Update README.md

README.md

@@ -11,7 +11,6 @@ tags:
 - factuality
 - structured-extraction
 - paper-understanding
-- claim-extraction
 - paper-to-code
 task_categories:
 - question-answering
@@ -26,11 +25,12 @@ language_creators:
 
 # SemanticAlign-Bench
 
-A benchmark dataset for evaluating AI agents' ability to extract **structured claims** from top-tier machine learning conference papers. The dataset tests whether agents can accurately capture four dimensions of paper content — numerical facts, method formulas, experimental protocols, and procedural pipelines — without hallucination, omission, or misordering.
+A benchmark for evaluating AI agents on **structured claim extraction** from top-tier ML conference papers. Each paper is decomposed into Semantic Alignment Units (SAU) — atomic, self-contained implementation propositions — across four diagnostic dimensions spanning numerical precision to pipeline-level workflow. Agents are evaluated on whether they can reproduce these claims without hallucination, omission, or misordering.
 
 ## Dataset Description
 
 - **Papers**: 30 papers from **ICLR 2025**, **ICML 2025**, and **NeurIPS 2025**, spanning 5 domains (6 papers each):
+
   | Domain | Count |
   |---|---|
   | Probabilistic Inference / Generative Models | 6 |
@@ -38,7 +38,8 @@ A benchmark dataset for evaluating AI agents' ability to extract **structured cl
   | Computer Vision | 6 |
   | NLP / LLM | 6 |
   | Numerical Methods / Scientific Computing | 6 |
-- **Total SAU Claims**: **1,426** (Semantic Alignment Unit — the atomic unit of paper understanding)
+
+- **Total SAU Claims**: **1,491**
 - **Size**: ~519 MB
 
 ### The Four SAU Dimensions
@@ -47,14 +48,14 @@ Each paper is decomposed into claims across four diagnostic dimensions, ordered
 
 | Dimension | Name | Count | Definition |
 |-----------|------|-------|------------|
-| **D1** | Numerical Precision | 497 | Hyperparameters, configuration values, thresholds, scaling factors |
-| **D2** | Formulas / Algorithms | 476 | Mathematical formulas, algorithm steps, architectural mechanisms |
-| **D3** | Experiment Protocols | 291 | Datasets, baselines, evaluation metrics, experimental scope |
-| **D4** | Pipelines / Procedures | 162 | Multi-step execution order: phase ordering, algorithm step sequencing |
+| **D1** | Numerical Precision | 523 | Hyperparameters, configuration values, thresholds, scaling factors |
+| **D2** | Formulas / Algorithms | 503 | Mathematical formulas, algorithm steps, architectural mechanisms |
+| **D3** | Experiment Protocols | 300 | Datasets, baselines, evaluation metrics, experimental scope |
+| **D4** | Pipelines / Procedures | 165 | Multi-step execution order: phase ordering, algorithm step sequencing |
 
-### Paper Venue Distribution
+The D1--D4 hierarchy is universal across all evaluated configurations: D1 > D2 > D4 > D3 in score holds invariant for all 12 generator setups (Claude/DeepSeek/Gemini/GPT-4o × BasicAgent/PaperCoder/OpenHands). D3 (experimental protocol) is the dominant bottleneck, with only 0.7% perfect-score rate — 14× lower than D1. D4 exhibits a distinctive pattern: lowest zero rate (33.7%) but only 5.9% of claims score ≥0.5, meaning agents almost always attempt ordering constraints but rarely get them right.
 
-All 30 papers are from top-tier 2025 ML conferences:
+### Paper Venue Distribution
 
 | Venue | Count |
 |-------|-------|
@@ -73,7 +74,7 @@ All 30 papers are from top-tier 2025 ML conferences:
   paper.pdf         # Original PDF
   sau.json          # SAU claims — the core annotation file
   images/           # Paper figures extracted from PDF
-  blacklist.txt     # Tokens excluded from extraction (e.g., author names)
+  blacklist.txt     # official repo url
 ```
 
 ### SAU Claim Format (`sau.json`)
@@ -96,13 +97,13 @@ All 30 papers are from top-tier 2025 ML conferences:
 ```
 
 Each claim includes:
-- `id`: Unique identifier for cross-referencing (`{paper}-{dimension}-{number}`)
-- `claim`: The extracted claim text (natural language, self-contained statement)
+- `id`: Unique identifier (`{paper}-{dimension}-{number}`)
+- `claim`: Self-contained implementation proposition in natural language
 - `source`: Paper section where the claim originates
 
 ### Annotation Quality
 
-All 1,426 claims have undergone **multi-version human review** with systematic error checks:
+All 1,491 claims have undergone **multi-version human review** with systematic error checks:
 - Verification against source paper for factual accuracy
 - Format normalization and consistency validation
 - Cross-reference integrity checks between dimensions
@@ -110,9 +111,7 @@ All 1,426 claims have undergone **multi-version human review** with systematic e
 
 ## Supported Tasks
 
-The benchmark supports the following evaluation scenarios:
-
-1. **Claim-Level Factuality**: Given a paper, can the agent accurately extract or recall a specific numerical value, formula, experimental detail, or procedural step?
+1. **Claim-Level Factuality**: Given a paper, can the agent accurately extract a specific numerical value, formula, experimental detail, or procedural step?
 2. **Dimension-Level Completeness**: Can the agent achieve full recall across all four SAU dimensions for a given paper?
 3. **Cross-Dimensional Consistency**: Are claims in D4 (pipelines) consistent with D2 (formulas) and D3 (experiments)?
 4. **Hallucination Detection**: Can the agent distinguish paper-supported claims from plausible but fabricated ones?
@@ -121,61 +120,48 @@ The benchmark supports the following evaluation scenarios:
 
 ### Source Data
 
-The 30 papers were selected from top ML venues (ICLR, ICML, NeurIPS 2025) to cover:
-- **Domain diversity**: 5 domains with equal representation
-- **Task diversity**: Classification, generation, reinforcement learning, theory, scientific computing
-- **Quality diversity**: Papers span both regular and spotlight-level acceptances across three top-tier venues
-
-### Annotation Process
+30 papers selected from ICLR 2025, ICML 2025, and NeurIPS 2025, covering 5 domains with equal representation across task types (classification, generation, RL, theory, scientific computing).
 
-1. **Initial Extraction**: AI agent extracts SAU claims from `paper.md` via a multi-agent extraction pipeline (3 specialized extraction agents targeting D1/D2/D3, with per-section parallel sub-agent spawning to avoid attention degradation)
-2. **Human Audit**: Expert reviewer verifies each claim against the source paper, correcting hallucinations, omissions, and misorderings
-3. **Iterative Refinement**: Claims undergo 4–5 review cycles (versions v1→v5) until all systematic errors are resolved
-4. **Cross-Paper Validation**: Consistency checks across papers in the same domain
+## Evaluation Results
 
-### Annotations
+In a benchmark study evaluating 360 paper-level runs (12 generators × 30 papers):
 
-- **Creators**: Expert-generated (AI extraction → human verification pipeline)
-- **Review Depth**: Multi-pass — each paper received 4–5 review iterations
-- **Quality Metrics**: Fairness audit scores (all 30 papers grade A)
+- **Overall SAS**: mean 0.221, median 0.200. 82.4% of SAU claims score ≤0.25.
+- **Model dominance**: Model choice drives 2.35× more score variation than scaffold choice (1.15×). Top 5 configurations all use Claude or DeepSeek; bottom 3 all use GPT-4o.
+- **Scaffold asymmetry**: PaperCoder (+0.116 for GPT-4o) provides more benefit to weaker models. OpenHands adds near-zero value without minimum planning competence.
+- **Failure pattern**: 81% of zero-scored claims contain partial but incorrect code; only 5.7% are completely absent. Improving scores requires better comprehension, not broader coverage.
+- **Paper difficulty**: Numerical methods/PDE papers dominate the easiest tier; multi-modal systems and complex training pipelines the hardest.
 
 ## Considerations for Using the Data
 
-### Biases and Limitations
+### Limitations
 
-- **Venue Bias**: All papers are from 2025 top-tier conferences; older, lower-tier, or non-English venues are not represented
-- **Domain Coverage**: 5 domains with 6 papers each — useful for balanced evaluation but not comprehensive
-- **Single Annotator**: Claims were verified by one expert reviewer; inter-annotator agreement is not available
-- **Derived Work**: SAU claims are interpretation-laden; different readers may reasonably disagree on claim boundaries
+This is a **static benchmark**: claims test specification fidelity (did the agent encode the right parameters, formulas, and protocols?) rather than runtime correctness. The benchmark does not include execution-based evaluation or dynamic testing.
 
 ### Intended Use
 
-This dataset is designed for **evaluation**, not training. Use cases:
 - Benchmarking LLM factuality on scientific content
 - Measuring agent understanding of structured paper content
 - Stress-testing retrieval-augmented generation (RAG) over academic papers
 
 ### Out-of-Scope Uses
 
-- Training data for production LLMs (limited size, single annotator, venue bias)
-- Legal or regulatory compliance evaluation
+- Training data for production LLMs (limited size, single annotator)
 - Automated paper review or acceptance prediction
 
 ## Additional Information
 
 ### License
 
-The SAU annotations in this dataset are licensed under **CC-BY-4.0**. The underlying papers are subject to their original copyright terms as posted on arXiv and the respective conference proceedings.
+SAU annotations are licensed under **CC-BY-4.0**. Underlying papers are subject to their original copyright terms as posted on arXiv and respective conference proceedings.
 
 ### Citation
 
-If you use SemanticAlign-Bench in your research, please cite:
-
 ```bibtex
 @inproceedings{semanticalign_bench,
   title     = {SemanticAlign-Bench: Evaluating Semantic Alignment in LLM-Based Paper Reproduction},
   author    = {Anonymous Author(s)},
-  year      = {2026},
+  year      = {2025},
   note      = {Benchmark dataset at \url{https://anonymous-hf.up.railway.app/a/rrgn430zpfui/}}
 }
 ```