gnosis-llm-self-awareness

name	gnosis-llm-self-awareness
title	Can LLMs Predict Their Own Failures? Self-Awareness via Internal Circuits
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2512.20578
keywords	["LLM Reliability","Self-Awareness","Uncertainty","Internal States","Interpretability"]
description	Enable frozen LLMs to predict their own correctness by decoding signals from internal hidden states and attention patterns, achieving reliable self-verification without external judges—adding only 5M parameters while reducing inference cost and improving calibration.

Overview

Gnosis is a lightweight self-awareness mechanism enabling frozen LLMs to perform intrinsic self-verification by inspecting internal computation states. It addresses a critical problem: LLMs generate confident but incorrect outputs, and detecting failures typically requires expensive external judges or multiple inference passes.

Core Innovation: Rather than relying on external verification, extract correctness prediction signals directly from the LLM's internal states—the hidden representations and attention patterns that encode the model's confidence and reasoning process.

Architecture

State Extraction Layer

Capture internal states at each layer of the transformer:

Hidden States:

• Extract final token's hidden representations from each transformer layer
• Retain information about accumulated computation and reasoning

Attention Patterns:

• Track attention weights across heads and layers
• Reveal what model attends to when reasoning
• Indicate confidence through attention concentration

Multi-Layer Analysis:

• Early layers encode syntactic/surface information
• Middle layers perform semantic reasoning
• Later layers prepare for output generation
• Different layers provide complementary correctness signals

State Compression

Compress extracted states into fixed-budget descriptors:

def compress_internal_states(hidden_states, attention_patterns):
    """Compress multi-layer states to fixed-size descriptor."""

    # Extract key statistics from each layer
    descriptors = []
    for layer_idx, (hidden, attn) in enumerate(
        zip(hidden_states, attention_patterns)
    ):
        # Hidden state statistics
        mean_activation = hidden.mean(dim=-1)
        activation_variance = hidden.var(dim=-1)

        # Attention statistics
        attention_entropy = calculate_entropy(attn)
        attention_concentration = attn.max()

        # Temporal statistics (change across tokens)
        activation_change = (hidden[:, -] - hidden[:, -]).norm()

        layer_descriptor = {
            : layer_idx,
            : mean_activation,
            : activation_variance,
            : attention_entropy,
            : attention_concentration,
            : activation_change,
        }
        descriptors.append(layer_descriptor)

     descriptors