Show HN: Detect LLM hallucinations via geometric drift (0.9 AUC, 1% overhead)

[yubainu]

I built SIB-ENGINE, a real-time hallucination detection system that monitors LLM internal structure rather than output content.

KEY RESULTS (Gemma-2B, N=1000): • 54% hallucination detection with 7% false positive rate • <1% computational overhead (runs on RTX 3050 with 4GB VRAM) • ROC-AUC: 0.8995

WHY IT'S DIFFERENT: Traditional methods analyze the output text semantically. SIB-ENGINE monitors "geometric drift" in hidden states during generation - identifying the structural collapse of the latent space before the first incorrect token is sampled.

This approach offers unique advantages: • Real-time intervention: Stop generation mid-stream • Language-agnostic: No semantic analysis needed • Privacy-preserving: Never reads the actual content • Extremely lightweight: Works on consumer hardware

HOW IT WORKS: SIB-ENGINE monitors the internal stability of the model's computation. While the system utilizes multiple structural signals to detect instability, two primary indicators include:

Representation Stability: Tracking how the initial intent is preserved or distorted as it moves through the model's transformation space.

Cross-Layer Alignment: Monitoring the consensus of information processing across different neural depths to identify early-stage divergence.

When these (and other proprietary structural signals) deviate from the expected stable manifold, the system flags a potential hallucination before it manifests in the output.

DEMO & CODE: • Demo video: https://www.youtube.com/watch?v=H1_zDC0SXQ8 • GitHub: https://github.com/yubainu/sibainu-engine • Raw data: raw_logs.csv (full transparency)

LIMITATIONS: • Tested on Gemma-2B only (2.5B parameters) • Designed to scale, but needs validation on larger models • Catches "structurally unstable" hallucinations (about half) • Best used as first-line defense in ensemble systems

TECHNICAL NOTES: • No external models needed (unlike self-consistency methods) • No knowledge bases required (unlike RAG approaches) • Adds ~1% inference time vs. 300-500% for semantic methods • Works by monitoring the process not the product

I'd love feedback on: • Validation on larger models (Seeking strategic partnerships and compute resources for large-scale validation.) • Integration patterns for production systems • Comparison with other structural approaches • Edge cases where geometric signals fail

This represents a fundamentally different paradigm: instead of asking "is this text correct?", we ask "was the generation process unstable?" The answer is surprisingly informative.

Happy to discuss technical details in the comments!

[yubainu]

I’ve been exploring why LLMs "break" during inference. Most current hallucination detection methods look at the final text (semantic analysis) or use another LLM to double-check (self-consistency). These are effective but extremely slow and expensive.

SIB-ENGINE is my attempt to solve this at the geometric layer. By monitoring the "Anchor Drift" (how hidden states deviate from the prompt’s latent trajectory), I found that hallucinations often manifest as a structural instability before the token is even sampled.

The Numbers:

Recall: 53.89% (It catches about half, but it's consistent)

Precision: 88.52% (Low false-alarm rate is my priority)

Overhead: <1% (Running on an RTX 3050 with 4GB VRAM)

AUC: 0.8995

I've released a Lite version (1-axis) on GitHub so you can see the fundamental logic and run it on your own machine. I’ve also included the raw_logs.csv from my N=1000 test run on Gemma-2B for full transparency.

I’m particularly curious if anyone here has experimented with similar geometric approaches or has thoughts on how this might scale to 70B+ models where the latent space is significantly denser.

Happy to dive into the technical details!