sympathyai 0.1.1

Structure-aware evaluation and alignment for LLM reasoning

SympathyAI

Structure-aware evaluation and alignment for LLM reasoning

---

Motivation

Large language models optimised via likelihood maximisation produce fluent text — but fluency is not reasoning.

In clinical settings, a response can be grammatically perfect, statistically plausible, and still:

• contradict itself internally,
• ignore established diagnostic guidelines (DSM-5, NICE, AHA),
• fail to link symptoms causally to a diagnosis.

SympathyAI introduces a sympathetic evaluation framework: a lightweight, composable set of structure-aware metrics that go beyond perplexity to measure whether a model response actually reasons correctly.

Inspired by François Chollet's critique of gradient-based optimisation as a sufficient proxy for intelligence.

---

Three-Line Quick Start

from sympathy import evaluate, rerank

# Score a single response
score = evaluate("Symptoms suggest MDD; DSM-5 criteria are met.", keywords=["depression", "DSM"])

# Select the best response from a candidate list
best = rerank(responses, keywords=["myocardial infarction", "ECG", "troponin"])

---

What SympathyAI Measures

Metric	What it captures	Fast heuristic	LLM-judge
Coherence	Internal consistency; absence of contradictions	✅	✅
Guideline adherence	Coverage of required clinical concepts	✅	—
Causal alignment	Symptom → diagnosis linkage	✅	—

All three are combined into a single sympathetic score:

sympathetic_score = 0.4 × coherence + 0.4 × guideline + 0.2 × causality

Weights are configurable.

---

Benchmark Results

Evaluated on 5 curated clinical vignettes (MDD, MI, Pneumonia, GAD, T2DM). Each case has a clinically correct response and a plausible-but-flawed response.

Case	Condition	Good ↑	Poor ↓	Δ	Re-rank ✓
MDD_001	Major Depressive Disorder	0.92	0.28	+0.64	✅
MI_001	Myocardial Infarction	0.80	0.36	+0.44	✅
PNE_001	Community-Acquired Pneumonia	0.80	0.40	+0.40	✅
GAD_001	Generalised Anxiety Disorder	0.89	0.28	+0.61	✅
DM2_001	Type 2 Diabetes Mellitus	0.88	0.36	+0.52	✅

Re-ranking accuracy: 5/5 (100%) · Average score delta: +0.52

Re-ranking consistently selects the clinically correct response without any additional model training.

---

Installation

git clone https://github.com/your-username/sympathyAI.git
cd sympathyAI
pip install -e .

No external dependencies required for the heuristic pipeline.

For LLM-as-judge coherence scoring:

pip install -e ".[llm]"

---

Project Structure

sympathyAI/
│
├── sympathy/
│   ├── scoring/
│   │   ├── coherence.py      # Heuristic + LLM-as-judge coherence
│   │   ├── causality.py      # Symptom-to-diagnosis causal alignment
│   │   └── guidelines.py     # Clinical keyword coverage
│   │
│   ├── rerank/
│   │   └── reranker.py       # Composite scoring + re-ranking
│   │
│   ├── datasets/
│   │   └── clinical_cases.py # 5 curated clinical vignettes
│   │
│   └── evaluation/
│       └── evaluator.py      # Benchmark runner + table printer
│
├── examples/
│   └── clinical_demo.py      # End-to-end demonstration
│
└── tests/
    └── test_sympathy.py      # Full unit test suite

---

Run the Demo

python examples/clinical_demo.py

Expected output:

Response : Symptoms are consistent with major depressive disorder...
  Coherence  : 1.00
  Guideline  : 0.75
  Causality  : 1.00
  â–¶ Total    : 0.90

Re-ranking accuracy: 5/5 (100%)
Average score delta (good − poor): +0.523

---

Run Tests

pytest tests/test_sympathy.py -v

---

Roadmap

• LLM-as-judge coherence scoring (GPT-4 meta-evaluation)
• Clinical grounding with DSM-5 and NICE guidelines knowledge base
• Contradiction detection via NLI models
• Prompt perturbation robustness analysis
• REST API / Python wrapper for integration into inference pipelines
• Extended benchmark (50+ clinical cases with expert annotation)

---

Citation

If you use SympathyAI in your research, please cite:

@misc{sympathyai2025,
  title  = {Beyond Likelihood: Evaluating Structure-Aware Reasoning in Clinical Language Models},
  author = {[Authors]},
  year   = {2025},
  note   = {https://github.com/your-username/sympathyAI}
}

---

License

MIT — free to use, extend, and publish with.