content(table)
table = self .reconstruct_segmented_tables(table)
return table
3. Implement Program-of-Thoughts Reasoner
Generate executable Python code for computation.
class ProgramOfThoughtsReasoner :
def __init__ (self, code_llm, execution_env ):
self .code_llm = code_llm
self .exec_env = execution_env
def generate_program (self, question, table_df ):
"""Generate Python code to answer question"""
schema = self .get_table_schema(table_df)
code_prompt = f"""Write Python code to answer this question:
Question: {question}
Table schema:
{schema}
Available variable: 'df' is a pandas DataFrame with the table.
Code (complete and executable):
"""
code = self .code_llm.generate(code_prompt, max_tokens=500 )
return code
def get_table_schema (self, df ):
"""Describe table structure and dtypes"""
schema = f"Columns: {list (df.columns)} \n"
schema += f"Dtypes: {dict (df.dtypes)} \n"
schema += f"Shape: {df.shape} \n"
schema += f"Sample row: {df.iloc[0 ].to_dict()} "
return schema
def execute_program (self, code, df ):
"""Execute Python code in sandbox"""
local_context = {'df' : df, 'pd' : pd, 'np' : np}
try :
exec (code, {}, local_context)
if 'answer' in local_context:
return local_context['answer' ]
elif 'result' in local_context:
return local_context['result' ]
else :
return self .extract_final_value(local_context)
except Exception as e:
return f"Execution error: {str (e)} "
def extract_final_value (self, context ):
"""Extract final computed value from context"""
non_builtins = {k: v for k, v in context.items() if not k.startswith('_' )}
for k, v in sorted (non_builtins.items(), reverse=True ):
if isinstance (v, (int , float , str )):
return v
return None
def reason (self, question, table_df ):
"""Full reasoning pipeline"""
code = self .generate_program(question, table_df)
answer = self .execute_program(code, table_df)
return answer, code
4. Integrate Three Stages
Combine pipeline into unified interface.
class TabDSRPipeline :
def __init__ (self, decomposer, sanitizer, reasoner ):
self .decomposer = decomposer
self .sanitizer = sanitizer
self .reasoner = reasoner
def answer_question (self, question, table ):
"""Answer tabular question using three-stage pipeline"""
sub_questions = self .decomposer.decompose_query(question, table)
clean_table = self .sanitizer.sanitize(table)
if len (sub_questions) == 1 :
answer, code = self .reasoner.reason(question, clean_table)
else :
sub_answers = []
for sub_q in sub_questions:
sub_answer, _ = self .reasoner.reason(sub_q, clean_table)
sub_answers.append(sub_answer)
answer = self .combine_answers(question, sub_answers, clean_table)
code = "# See sub-questions above"
return {
'answer' : answer,
'code' : code,
'sub_questions' : sub_questions,
'cleaned_table' : clean_table
}
def combine_answers (self, original_question, sub_answers, table ):
"""Merge sub-question answers into final result"""
combine_prompt = f"""Combine these sub-answers to answer the original question:
Original: {original_question}
Sub-answers: {sub_answers}
Final answer:"""
combined = self .reasoner.code_llm.generate(combine_prompt)
return combined
Practical Guidance
Numerical QA over business/scientific tables
Tasks with complex multi-hop reasoning
Settings where accuracy is critical (hallucinations unacceptable)
decomposition_threshold: Complexity level triggering decomposition
null_replacement: How to handle missing values
code_max_tokens: 500-1000 depending on problem complexity
Simple lookup questions (single cell retrieval)
Tables with highly irregular structure
Domains where code generation fails (expert knowledge needed)
Code generation hallucination : Model may generate syntactically valid but logically wrong code; validate with intermediate resultsOver-decomposition : Some questions don't benefit from decomposition; check impact on resultsTable cleaning over-aggressiveness : Over-sanitizing can remove important nuance Integration Strategy : Use as fallback for numerical questions. Combine with semantic matching for hybrid QA systems.
Reference 