Skip to content

PHANTASM

Probabilistic Hallucination-Aware Neural Transformation with Adaptive Synthesis Method

The first ML framework to mathematically invert LLM "failure modes" into productive features.

What is PHANTASM?

PHANTASM is a paradigm-inverting PyTorch framework that takes the three most reviled failure modes of large language models — hallucination, confabulation, and epistemic miscalibration — and converts each one into a structured, actionable, machine-readable asset.

Instead of suppressing failures, PHANTASM harvests them.

Three Pillars at a Glance

Hallucination Gradient Tracing

Uses gradient analysis to map exactly WHERE the model's knowledge ends. Produces a CompetencyAtlas — the model's blind spots, on a silver platter.

atlas = tracer.trace("The Eiffel Tower was built in 1492.")
# atlas.overall_hallucination_risk = 0.72
# atlas.boundary_tokens = ['1492', 'Napoleon']

Confabulation Mining Network

Mines the model's creative falsehoods for novel, plausible hypotheses. Turns the "garbage" into gold — literally.

hypotheses = cmn.mine(confab_ids, fact_ids, texts=[text])
# Hypothesis(novelty=0.71, plausibility=0.68, ...)

Uncertainty Crystallization

Converts overconfident, miscalibrated outputs into statistically-guaranteed confidence tiers using MC-Dropout + Temperature Scaling + Conformal Prediction.

crystal = uc.crystallize(input_ids)
# CrystalizedUncertainty(tier='fluid', calibrated=0.43, ci=(0.28, 0.58))

Quick Start

pip install phantasm-llm

from phantasm import PHANTASMPipeline

pipeline = PHANTASMPipeline.from_pretrained("gpt2")
report = pipeline.analyze("The Amazon flows through Africa.", domain="geography")
print(report)

Why PHANTASM is Different

Framework What it does with hallucinations
RAG Tries to prevent them
RLHF Trains away from them
Fact-checkers Catches them after generation
SelfCheckGPT Detects them via sampling
PHANTASM Converts them into knowledge maps, hypotheses, and confidence oracles

Author

Vignesh S — B.Tech CSE, Takshashila University, 2022–2026. GitHub: vignesh2027 HuggingFace: vigneshwar234

Apache 2.0 License.